Where this is going

Ask support

Send a question to our team from the Ask box, with what you’re working on attached; the reply comes by email and shows here too.

Start the work on your site

One press on your project page starts the Evidence Engine on your own site, with what it spends shown first. No Google account is needed.

One set of first steps

Welcome, Help, Connections and the tour now show the same five first steps, and Needs you shows the same count everywhere.

Campaign end dates

Campaigns with a duration now show the date they end on the Campaigns list and their page, you get a notice two days before, and Edit campaign lets you set the duration, with 0 for no end.

Start here on every project

A project now opens on what we are doing for your site, in four lanes (Needs you, In flight, Measuring, Done), with the pages fetched and cited; the less-used tools sit under “More tools” in Create until you open it, and it stays open once you do.

Evidence Engine counts that take you there

Each count on the Evidence Engine page opens the list it counts, Needs you shows each item waiting on you, and before any gaps are found the page lists every way they are found, each with a link to start it.

Every engine on its own line

Run details, All clients and the before-readings now show each engine’s own count and range, in one colour per engine across every screen. Trained recall’s Cited says “not measured”.

Who leads each topic

For each topic and engine, see whether AI answers cite you most (Owned), share it (Contested) or cite others (Unowned), with the leading site and the counts behind it.

Your published pages, read back

We read each page we put on your site back about weekly and show where its markup, dates or canonical no longer match what we sent.

Buttons that say what they need

A gap that needs a job you did takes you to add it; every AI button shows its key and cost before it runs; a rejected AI key pauses with a “try it again” button.

Campaigns, entities and settings in the same clear layout

Campaign pages, the weekly report, the Entity pages, the SmartStack hub, AI Settings and Connections now open with tiles and one chart. Disconnect asks you to confirm in one clear step, and a failed Cloudflare read says what happened in plain words.

One count of what needs you

The Overview, its header and the work board now give the same number. Its card lists the first items with their buttons, and each count opens its lane. The engine chart opens on Cited.

Engine figures — with and without your brand

Each engine’s tile shows how often it names you on questions that carry your name, and on questions that don’t. Click either one to see only that, on the tiles and the chart; click again for both. Up or down is called only when the same questions were asked.

Clearer pages throughout

Every project tab, report and AI tool now opens with tiles and one chart. The tables, and how we count, are a click away.

Engine figures — with and without your brand

Each engine’s tile now leads with how often it names you on questions without your name, with all questions underneath. Up or down is called only when the same questions were asked. One chart shows every engine; pick one to see its range.

Overview at a glance

Three plain sentences up top: what changed for each AI engine, what we published this week against last, and what needs you. Then a tile per engine and one chart, with how we count a click away.

Clearer site check

Your scores, your top 3 fixes, and charts by area up front, with how we checked each one a click away. Each AI question says why we asked it.

Your first reading in minutes

Check your site, confirm what you do, and see what AI says about you. No campaign to set up and no Google account needed, and a trial shows a first look before you add an AI key.

Choose a cheaper AI for analysis

AI Settings has an Analysis order: pick which keys and models do the analysis (sorting topics, classifying sources, reviews), apart from the ones that write. We suggest a cheaper model from the same maker, and every AI button now says which key and model it will use before you press.

How we measure, in full

How we measure now lists every setting behind your figures: temperature, location, fallback, and the rules for position, tone and recommendation. Each engine’s trend line now starts again when its model changes.

No recorded job yet? An approach article

The Evidence Engine writes an approach article in your own voice: how you work, built from your own material, never an invented customer or figure.

Fixed it? Check again

Cards about your robots.txt or llms.txt in What to do next now have a Fixed/Recheck button. It reads that file again in seconds, with no full check and no AI, and clears the card once it’s fixed.

The sentences AI drew on

Sites AI cites now keeps a dated history, for each cited page, of the words each engine drew on and the question it was answering, with your own pages first.

More facts in your fact Sheet

Your fact Sheet now includes your legal name, other names, logo, address and contact details, Wikidata item, what you offer and who is behind the business, and every row says where it came from.

More gaps found for you

Each measurement run now finds gaps topic by topic, from answers that don’t name you, and topics added by discovery are checked too.

Sharper citation counts

A page counts as cited when the answer points to it; pages an engine read but didn’t cite show as looked at. Gap credit now reads “cited in 2 of 3 runs that asked about it”.

Searches the engines ran

Each topic shows the searches the AI engines ran while answering, with which engines ran them, and Add as a topic.

Cited, not named

The AI answers tab lists answers that cited your page without naming you, with the fix that applies.

Grok can search X

A watchlist switch lets Grok search X posts as well as the web.

More citations protect your pages

A page cited in the ChatGPT app, the Gemini app or Microsoft’s AI Performance report now waits for your approval before any rewrite.

See the page before you send it

Close it now shows the full page first, and update cards show exactly what will change.

Sources used the way their sites allow

Writers follow each cited site’s AI-use settings, including pages they find themselves.

Your entity hub, on the project page

The Pages we publish tab now holds the project’s entity hub: the entity page on your own site, the facts it carries and the llms.txt download, with Open and Edit going straight to this project’s entity. Settings links to it.

Connect Cloudflare by signing in

Under Connections, Cloudflare, press Connect Cloudflare and sign in, instead of creating and pasting a token. Cloudflare shows which account you are sharing and what we ask for, and we match each project to the zone that serves its site; where more than one fits, you choose. The nightly crawler read then runs through that sign-in. Tokens you pasted before keep working, and Disconnect withdraws our access at Cloudflare.

Searches that could become topics

Your project’s Search & topics tab now lists the searches Google and Bing showed your site for, each engine’s figures side by side and never added, with the stage the words point at and what each already is on your site. Add any as a topic, set one aside with Not this one, or track a question in its own words, with its cost on your own AI key shown first. Once a site has ten searches that are not your name, each shown at least 20 times by one engine, topic discovery adds up to five a pass on its own, until you have 60 main topics.

AI citation share on the Activity report

For each AI engine your runs measured, a ring shows how the sites its answers cited were shared between you, your competitors and every other site, each slice a count with its whole, such as “14 of 166 citations”, and engines are never added together. Tick “Include AI citation share” when you download the report’s PDF, or make a share link a client can open without signing in, for 7, 30 or 90 days and withdrawable at any time.

Your project page, in tabs

Each project now opens on an Overview with one To do list: what to do next, the review’s recommendations, and anything else waiting for you. The rest sits in tabs: AI answers, Pages we publish, Search & topics, Off-site and Settings. The header shows your next action and how many things need you on every tab, and a link from an email or the dashboard opens the tab its card is on.

Answer accuracy

Your project page now checks what AI answers say about your business against your own facts: prices, phone, founding year and address, and only the ones you have stored. When an answer says something different, you see what it said, which engines said it and the page it cited, read again today: if your page still says the old figure, fix it; if it no longer does, ask the engines to read it again. In an answer that compares several businesses, a price counts as yours only in your own sentence, table row or column.

New versions of the videos that are working

When a video of yours is cited in an AI answer, carried in a Google AI Overview or found on Google’s page one, the project page lists it under “Earning attention: videos worth remaking”, with a brief for your video maker to make a public or better version. Once theirs is on YouTube, add its link: we watch it for citations, and the original stays up as it is. For an unlisted campaign video, we can render a quick branded version instead, after showing what it costs.

Conversations to join

The project page’s Off-site tab now starts with forum threads to join on the questions where AI answers did not name you: Reddit, Quora, Stack Exchange and community forums, and X only where no forum thread passed. Each one is checked to take replies, to be from the last 6 months, and, on your own AI key, to be someone choosing or comparing what you sell. “Comments closed” sets one aside. Pages to pitch follow, with “It’s a competitor” for a publisher a rival owns.

Recommendations from a nightly review of each project

Every night, with no AI, each project is read for what is worth changing: tracked keywords that match no topic and never rank (“Still watching it?”), keys out of credit, cheaper like-for-like models, crawlers blocked in robots.txt, and more. Each comes with the figures behind it, nothing changes until you press a button, and an optional AI review on your own key puts the three to do first at the top.

Linkable sections, and your reviewer named

Every section of a page we publish on your site or your cloud pages can now be linked to directly. Once someone confirms a “Reviewed on” review, the page’s structured data names your project’s Content Author as its reviewer.

Publish a project’s pages to its cloud site instead of your blog

Once a project has a Cloud site set, “Where this project publishes” offers it beside your WordPress sites and other connections. Choose it and the Evidence Engine’s pages go to the cloud site’s pages/ folder, listed on its index page and sitemap and sent to IndexNow. A page sent again for the same subject replaces the one there, and with Review before publishing on, each page waits for your approval first.

Read now for AI crawlers

On AI crawlers on your site, Read now reads your Cloudflare zone straight away instead of waiting for tonight, and the card shows when the zone was last read.

Set up your cloud site from the project screen

Cloud site now lists your buckets (and says why when it can’t), makes a new public website bucket for you, lets you pick the folder from what’s already there, and checks your domain’s DNS, http and https live.

A running total for your keywords

The Keywords panel now shows how many of your tracked keywords are at #1, in the top 3 and the top 10, each counted once, with the week’s change and a daily chart. Beside it, every search Search Console shows your site for, and how many land on pages we published for you, tracked or not. Track these adds the top-10 searches you don’t track yet.

Your reports now open with what our pages rank for

The weekly email, What’s working and the Activity report each start with one line: how many searches in Google’s top 10 land on pages we published for you, tracked or not, how many of those you don’t track, and the change since last week. It is Search Console’s own 28-day average, never added to your rank checks.

Your entity page can live on your project’s cloud site

On the entity’s Canonical page card, “Where the page goes” now offers Your site (as before, and still the default) or Your cloud site. Choose the cloud site and press Publish: the same About page, with the same schema, goes to about/your-name/ on the cloud site set for the project, is listed under About on its index page and in its sitemap, and is rewritten in place whenever the entity changes. Its schema still names your business by your own website. One cloud site carries one business’s About page.

Choose what each Google account is used for, and give each location its own

Under Edit campaign, each Google account linked to a campaign has a “Used for” choice: Everything, Business Profile, Blogger (and which blog) or Drive & Docs. Blogger posts go out only from the Blogger account, Docs, Sheets and Slides are made only with the Drive account, and Business Profile posts come only from the Business Profile accounts. A campaign where nothing is chosen works exactly as before. For a brand with many locations on one website, each branch can have its own Business Profile account and location, Blogger account and blog, and Drive account; “Create campaigns for these branches” makes one campaign per branch with those accounts, its landing page, its town or city and a cloud folder of its own.

Each project can have its own cloud site

In Manage project, under Where this project publishes, Cloud site picks the storage connection, the bucket, the web address that serves it and an optional folder. The project’s cloud pages, case studies, Cloud Stack and Affiliate Stack all publish there, and the screen says before anything runs whether robots.txt and llms.txt will be written at the address’s root, or that the bucket needs a custom domain first.

Add WordPress categories from the category panel

Under each WordPress site, Add a category makes one on your site. When your site has fewer than three usable categories, the panel suggests up to eight from your topics and the kinds of page we publish there; tick the ones you want, change any name, and press Create. Nothing is created unless you press.

Your cloud folder is now one growing site

Every page we put in a folder of your cloud storage - from a campaign’s Cloud step, Cloud Stack, case studies or the Affiliate Stack - is listed on the folder’s index page and in its sitemap, earlier runs included, by section with titles and dates, and every page links back to that index page.

Tidy your AI Overview list - stop a search, or put it away for now

“Tidy this list” on Tools → AI Overview suggests searches to stop watching or put away, per site, when you press it - never on a schedule and with no AI call. Suggestions come only from what you told us (a topic you removed) or from the same search watched twice on one site; never because Google does not name you or shows no AI answer, which is where the work is. A stopped search keeps its row and its checks, so Watch all and the bin leave it off. A put-away search is not checked or counted in your cost or your list’s limit until you bring it back; you choose when to be reminded (3 months by default), and publishing a page on its subject prompts you too.

Catch-up offer

When a project first gets somewhere to publish, the Completion Hub shows the pages its evidence never got. Start them, a few a day, or say not now.

Choose how WordPress posts are filed

Each WordPress site can pick a category per post, always use your default, or let AI choose from your own categories, and each project can set its own. Every route that posts to WordPress uses it, and we never create a category.

Your AI spend, day by day, with a warning when it jumps

AI Settings now shows what your AI calls cost on your own keys for the last 30 days, by provider and model. If a provider’s day costs at least three times its usual and at least $3 more, you get a notice the next morning that names the model, says when it changed and what that model costs against the one before, lists what the calls were for, and names a cheaper model of the same provider where there is one. Nothing is switched for you. And when you save a model that costs more per million output tokens, the save says so.

Every topic now competes on your own site

Where the models think a topic belongs to someone else’s kind of page - a review, a roundup, a tool - your page is still written in the shape that wins and published as normal, and the places they named are shown as advice (“Also worth being on”). They are now told which of the pages AI cites are your competitors’ own, and a topic answered mostly by competitors’ own pages counts as one your site can take part in. If you would rather place a topic elsewhere, switch it to “Place it elsewhere” on the topic list or the answerability card: its placement becomes a task, as before.

Fairer “found without your name” figures

Your own questions now count when they do not name you. A local business is asked about its own town or area (from a Business Profile link or a LocalBusiness entity, never guessed), and the trend marks the first run that names it. A brand name that cannot be right (a common word such as “Hjem”, a page title, a web address) is flagged with the name your site uses, and changes only when you confirm. Figures across several sites show each site, and the median site, beside the plain total.

Your brands are now on your About page

Link a parent company and its brands once - in Entity Factory or in a project’s “Parent company” setting, which are now the same link - and the parent’s About page, hub and schema list its brands, and each brand’s page names its parent.

Every sub-topic now has a job: covered by its main page, written into it, or built on its own

Each sub-topic under a main topic now takes one of three routes, shown on its chip. If the main topic’s page already answers it, the chip says so and quotes the passage that does. If the main topic has no page yet, its sub-topics become sections that page must answer when it is written, and are checked once it is published. If a sub-topic needs a page of its own, it is added to your gaps and built like any other. The platform decides each one and acts, and you can change any of them from the chip. Main topics now show their family’s figures too, and “Nothing built” counts what was built for any of its sub-topics.

See where the pages we published for you rank in Google and Bing - and which are climbing

Each project now shows how many of the pages we published on your site sit at #1, in the top 5, top 10 and top 100, with the share of the total beside each count, and the same for your Google Docs, videos, Blogger posts and other pieces on their own lines. A “This week” strip names the pages that entered the top 10, the ones climbing toward it from #11-30, and any that left it, each with where it was and where it is now. Every page is listed with the search it ranks for, how many searches it shows up for, its position, where the position came from and when, and a chart shows the pages in each band day by day with a line for the top 10. Where Search Console shows a page ranking for searches you do not track yet, you can start tracking them with one click. Bing appears as its own view when your Bing Webmaster Tools is connected. The dashboard’s project cards lead with the week’s movement, the Activity report carries the whole section (screen and PDF), and the weekly email says it in one line.

Site check: challenges, missing pages, trust pages and llms.txt links

The site check now fetches your page as OAI-SearchBot too, seven AI agents in all. When your firewall answers one of those fetches with a bot challenge, it says the agent was challenged when fetched from our server, and never that AI agents are blocked: a firewall judges a request by where it comes from, and the real crawlers come from their own companies’ networks. It also asks for an address that cannot exist, which should answer 404; looks for your privacy policy, terms, contact or about page and pricing page, and whether each is readable before any JavaScript runs; follows up to 5 of your llms.txt’s links to your own pages and notes its length; and notes whether you serve llms-full.txt, ard.json and ai-catalog.json. None of it changes a score. The free check at /audit shows the same rows.

A developer page, and the free site check as an API

/developers lists both MCP servers - your account’s and the free check - how to connect, every tool, the scopes, the limits and the error shapes. The free site check can now also be run with one GET to /api/public/site-check, described by an OpenAPI 3.1 document at /openapi/site-check.json.

Was cited, now isn’t: per engine, and a swap is not a loss

A page is now listed only when the AI engine that cited it asks about its subject again and does not cite it; a check by another engine never counts. After one such check the page shows as "Not seen in the latest check", apart from the rest. After two in a row it is "Was cited, now isn’t" when that check cited no page of your site, or "Cited another of your pages instead" when it cited a different page of yours on the same subject, linked. The project page, the weekly email and the agent tools list the same pages in the same three groups, with the engine and the dates of its checks.

Up to 100 of your own products and services, and a list to manage them

A project can now describe up to 100 of its own products, services and named features, each with a line in your words. The What you do screen lists every one with a count out of 100: tick the ones you no longer want and press Remove selected. On the project page, a name that is not one of yours has “Not one of ours”: it leaves the list, is never suggested again, and nothing is written about it, with Undo for a few seconds.

A refused AI key is named, and your other engines keep measuring

When an AI provider rejects your key, refuses it access to a model, or reports your account disabled or its billing off, the box at the top of your dashboard says which provider and what to do, and you get one notice a day. A visibility run carries on with your other engines: that provider’s figures are marked missing, not zero, until its key works again, and tracked prompts carry on without it. A run whose every engine uses the refused key stops, as before.

Every earlier release on the Roadmap

What’s New now shows only this release. Every earlier release, with what changed in it, is listed on the Roadmap page under Release notes.

One check before anything we write goes public

Posts from the WordPress AI campaign, the content drip (all five channels), Cloud Stack and the Affiliate Stack’s cloud pages now pass one publishing check before they go out: a figure or a named source we were never given is cut, a “best”, “leading” or “cheaper than” about your business with no source is cut, and a post that repeats one on the same site in the last few weeks is not sent. With Review before publishing on, these posts now wait for you too, in Evidence Engine under Held for review, with Publish and Discard. If more than three sentences have to go, the post waits for you whatever the setting.

What we published for you is recognised as yours on more pages

Blog Stack posts, Entity and Affiliate Stack posts and cloud pages, topic-cluster Blogger and LinkedIn posts, SmartStack WordPress articles, Business Profile posts and posts you send by hand to a webhook, Ghost, Wix or Shopify are now recorded as ours when they go out, so an AI answer citing one never counts it as independent. A cited page on a site you connected for publishing that we cannot prove we wrote is now counted as “could not be traced”, never as independent.

Our MCP servers in an Agentic Resource Discovery catalog

Our two MCP servers are now listed in an Agentic Resource Discovery catalog at /.well-known/ard.json, with an MCP Server Card for each.

Four AI engines a run

A visibility watchlist can now ask up to four of the six engines, your choice, in Set up a run and on each saved watchlist, and “Use these engines for every watchlist” takes four as well. Every question is asked of each engine, so a fourth adds a third more answers: ten questions are 40 answers instead of 30. At our estimate of at least $0.022 an answer, that is about $0.95 a month more for a weekly run of ten, on your own key, plus that engine’s search fee; tracked prompts and gap before-readings on that watchlist ask the fourth engine too. A watchlist you already have keeps its engines until you add one, and a fourth engine starts its own line on the trend. A project newly measured weekly by default is given up to four of your engines that search the web, never one that answers from training to make up the number; a project already measured keeps the engines it was given.

Every project measured weekly, announced first

Each project with work published, or its set-up finished, is now measured once a week: ten questions buyers ask, two about the business itself and the rest about its measured topics, asked of up to three of your AI engines that search the web, on your own AI keys. Before anything is spent we email the account holder what will be asked, on which engines, and about what it will cost in money. The first run is no sooner than 24 hours after that email; if the email cannot be sent, nothing runs. The email and the project page have “Run it now” and “Don’t measure this project”. A no is remembered, and only “Measure weekly again” on the project page switches it back on. A site you already measure, or chose not to, is left as it is. It needs at least three measured topics and an AI key for an engine that searches the web, and nothing runs for a read-only account.

Remove a campaign’s files from Google: one, a chosen set, or all of them

In a campaign’s Activity tab, “Remove from Google” on a file’s row takes that one file down; “Choose files to remove” lets you tick files, or pick them by when they were made, their kind (Docs, Sheets, Slides, PDFs, images, shortcuts, other files) or the words of their title, and shows a preview with the counts first. Each file is unpublished, has its link sharing removed, is moved to Drive’s bin and is checked afterwards, and the shortcuts the campaign made to it come down with it. Posts still queued for your Business Profile or LinkedIn that link to it are cancelled. A file that something outside the campaign still uses, such as a page published from it or a draft waiting for your review, is left as it is and listed with what uses it. The preview also names any page that stays up and links to a file you remove, because that link will then open a file in Drive’s bin. The campaign keeps running and making new files - a new data sheet, if you remove every Sheet - but never re-shares, repairs or replaces the files you removed, and new files it makes will not link to them; links other files already hold to them are not edited. Folders are never unshared when you remove chosen files, because the campaign’s other files depend on them; a file in a folder that is still shared is removed like any other. For the rarer case there is “Whole campaign…” in the “Withdraw from Google” menu: every Drive file the campaign recorded is taken down the same way, the folders it made lose their link sharing before it finishes (except one that a file still in use sits in), the campaign’s name is typed to confirm, and the campaign stays as a record that never runs again. Blogger posts, Calendar events, YouTube videos, Business Profile posts and cloud pages are not removed. A file Google will not confirm is down, or whose Google account is disconnected, is listed as not removed and “Try the failed ones again” tries it again. It works after a trial has ended, and it cannot be undone from the app.

Connect Bing Webmaster Tools

Under Connections, Bing, connect the login you use for Bing Webmaster Tools. It is read only. Each project then says whether its site is verified in Bing, or how to add it (Bing can import your sites from Search Console). For pages we published on a verified site, each item in your work lanes shows Bing’s index status on its own line — “Indexed in Bing” or “Not in Bing’s index yet” — read at most once a day. A project’s searches show Bing’s figures beside Search Console’s, by Bing’s weeks, with Google, Bing and Both buttons, and Pull from Bing sits beside Pull from GSC. Upload Bing’s AI Performance export under Reports & Tracking, Microsoft AI citations, and What’s working and your work lanes show Microsoft’s own count of citations of your pages in its AI answers (Copilot, Bing’s AI summaries and partner integrations, combined). Each is Bing’s or Microsoft’s own figure: never a rate, never added to Google’s, and never part of an engine’s figure or your visibility score.

Microsoft’s count in your Activity report and your AI assistant

With Bing Webmaster Tools connected and an AI Performance export uploaded, the Activity report shows Microsoft’s own count of citations of your pages in its AI answers for whichever period you pick, beside Google’s, with the period before as a count and the pages Microsoft cited, each marked as seen cited by our own checks or not. Your AI assistant sees it on each published page in get_published_evidence. It is Microsoft’s count, never a rate, and never added to your citation counts or to Google’s.

Cloudflare under Connections

Connections now has a Cloudflare entry in the menu. It lists each of your projects: connected (with the zone it reads), not connected, or not on your plan, and opens that project’s Cloudflare card in its settings. Reading which AI crawlers fetched your pages from your own Cloudflare zone, with a read-only token you create, is part of Growth and above; if your site runs on WordPress, our plugin does the same on every plan.

The site check asks about Search Console’s AI setting

Google’s Search Console has a setting, under Settings, then Search generative AI, that decides whether AI Overviews, AI Mode and Discover’s generative features may use your pages. We know of no way to read it from outside, so the site check now asks you. Under “Can AI reach the page”, press Include, Exclude or Not sure. Your answer is kept with the project and dated. Exclude is listed to fix first, unless you meant it. Not sure says to check it yourself. The free check on the public page can only say check this yourself.

AI Visibility: each engine’s naming when the question leaves your name out

Open a run on AI Visibility and a new panel, “Named when the question did not name you”, shows for each engine how many of its answers to questions without your name named you, and how many recommended you, as counts such as “3 of 12” with their 95% range. Most questions carry your name, so the per-engine “Mentioned” figure says mostly that the question named you; this is the reading of whether an engine finds you when it is not told who you are. Each engine is counted on its own and nothing is added across engines; with three answers or fewer there is no range. It is read from the answers each run already stored, so every run shows it at no AI cost.

AI Visibility: the competitor recommended instead of you, per engine

Open a run on AI Visibility and “Recommended instead of you” lists, for each engine, the competitors you track that it recommended in an answer that did not recommend you, as counts such as “4 of 20 answers read”. Your competitors are the watchlist’s list plus any site you marked as a competitor on Sites AI cites. Answers stored before 28 Sep 2026 and not re-read were not read for competitors, so each count is out of the answers that were, and an engine with none read says so rather than “none”. Each engine is counted on its own.

Sites AI cites: the sentence Claude quoted, under the page

When Claude, asked directly on your own Anthropic key, cites one of the pages on Sites AI cites, the sentence it quoted from that page now shows under the page, with the date. Open a site to see its pages; a site with one says “a sentence quoted”. Claude through OpenRouter and the other engines hand back only the page’s address, so most pages will have none — and until a direct Anthropic key answers, none will.

Project page: was cited, now isn’t

The project page now lists your site’s pages that an AI answer cited before and that a later check of the same subject did not cite, with the date each was last cited and the date of the check that did not cite it, each linking to the page. It is the same list the weekly email names week by week, kept on the page until the page is cited again. A page no check has asked about since is never listed: that is a question nobody asked, not a loss. Nothing shows when there are none.

Bars and dates on more of your screens

All clients now draws a bar beside each client’s “named of measured”: its length is the answers read in that client’s latest run, on one scale for every client, and the dark part is the answers that named them, so a client measured on three prompts is plainly a shorter bar than one measured on fifty. Nothing is averaged across clients. Competitor gaps draws a bar for each rival: its length is how many of your tracked searches it holds, with the part where you are absent shaded. Google’s AI impressions is now a bar chart with dates, one bar a day (a week when the history is long); a day Google counted as zero is a short grey tick, and a day no upload covers is hatched as “no reading”, never a line across it. Its pages are bars too, each marked as one we published and as one our own checks have seen cited. Microsoft AI citations gets the same: a bar a day for Microsoft’s own count, with the days no Overview covers hatched as “no reading”, and the cited pages as bars with the same two marks. The Entity hub’s health trend is a dated chart with one point a validation, and the project page’s three Signals sparklines are one chart of signals a day. Hover over any of them, or tab to a bar or a point and use the arrow keys, for the date and the count.

One weekly email per site, with the proof

Every Monday each of your sites gets one email: what AI answers cited for the first time (which engine, which model, which question), what we published and did, what needs you, and what each engine said, counted out of what was asked, plus Google’s own AI count from your upload. Anything we did not measure says so, and why. A client address on a project gets a version with the proof links only. A quiet week sends nothing. It is on by default for every site, on every plan. An account made before this release gets a notice first, on a Monday, naming every address that will get the new email, site by site; the email starts the Monday after, and the old weekly email and the Monday summary stop that day. Settings → Weekly report shows when yours starts. Switch a site off there, or use the unsubscribe link in any email.

Publish your held Google files together, or a few a day

On the Evidence Engine’s list of Google Docs, Sheets and Slides waiting for your review, tick the ones you are happy with and press Publish selected or Discard selected. Each goes through the same checks as its own Publish button, and any that cannot go says why and stays in your list. Or publish them over the next few days instead of all at once: choose how many days (30 unless you say), and before anything starts you see the plan – how many a day, never more than 10 for a campaign, the last day, and today’s pieces, highest impact first. Then a few go out each day, and the list shows what has gone and what is next, with Pause, Resume and Cancel; cancelling leaves the rest waiting for you. A piece that would repeat one already out on the same subject is not published and says so, and the rest carry on. A file only counts once Google Drive has shared it, and if a schedule has to pause – your Google connection, say – you get a notification saying why. Pieces on the same subject worded differently, such as “auto local SEO” and “automated local SEO”, are shown together with the one we suggest keeping, and you decide. The list shows 20 at a time. Your AI agent can see them too: list them, approve one, or plan and start a release on your say-so.

Record the pages you got onto

Got onto a review site, a directory, a press article or a forum thread? Record it on Sites AI cites, or press "I got onto this page" on the outreach list. We check once whether the page names you, then watch AI answers for it from that day. Its citations show on their own line on What’s working, never added to the pages we published.

AI Visibility: why a line changed

The trend chart now marks the other things that can move a line besides your visibility. A diamond at the top of the chart marks a day when some of your questions were reworded, the whole question set was replaced, or the search mode changed, with one diamond a day listing everything that changed. A replaced question set starts the line again, as a change of engines does; some questions reworded, or a change of search mode, leaves the line running and only marks the day. Hover over a diamond, or move to it with the keyboard, to see what changed. The by-engine charts break and mark the same way.

Tracked prompts: the searches ChatGPT ran

Open a tracked prompt the ChatGPT app has answered and a small block under its answer strip, "Searches ChatGPT ran", lists the searches the app ran to answer it and how many of its answers ran each, as a count such as "4 of 5 answers", never a percentage. Each search says whether one of your topics already covers it; one no topic covers has an "Add as topic" button. Nothing is added unless you press it, and it goes through the same add-topic route as the project page, with its 250-topic limit. It counts only the answers that reported their searches, only tracked prompts, only the ChatGPT app, only where its switch is on, and answers vary call to call, so read it as a pointer to what to cover, not as a measure.

One frame for every panel

Every panel and card in the app now shares one frame: the same corner radius, the same header and body inset, the same gap between panels, and no drop shadows, on every signed-in page. Nothing moved and nothing was renamed; the pages now read as one surface.

AI Visibility: by engine over time

One small chart per engine across your score runs, each point that engine’s named, cited or recommended answers in one run, with its 95% range as a band. Hover or tab to a point for “12 of 25 answers”, the model and the surface; a run that did not ask an engine leaves no point, and the line breaks where the engine’s surface changed.

What’s working: each rate as a dot with its range

What’s working draws each indexed and cited rate as a dot with a whisker for its 95% range, on the same 0–100% scale in every row, so you can see which results overlap and which stand apart. The counts, the range and the sample stay written beside every dot, and rows with too few items still show counts rather than a rate.

Tracked prompts: an answer strip per prompt

Open a tracked prompt and its answers are a strip: one cell per answer by date, a row per engine, with the ChatGPT app, the Gemini app and Google AI Mode on rows of their own. Darker means a page of yours was cited, filled named, outline not named, hatched no answer — so a prompt turning from outline to filled is visible at a glance, with every reading on hover.

AI crawlers on your site: the stages as a bar, and two lines by week

The AI crawlers card on each project page draws its funnel as a bar — not fetched, refused, crawled, retrieved and cited, with the count on each — and, under it, two lines by week: the pages a crawler first fetched and the pages an AI answer first cited. Hover or tab along a week for the running counts and the pages behind that week’s step with their next action; a page nothing has checked is drawn hatched, never as a no.

What’s working opens on a timeline of citations

The report now starts with a chart: the pages we published for the site that an AI answer cited, by the day an answer first cited each — by the week over a longer history. Hover a bar, or move to it with the keyboard, for the engines and the pages behind it, with the question each page was cited for where we know it. A page counts once, on its first day; a page nothing has looked at is not a miss.

The Activity report’s tiles show the period before

Each of the five “What happened” counts — published, submitted, cited by AI, gaps closed, re-measured — now carries the count for the period of equal length immediately before the one you chose, as two bars with the difference in words, and the report names that period. Choose both a start and an end date to see it: an open-ended period has no equal period before it.

AI crawlers on your WordPress site, on every plan

Our WordPress plugin records visits from the AI crawlers on our list and shows them in “AI crawlers on your site” — no Cloudflare needed. Each visit is checked against the addresses Google, OpenAI and Perplexity publish (Verified or Unverified; Claude’s cannot be verified). It records nothing about people. If your site has a page cache, counts read “at least”. Download it and make its key under the project’s Settings → Connect WordPress.

Tell us what customers ask you

Setting up a project now has an optional box, “What do customers ask you?”. Write the questions customers ask before they buy, one per line, and each is saved as a weekly tracked prompt: asked on your AI engines, on your own keys, with every answer kept so you can see whether they name you. Within the 20 a project keeps; leave it empty and nothing changes.

Pages due for review, and a “Reviewed on” date you earn

Every 3, 6 or 12 months (6 unless you choose) a page we published for you is listed under “Due for review” on its project. We check its source links for you; you confirm they open, the figures and dates are still true, and anything that changed has been edited. Only then is it recorded as reviewed. A review never changes the page’s own date: that moves only when its words do.

Sites that refuse AI use are not quoted

Before an outside page is offered to a writer, we read its site’s robots.txt. A site that says ai-input=no, or does not let our crawler read the page, is left out and named with the reason.

How firm a rate will be, beside the prompt count

The run form now says what a rate from that many prompts can tell you. With 25 prompts, a 20% rate on one engine could really be anywhere from 9% to 39%; with 100, from 13% to 29%.

Indexed, and looked at but not cited

Beside each page’s stage in the AI crawler report: whether Google has it indexed, read through your Search Console, and whether an engine’s search returned it for an answer that did not cite it. Only OpenAI, and Claude asked directly, report that, so most pages read unknown — never no.

Google’s own AI count on What’s working and for your AI assistant

If you upload the Search Console generative AI export, What’s working now shows it for the site you pick: how often AI Overviews and AI Mode showed your pages on each day the uploads cover, and the pages Google showed, marked as ones our own checks have seen cited or not. An AI assistant connected through the agent tools sees Google’s count on each published page, with the dates of the export it came from. It is Google’s count, kept apart from the citations we measure and never added to them.

Ask Google AI Mode your tracked prompts

Tick “Also ask Google AI Mode” on the Tracked prompts card and each prompt is also asked of Google AI Mode, through DataForSEO on your own DataForSEO credit, from your project’s location, on the prompt’s own schedule. Its own line, never added to your score or AI Overview figures; a page of yours it cites is recorded as cited in Google AI Mode. About $0.0012 a question, with the monthly estimate beside the switch.

The Visibility trend is an interactive line

One point per run by date, in plain engine names, broken only when an engine is added or dropped. Hover for the answers and models; click to open the run.

AI Visibility, one project at a time

The page follows the project you pick: tiles for each way we watch AI answers, then the trend, what happened run by run, and each run’s detail.

Runs named by what started them

What happened lists each run as scheduled, by hand, by an AI agent or a quick check. Gap before-readings fold into one line a day, with no grade.

Copy to a new campaign

A campaign’s page can open New Campaign filled in from it — project, target, business, keywords, languages and accounts. You pick the recipe; the original is not changed.

Fold any large panel

Large panels on the project page, project settings, the Evidence Engine and AI Overview fold to their header. They open by default and stay as you left them.

The AI crawler report is on the project page

Which AI crawlers fetched which page now sits on the project’s overview. Settings keep only your Cloudflare zone and token, and the campaign list is on the overview too.

Withdrawn campaigns tucked away

Campaigns withdrawn from Google leave the Campaigns list; Show withdrawn brings them back. Each is kept as the record of what it published.

Models listed cheapest first, with a cost per answer

The engine picker shows each model’s estimated cost per answer from published prices, cheapest first, with the engine’s search fee where it is published. Unpriced models say so.

One click: the same engines on every watchlist

Put the engines you have chosen on all your watchlists, or the ones you tick. It lists what changes first, since a changed watchlist’s trend starts new lines.

Grok searches the web in visibility runs

Grok 4 and later models now look pages up with xAI’s web search as they answer, at most one search an answer, and what they cite is recorded. A Grok line may step where this starts.

Campaigns can aim at your Maps listing

Each campaign keeps its project, which you can change in Edit campaign. Its target can be your Google Maps or Business Profile listing, and it still sits under that project.

Your listing link on the project

Keep your Google Maps or Business Profile link beside the domain in project settings. Local Stack and GBP Stack campaigns offer it as their target.

Put your projects in your own order

Move projects earlier or later on the Projects page; every project menu and the dashboard follow that order. New projects go at the end.

Every Cloud page in your published items

Each page a campaign puts on your cloud storage — its article pages and the index page beside them — is now listed with your published items, with its address, topic and the date it was made, like a Google Doc. It shows in the Ledger, the reports and what an AI agent reads, and a citation of it is recorded like one of a Doc; a page sent for indexing says when. A page is listed from the run that makes or rewrites it, and a page rewritten on a later run stays one item. The sitemap and robots.txt are not listed.

Cloud says when it needs a storage account

Ticking Cloud pages, or choosing Cloud Stack, Local Stack, Competitor Stack or Brand Stack, does not choose a storage account. Without one, or when the one chosen has since been deleted, Edit campaign, the wizard and Recipe Studio say “Cloud needs a storage account: connect one, or Cloud is skipped”, and the campaign page says it until one is chosen, with Choose one beside it. Saving still works and Cloud stays ticked; each run that skips Cloud for this says so in Job Logs.

Translate a campaign into up to three more languages

A campaign made from now on writes in your project’s content language, or your account’s; Edit Campaign says when one made before writes in another, and offers to switch. Add up to three more languages in the wizard’s Launch step or under Edit Campaign, Languages: once a pass’s places have published, it makes the pieces they published — its article (the Drive Doc and cloud page), analysis Doc, Blogger post, calendar release, and Local and Competitor Stack’s own Docs, where the campaign publishes them — in each of them, each published where its original went as a piece of its own. Adding a language switches Translation on; switch it off under Signal types to keep the languages and translate nothing. Translations are written with your own AI key just before they are published — about 1 US cent per Doc, post or page per language at Gemini Flash prices — and keep your links and business name. They count toward the campaign’s signals per run and go out only while the run’s share has room, so they never stop a place from publishing, and a small share can leave room for none. No separate translation key is needed, so Settings no longer asks for a Google Translate key.

Recipes run on their own schedule, and as described

A campaign takes one recipe and runs on that recipe’s own schedule for its own number of days — weekly for most recipes, daily for Calendar Stack and Launch Sprint, once for Doc Stack, Doc Chain, Cloud, Site, Slide and Photo Stack — at the recipe’s own pace. It is created as a draft: press Launch on its page to start it. A recipe that runs once runs everything it switches on. A Calendar Stack campaign adds dated information releases about your services — all-day entries titled “Update: … – published” with the date, with no time or place — and a GBP Stack campaign posts updates, never Event posts. Local Stack publishes its local article as a Google Doc and its local FAQ on its cloud page, for the location you give it; Competitor Stack publishes its comparison as a Google Doc, written only from your description and the own page of a competitor you give by its web address (the wizard asks for one); and Brand Stack’s Slides deck is its brand deck, written from your description. No recipe switches on Google Sites, which no software can create: Site Stack publishes the Google Docs for Sites you build with the guide in Help. Photo Stack’s album is private, and no other recipe makes one. Every run writes new text on your own AI key. Campaigns made before keep their switches and schedule, and their next runs work as above; one on a schedule no longer uses Photos or Google Sites after its first run. Entries and posts made before are left as they are.

A guide for every recipe, and one name for each

All 17 recipes in the campaign wizard now say what a campaign made from them makes, what they need connected, what their first run uses, and what they do not do — each with its own guide in Help, linked from its card. Recipes and the tools that do the same job share one name: Doc Stack, Sheet Stack, Slide Stack, Calendar Stack, Cloud Stack, GBP Stack and Blog Stack. Every recipe is on every plan.

Your own site, first on every project

Each project page now opens with your own site, for the last 30 days: how many subjects have a page there, how many of them an AI engine was asked about, how many of those pages an answer cited, and how many were not asked yet - with Measure them, which lists the subjects named and adds any you pick to the questions your visibility runs ask. Below that, how far every page has got since it went live: not fetched yet, fetched only by AI training crawlers (which feed no search index), fetched by search crawlers such as Googlebot and Bingbot, fetched by an assistant for somebody’s question, or cited. Fetches show once your site’s Cloudflare is connected; until then the card says they are not visible rather than counting them as zero. The Docs, videos and other evidence pieces we publish elsewhere come second, with how many were cited and whether the answers that cited them already named you.

Pages AI engines cite wait for your approval

We never rewrite the text of a page an AI engine is citing until you have approved the change. When new material arrives for a cited page, when you press “Rebuild in the right shape” on one, or when your About page or its Google Doc would be written again, the new version waits for you under “A page update waits for you” on the project page, saying why it waits. Read it, then publish it in place or discard it; if you leave it, the page stays as it is. One update from new material waits per page, and a rebuild or re-align of the page replaces it. A rebuild and a re-align of the same page can both wait - the project page says so - and each replaces the whole page if you publish it, so publish the one you want. A discarded About page update is not offered again until your entity’s facts change. This holds whether or not Review before publishing is on, and pages no engine has cited yet still update by themselves. The links we keep at the foot of your pages still update as before. And once we have written your About page on or after 7 October, it is not rewritten by itself over a change made to it on your site.

One piece per subject, per place

The Evidence Engine no longer publishes a second piece on the same subject to the same place within 30 days: no second guide on a subject in Google Docs, however the subject is worded, no second page on it on your site, no second post on your Blogger blog. A different kind of piece - an FAQ beside a guide - or the same piece somewhere else, such as the page on your own site that follows a Google Doc, still goes ahead. A gap whose campaign already has the piece it asks for is closed on that piece; any other gap it would repeat is set aside with the reason and when it could be made again, and a page already written for it stays on the gap, to be sent if you reopen it then. An AI-visibility topic now sends one supporting cloud page, then none on that topic for the 30 days. Change the 30 days under Settings, to anything from a week to a year. The classic recipes’ own generators are not affected; a topic cluster’s pieces are, in any campaign.

Ask the ChatGPT and Gemini apps your tracked prompts

Switch them on for a project and each of its tracked prompts is also asked of the ChatGPT app and the Gemini app themselves, through DataForSEO on your own DataForSEO account, from your project’s location: ChatGPT from its country, Gemini from the place itself. Each app has its own line under the prompt - named and cited, out of its answers - kept apart from your engines and your score, and the prompt’s history keeps every answer with what it cited and, for the ChatGPT app, the pages it looked at and did not cite and the searches it ran. An answer that asks you a question back is counted apart. It costs about $0.0012 a question on your DataForSEO credit, outside your AI Overview budget, and the card shows each app’s monthly estimate before you switch it on. To start, add your DataForSEO login in Settings and a location under Manage Project, then use the two switches on the Tracked prompts card on the AI Visibility page. The help article “Asking the ChatGPT and Gemini apps” goes through it step by step.

Rank checks carry on when Serper runs out

If you have both a Serper key and a DataForSEO account, Serper is used first, and when its credit runs out your rank checks, Reddit thread searches and news-article lookups move to DataForSEO by themselves - so you do not need to keep topping Serper up, and nothing stops. DataForSEO checks cost about $0.60 per 1,000 and its credit does not expire. A DataForSEO account on its own now runs rank checks too. And if Serper runs out with no DataForSEO account, you are told once and then weekly, not every night.

Room for 250 topics

Each project can now keep up to 250 main topics, on every plan - up from 60 - so you can add and measure the topics you care about however long your list already is. Discovery still adds topics on its own only until a project has 60; past that, a topic joins because you added it or measured it from a gap. A longer list does not cost more: each run asks as many questions as before, so each topic is asked a little less often.

Set where your customers search

An AI Overview depends on where the search is made. Searches with no location set are now read from the United States (they were read from the United Kingdom), and the AI Overview page says so on each one, with a link to set your site’s location. And a search’s figures now count only the readings from where it is checked now: earlier readings from another place are counted apart, never mixed in.

The free site check reads more of what AI agents are told

It now reads robots rules in your server’s X-Robots-Tag header as well as the page’s meta tag — noindex, none, nosnippet and noarchive among them — and says which AI agents were sent a header; a rule aimed only at an AI crawler is noted separately, never as a search-engine block. It also shows how quickly the page answered each AI agent, the first words GPTBot read, whether http:// and www lead to one https:// address, heading order, links to sources, alt text, the page’s language, the names it gives your business and, only for a page that shows a street address with a phone link, map or opening hours, LocalBusiness markup. Every point the SEO score takes off now has a line in Everything we checked saying how many points it took.

AI Overview checks go out only from real places

AI Overview checks are now sent only from places on DataForSEO’s own list. As you type a location we check it on your own DataForSEO key: where it fits more than one place you can pick the one you mean (your wording is kept), postcodes are flagged as not used, and the AI Overview page says where each reading was taken from. One refused search no longer stops the rest.

Two more answers for a site AI cites

“Not related” is for a different business or subject with a similar name, cited on a question that named you. It leaves the naming rates and the outreach list, as your own sites do. “Neighbouring field” is for a business that sells to your customers without selling what you sell. It gets a rate of its own instead of blurring your competitors’ or your publishers’.

The robots.txt we give you lets Gemini in, and our checks read the same list

Google documents Google-Extended as the switch for whether Gemini may use your pages to ground its answers as well as to train, so the robots.txt in your fixes now allows it. The file, the GEO audit and the site check read one list: a blocked answer or search crawler (OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot, Perplexity-User, Google-Extended, Googlebot, Bingbot) is a Blocker; a blocked GPTBot or ClaudeBot is reported as a choice, because it does not stop ChatGPT or Claude citing you; other companies’ training crawlers are not marked either way. OAI-SearchBot, which decides whether you appear in ChatGPT search, is now read. If you installed the earlier file, swap in the new one.

Google’s own AI count, charted in your Activity report

If you upload the Search Console generative AI export, the Activity report now uses it. For whichever period you pick you get a bar for each day Google’s AI Overviews and AI Mode showed your pages, the total, and the period before for comparison. A day no upload covers is drawn as “no reading”, never as zero. Under the chart are the pages Google showed, each marked as one our own checks have seen cited or not - so you can see the pages AI is using that our checks never came across - and as one we published or not. The same report now also shows your Search Console clicks and impressions against the period before, and your AI Overview watches; both were only in the PDF. Nothing here is added to the scores we measure by asking engines: it is Google’s count, shown beside ours.

You approve the text you read, not whatever is waiting

An update to a live page waits for you as one item per page, and when more material arrives it replaces the text that was waiting. Publish now carries the version of the text you were shown: if the text has changed since, nothing is published, and the card says so and shows you the text that waits now. The same goes for an AI assistant with write access: approve_publish must hand back the version list_waiting gave it. Separately, text we take from other sites for a writer to quote now arrives marked as quoted material and not as instruction, so a page written to give orders to an AI is read as a page.

Looked at, not cited

Asked directly on your own OpenAI or Anthropic key (not through OpenRouter), OpenAI and Claude say which pages their web search returned for an answer, not only which ones the answer cited. A visibility run now keeps both: under each answer you see the sites the search returned and the answer did not cite, and above the answers how many of them passed over a page of your own. That is a different problem from not being found - the search reached your page and the answer chose other sources. It shows the search returned the page, not that the model read it. Gemini and Perplexity do not separate the two, so their answers are not counted, and runs made before today have no list. Also from today, an OpenAI answer that searched the web and cited nothing counts as a live search answer.

Fixed your site? Re-score the run without spending AI credit

When a visibility run points at something on your site - no llms.txt, a robots.txt that leaves AI crawlers out, no structured data - you no longer need a new run to see the fix counted. On your latest run, press “Re-read my site and re-score”: we read the site again and work the score out again with that run’s own answers. No AI is asked, so nothing is spent, and only the site’s 60% of the score can move. The run then says when it was re-scored and what it scored on the day. Earlier runs keep the score of their day.

IndexNow: every site has its own key, and you can use yours

Each site now has an IndexNow key of its own - made for it, or the one your SEO plugin already serves, which you can paste in. The card is in each project’s settings as well as on the WordPress page, so a site you publish to by webhook has one too. “Check now” fetches the key file from your site and then asks IndexNow itself, so the card shows what IndexNow answered - working, or refused and why. Google Docs, Sheets, Slides and Blogger posts are not sent: a key file cannot be placed on Google’s hosts. A WordPress site set up before today keeps the key it had until you press “Give this site its own key”.

Competitor gaps opens on the pages worth writing

The list of businesses competing with you now starts closed, naming the three taking the most of your searches, so the searches worth writing a page for are on your first screen. Click its heading to open it: it shows the first twenty, the rest on request, and stays open next time if you leave it open.

Quicker overnight checks

The nightly checks on your Google files, and the stages that build on them, now read only what they use - each file's address and health - rather than every stored document in full, and save only what they change. The app stays quick while they run.

Track the exact questions you care about

The visibility score is a composite of many questions and your site, so it cannot show what happened to one question. Tracked prompts can: up to 20 per project, each asked on your visibility engines daily, weekly or monthly, with every answer kept - named or not, where, cited or not, and what it cited. Add them on the AI Visibility page, or press Track question on any topic. Tracked answers stay out of the score, and one search finds a topic across your runs and your tracked prompts together.

Sources are quoted from the article itself

Every writer is shown what its strongest sources say about the subject. Source pages are now read in paragraphs with the menus, header and footer left out, so the passage a writer sees is the article’s own words on your subject.

Your definitions now hold for the whole page

Where you have told us what something is, every writer is now told that your line holds for every section of the page, not only its opening, and that a source using the same words for something else is about something else - it gives the page no definition, comparison or example. And a note from the writer about which item of its brief a section covers ("this section satisfies item 1") is now cut before a page goes out.

Your visibility score, explained

The score is a composite: 40% how the AI answers treated you, 60% how readable your site is to AI. "How we measure" now works one through step by step with a made-up business, including why it can move when only the questions changed. With it, 20 points of the answer score now go to being named when the question leaves your name out, the "not surfacing in comparison prompts" warning is retired, and an old run shows the score it had on the day.

Competitor gaps can sort themselves - when you press

Pick a site on Competitor gaps and press "Sort these for me". Searches won by a video or a forum thread are marked to leave; then your own AI is asked once to mark the rest "Take it", "Fold under" one of your main topics, or "Same as" another search, each with its reason. It only runs when you press it, it never judges whether you offer a subject, and it adds nothing on its own. Each row now also shows its buying stage, and links your own page on it where you already have one.

Your topics change only when you say so

Two things used to happen on their own: a discovered topic asked about six times without ever being named was retired, and when the measured list was full an idle topic had its measuring switched off to make room for one we had just published for. Both are now suggestions. A small "suggestions" chip beside your topics opens them, each with its reason, to accept or keep as it is - and one you keep is not suggested again.

Competitor gaps: see what beats you, and measure what you add

Each search on Competitor gaps now links the rival page that is winning and says what kind of page it looks like - a list, a video, a forum thread - so you can see what you would have to beat. Adding a search now also makes it a measured topic by default, so your AI visibility runs ask about it before the page is written as well as after (untick the box above the table if you only want it built). "Not for me" takes a search off the list for good, and you can bring it back from the folded list underneath. The page is linked from each project's topics, and the 60-topic limit now counts main topics everywhere - sub-topics never take a slot.

Settled pages leave the answerability list

A page you chose to leave as it is, or one that was rewritten and graded again, now stays in the list for two days so you can see the outcome, then moves under "Done" at the foot of it with what was decided and when. A page rewritten by Re-align or a page update now shows "awaiting re-grade" instead of its old grades.

Your own facts on your entity page

Facts you keep on an entity under names of your own - your prices, what a plan includes, what you connect to - now appear in the key facts of its page on your site, in its Google Doc and in its llms.txt. The founding year the website scan finds now reaches the page as well; it was stored but never shown.

Gaps waiting for a campaign now find it

When several campaigns share a site and a gap’s topic is in none of their keywords or target pages - or no campaign there takes the Evidence Engine’s work - the gap waits in no campaign. The screens said that giving a campaign a keyword for the topic would move it there, but nothing did. Now every run of the engine does: each waiting gap goes to the campaign on its site whose keywords or target page name its topic, and until then its card says what would move it. The engine still never picks between campaigns the topic does not tell apart, and never gives work to a campaign kept out.

The free check inside Claude and ChatGPT

A second MCP address, https://therankingfactory.com/api/mcp/check, needs no account and no sign-in: its one tool runs the free site check on any address and answers with the findings, most urgent first, and a link to the full report. Both MCP servers now speak the current protocol revision as well as the older one, and every tool has a plain title and says whether it reaches outside your account.

Keep a campaign out of the Evidence Engine’s work

A campaign’s SmartStack settings now have a switch, off by default, that keeps the Evidence Engine’s gaps out of it - for a one-off or a test that should do only its own work. The engine then gives it no gaps and stops looking for new ones under it. Its open gaps move to another campaign on the site that covers them, and the switch says how many moved and how many were left unassigned.

The free check answers first

The free AI visibility check at /audit now shows its findings as soon as someone types a web address - no email needed. The full report, with what AI engines say about the site, still goes only to an address confirmed with a six-digit code, and follow-up emails go only to people who tick a box for them, each with a one-click unsubscribe. Starting a trial is simpler too: the company name is optional and the page says exactly what is charged, and when.

Sheets and Slides only where AI cites them

A standalone Google Sheet or Slides deck is now made for a subject only where AI answers, Google’s AI Overview or search results were seen citing that kind of file for it. A topic cluster still gets its Guide, nothing is spent on the Sheet or deck it skips, and a gap that asked for one is closed with the reason. A subject with fewer than 10 pages seen is made as before, and Google Docs are unaffected. To make them for every subject as before, tick “Make them for every subject” in the project’s settings.

Where to earn a place: pages that sit beside recommendations

Sites AI cites now ranks other people’s pages by how many of the answers that cited them recommended someone - you, or a competitor you track - rather than by how often they were cited, so a “best tools” roundup that sits beside recommendations comes before a definition cited in answers that recommend nobody. Each page shows which competitors were recommended beside it and how to approach it, and the whole list downloads as a spreadsheet. Your own sites, your competitors’ sites and the pages we published for you are left out. We contact nobody for you, and never suggest paying for a place or swapping links.

Pages take the shape AI answers cite for the topic

Before a page is written, your AI providers now see what AI answers and Google’s AI Overview cite for the topic today. When most of it is comparisons and best-of lists, the page is an honest comparison - competitors named fairly, any fact about one from their own site and dated; when it is definitions, a glossary. When it is free tools, the answer is your own tool or product page, so we hold the article for you with the reason instead of publishing it.

No second page for a subject you already have

A subject worded differently - "will SEO be automated" beside "can SEO be automated" - now counts as the subject your site already has a page for, so we add to that page instead of writing a rival to it. Your AI assistant can also see which of your pages share a subject and which to keep, and which pages were cited before but not by the runs that asked since.

Pages are filed where your site keeps that kind of page

A page we publish to your WordPress site now goes under one of your existing categories - Comparisons, Glossary, Guides and the like, or the one for its subject - instead of Uncategorized. Ghost posts get one of your existing tags, Wix posts one of your categories, and webhooks are told the section and category. We never create a category on your site; where nothing matches, the page goes where it always did, and your AI assistant can see where each one went.

Your pages link to the pages on their subject

The related-pages list we keep on the pages we placed on your WordPress site used to link the eight newest pages, whatever they were about. Each page now links to the page for its parent topic first, then the pages on its subject, then the newest - and a page an AI answer has cited, or an assistant has fetched, links your newest related pages first, so crawlers find them sooner. Your AI assistant can see how many of your pages link to each one.

Are the subjects you publish on being measured?

Your AI assistant’s get_topic_visibility now says, per site, how many of the subjects you published on a visibility run asked about in the last 30 days - and names the ones it did not, so a subject your watchlist is full or missing a topic for shows up by name instead of going unmeasured. It adds no prompts and costs nothing.

Every page you published, and where it is stuck

With your Cloudflare zone connected, each page we published on your site now shows one stage: not fetched, refused, crawled, retrieved by an assistant for someone’s question, or cited - counted over every night we hold, not only the last week - with the next step that stage calls for. "Refused" comes from the status your edge returned to the crawler, read from today; earlier nights say nothing either way. Your AI assistant sees the same through get_published_evidence.

“Recommended” is read more accurately

An answer that recommends you is now read sentence by sentence properly: a line break, a bullet, a semicolon or the citation numbers some engines write after a full stop all end a sentence, so a recommendation in one line is no longer cancelled by a “not” in the next paragraph. “Best for”, “the strongest option” and “I’d start with” now count too. The same rule reads your competitors. Your recommended figures from this run on are read this way, which is part of why the trend lines start again.

IndexNow: what is announced, said exactly

IndexNow tells Bing - and so ChatGPT search and Copilot - that a page is new. It covers pages we publish to your WordPress site once the key file is in place, and the Google and cloud pages campaigns create. Shopify, Wix, Ghost and webhook sites cannot serve the key file at the root of your domain, so their pages are not announced, and our pages now say so rather than implying every page is. The “Submitted” figure on the Activity report counts exactly these. Since 2 October 2026: each site has its own key, a site you publish to by webhook is announced once it serves that key, and Google pages are no longer sent.

Each visibility run now covers four topics, not one

A ten-prompt run used to spend eight prompts on fixed questions about your brand ("What is …?", "What is … known for?") and only two on your topics - one topic a run. Your topics now come first, each asked once without your name and once with it, and two brand questions follow, a different pair each week so all eight come round every four weeks. Because the questions changed, your trend lines start again from this run.

The weekly report reads properly

Problems now show their actual message instead of a version with every number replaced by “…”, the category is no longer printed twice, the work the engine did is no longer listed as problems we “didn’t repair”, and the coming actions are written in words rather than internal names. A fuller report - one email per site, with the answers that cited you and links into the app - is being built. (Since 11 October 2026 it is here - see the entry at the top.)

Which competitor AI recommended instead of you

Every visibility answer is now read for the competitors on your watchlist that it names, and whether it recommends each - by the same rule that decides whether it recommended you. The competitors are the ones on your watchlist plus any site you have marked as a competitor on Sites AI cites, and each says which. Your AI assistant sees it per engine: how often each competitor was named and recommended, and how often it was recommended in an answer that did not recommend you, each out of the answers read. Earlier answers are read from what was stored, at no AI cost.

Your schema no longer calls a Google Doc your business

sameAs tells an engine that an address IS your business, so it now holds only your profiles and listings elsewhere - a company LinkedIn page, X, Crunchbase, G2, Wikidata. Google Docs, Sheets, Blogger posts and single videos are pages about you and are listed as subjectOf instead; a person’s own LinkedIn profile goes on that person, not on the business. This applies to your entity page, its Doc and Sheet, the entity schema and the Organisation Schema from the Entity Stack. If you pasted that schema into your site before today, copy it again.

Page addresses keep accented letters

A title such as “SEO para fábricas: la guía práctica” used to become /seo-para-f-bricas-la-gu-a-pr-ctica - every letter outside a-z was dropped. New pages now get /seo-para-fabricas-la-guia-practica. Pages already published keep their addresses and are still found and updated under them.

A video is never credited with a citation from before it went up

Google’s AI Overview sometimes cites a YouTube video by its title alone, and a title can belong to an older video of yours. A title seen before a video was uploaded is no longer counted as a citation of it, and the few that had been are corrected by the next citation sweep.

Found unprompted, counted apart

Most visibility prompts carry your name in the question, so “named” mostly shows that the question named you. Your AI assistant now also sees, per engine, the prompts that never name you and how many of those named or recommended you, and, per topic, the runs in which a question without your name was answered with it.

Publish to your Shopify store

Connections → Publishing now takes a Shopify store. Create a custom app in your Shopify admin with the read_content and write_content scopes, paste its Admin API access token and your store’s myshopify.com address, then load your blogs and pick one. Pages go out as articles on that blog, credited to your business, with the same machine-readable provenance as any page on your own site. Publishing a page again updates the same article rather than making a copy, and an article of your own with the same handle is never touched. Your token is stored encrypted and never shown back to the browser. Articles go live straight away, as on Ghost and Wix, unless you review before publishing. Tags are not sent yet.

Each gap now says whether AI has cited what we published for it

On the project page, a gap whose evidence was published now reads “Cited in 2 of 9 runs checked since publishing”, with the engines that cited it. Nine means every completed run since publishing, up to the 30 most recent, not only the runs that asked that gap’s question. A gap no run has checked since says so rather than showing 0, and held or withdrawn evidence never counts. It shows an association, not proof that the page caused the citation. Your AI assistant sees the same figure through get_gaps.

Webhook posts that fail on a busy server are tried again

A post queued on the Drip Schedule that times out, cannot connect, or gets a 408, 429 or 5xx is now sent again, five attempts in all, after about 1 minute, 5 minutes, 30 minutes and 2 hours. If your server asks us to wait longer we do, up to 6 hours. Every retry carries the same webhook-id and externalId, so a receiver that already de-duplicates will not make a copy. Any other 4xx is taken as your site’s answer and not resent. The Drip Schedule shows “Retrying (attempt 2 of 5, next at …)” or “Gave up after 5 attempts: HTTP 503”.

Two optional readings on your own AI key, asked when you set up a project

Before-readings for new gaps take one small visibility check on a gap’s subject before anything is published for it, so a later reading shows a real before and after, each with its sample: up to 6 calls with 3 engines, at most 5 readings per project a day. The passage check has your cheapest engine read each section of a page for your own site on its own, one call per section and at most 12 a page, and flags any that do not stand alone without rewriting them or holding the page back. Both are off unless you turn them on in Settings, where the calls are stated beside each switch. New projects ask Turn on or Pass for each; existing accounts stay off.

The site check now checks the on-page basics

Title length (30–60 characters), description length (100–160), exactly one H1, a viewport that fits a phone, a canonical tag, Open Graph, and profile links a visitor can see. Each finding quotes what it read, and they cost points in the SEO score, so a score saved before this is not comparable with one taken now. A new group at the end of the checklist lists what the check does not measure — Core Web Vitals, how the page behaves on a phone, and links from other sites — as “Not measured here”, never as a pass.

Your Wikidata work is kept on your account

The Wikidata Q-number you tell us, the QuickStatements batch we last drafted for you, and whether you have submitted it are now saved per project on your account, so they are there on any computer. Press “I’ve submitted this batch” once you have run it (there is Undo). Anything your browser remembered moves across the first time you open the project. You still submit the batch yourself; we do not create or maintain the Wikidata item.

Blogger posts and cloud pages carry provenance too

The machine-readable record of when a piece was drafted and published, and by whom, now goes on the Blogger posts and hosted cloud pages we publish as well as on pages for your own site. Most carry the dates and your business as publisher but no source list, because they are not written from sources we fetched; case-study cloud pages list theirs.

Sections written to be quoted on their own

AI answers lift passages, not pages, so every writer is now told that each section must name its own subject rather than lean on “this” or “as above”. Reference tables and calculators published as Sheets open with a one-line summary under the title, and it is left out if it states a figure nothing supports. No extra AI calls.

Your AI assistant can read the sentence Claude quoted

When Claude is asked directly on your own Anthropic key and cites one of your pages, the exact sentence it quoted is kept with the citation. list_cited_sources returns the newest per cited site (quotedText, with quotedUrl naming the page). From Claude through OpenRouter and the other engines we keep only the page’s address, so it is empty for them. It is not on screen yet.

The robots.txt we give you, updated

The robots.txt in your fixes allows every crawler that fetches pages to answer a question, now including Claude-SearchBot, Claude-User and Perplexity-User. OpenAI’s and Anthropic’s training crawlers, GPTBot and ClaudeBot, stay allowed so their next models can learn about you, and sit under their own training heading. Other companies’ training-only crawlers are still blocked by default. The file now carries a Content-Signal line allowing search and AI answers, and names no retired agents. If you installed the old file, swap in the new one.

Gap results say how many answers they are based on

A closed gap’s note used to read “mentioned 25%” without saying out of how many. It now reads “mentioned in 3 of 12 answers (25%), cited in 1 of 12”, so a figure from four answers is not mistaken for one from four hundred.

Review Google Docs, Sheets and Slides before they go out — per project

In a project’s Settings, tick "Hold Google Docs, Sheets and Slides for my review". Each one is then written into your own Google Drive but not shared or counted until you look at it: open it from "Waiting for your review" on the Evidence Engine, edit it if you like, then Publish or Discard. A topic cluster built on a held document waits with it, so nothing public points at something you have not approved.

Google documents: checked, and spaced out

A Doc, Sheet or deck that only describes a subject instead of saying something about it is no longer written — the same check your own pages already had. And Google documents now go out at most one batch a day per project, rather than on every cycle.

Every project shows its own latest visibility run

On an account measuring several sites, Entity Health could say "No completed AI-visibility run yet" for a site that had one, because it only looked through the account’s ten newest runs. It now asks for each site’s own runs — and topic discovery, which had the same blind spot, learns from them again.

Help with pictures

The guided tour and the main help articles now show the screens they describe — work lanes, the dashboard, the Evidence Engine, publishing connections and the visibility audit.

Connect your AI assistant on any plan, Starter included

Agent access now comes with every plan, Starter included, and with the trial: connect Claude or any AI client that speaks MCP by signing in, or with a key from Settings → Agent access. It reads your measurements, gaps and published evidence, and makes changes only if you tick “Let it make changes too”.

Let your AI assistant make changes too - only if you say so

An AI client connected over MCP can now act, not only read - but only if you tick “Let it make changes too” when you connect it or create its key; it starts unticked, and every connection made before stays read-only. With that permission it can update your business’s entity (description, main topic, location, topics and profile links), tell us what one of your products is, work a gap now, approve a page or rewrite that is waiting for you, and start a visibility run - each through the same buttons the app uses, with the same checks. A connection given one project can change nothing else. Every attempt, allowed or refused, is listed under Settings → Agent access → What agents changed, and each connection is limited to 30 changes an hour.

Your business’s own page, on your own site, in the form AI engines read

Open your entity’s hub and use the new Canonical page card to publish one page that states what your business is - in plain sentences and as dated schema.org structured data - to your own WordPress, Wix or Ghost site, or through your webhook. It keeps one address, it is rewritten whenever the entity changes, and the Google Doc and Sheet now point to it as the authoritative version. The card shows where the page is, when it was last written, and whether it has fallen behind. You can also download the same facts as an llms.txt for your site’s root.

The site check now asks as ChatGPT, Claude and Perplexity fetching a page for a user

Besides the crawlers that build AI indexes, the site check now fetches your homepage the way ChatGPT, Claude and Perplexity do when they open a page live to answer somebody. If your site turns them away, you get a finding that says so - including when a Cloudflare setting for newly added domains may have switched the block on for you, and where to turn it off.

Webhooks are signed, and pages can be updated in place on your own site

Every message we send to a webhook is now signed the standard way, says whether it is a new post or an update, and carries our id for the post so a resend never makes a copy. Tick “Accepts updates” on a webhook connection once your receiver handles them, and rewritten pages replace the old ones in place. Where an update could not be delivered, Send again sits beside Copy the HTML. The exact format, with code to check the signature, is on the Help page “Webhook format - for your developer”.

Before we ask you what a product is, we look it up

When a topic names one of your products that nothing of yours explains yet, your web-searching AI engines now look it up first, in the context of your own site. An answer found on your own site is saved for you; anything else is offered as a suggestion for you to check and save. A name nobody could find is looked up again once a week.

Visibility runs on OpenRouter models now use live web search

Any model you run through OpenRouter now searches the web when it answers, and the sources it used are recorded as citations. Because live search and a model’s own memory are different measurements, runs on OpenRouter models start a new series from this week rather than being compared with the ones before.

One line now clears every page asking about the same thing

On the Re-align card, the line you give for a product now clears every page waiting for it, not only the one you typed it on.

Connect Claude by signing in - and give it one client, or all of them

Claude and other AI clients that support MCP sign-in connect to your account with no key to copy: add a custom connector with your account’s MCP address, sign in here, and choose what it may read - every project, or one. A connection given one project reads that project and nothing else, so you can connect on a client’s behalf without their Claude seeing your other clients. Your connections are listed in Settings → Agent access, where you can end any of them at once. Keys still work as before, on every plan with agent access.

Pages now pick up the Docs, Sheets and decks that arrive after them

A page on your site is put together from the Docs, Sheets and decks built for its subject. Until now that happened once, so anything that arrived later - often the deck - never reached the page. Now a late part is added and the live page is updated in place, with no new AI writing: the words are the ones those parts already had. It is never done over your own changes: if the page on your site no longer says what we published, or has text of yours added to it, it is left alone and the project’s activity says why. On a WordPress site we read the whole post; on any other site we can only tell your text from your theme’s inside the article, so if you add to our pages below their last paragraph there, turn on Review before publishing. And if you review before publishing, the update waits for you on the project page, with Publish in place and Discard, and you are told about it once.

A health note for the services your account runs on

A small note beside "Since you were last here" shows whether the services your account runs on are healthy: green when they are, red when one needs you and says which, and grey - never green - when the check itself could not be reached. Your AI keys are now checked with the providers directly, so a key that has been revoked or disabled is caught before anything tries to use it — rather than after a run has already failed. DataForSEO is asked for its balance each time you open the dashboard, which is a free call, and you are warned while there is still credit left rather than once it has run out. It also tells you how sure it is: most AI providers publish no balance anybody can ask for, so for those the only honest signal is what they told us when we last called, and the note says so rather than showing a tick it cannot justify.

Writers are now shown what their sources actually say

We already found the pages that AI answers and search results cite on your subject and checked every address still loads. Now the part of the three strongest that is about your subject goes in front of the writer, so a piece reports what a source says rather than what a model remembers about it. Every other source is still offered by name and address. A sentence that reproduces a source instead of reporting it is cut before publishing: a quotation is fine, the same words with the quote marks taken off are not. This is how everything is written now — there is nothing to switch on.

Google's own AI impressions, uploaded weekly

Google counts how often its AI surfaces showed your pages, and reports it in Search Console — but exposes it through no API at all, so it cannot be fetched. Reports & Tracking → Google's AI impressions takes the export: Performance → Generative AI → Export, then upload the .zip whole. It is the one number here that Google reports rather than we measure, so it is never blended with what we learn by asking engines questions. It carries no search terms, so it ties to a page and never to a prompt or a topic — and most days on a real site are zero, which is a reading rather than a gap.

Every site AI cites, in one screen you can filter and sort

The dashboard counts the sites AI cited across your whole account, but the only place you could label one showed a single project at a time — so labelling everything visible barely moved the number. Reports & Tracking → Sites AI cites is that place: every cited source across every project, searchable, filterable by kind and project, with a toggle for the ones still waiting on you. "Ours" is now a label too, so your own other sites stop being counted as third parties and inflating the stockist naming rate.

One search is one row on Competitor gaps

A search four rivals held used to fill four rows, each offering its own "Add to the mix" for the same single page. It is one row now, naming the leading rival and how many others hold it. Above the table, a summary says how many searches are already published, being worked, or still need a decision — and each one filters the list, so the question "what needs me" is that table with most of it taken away.

The Indexed column said "cited"

On What's working, a row could read "1 cited" in the Indexed column and "0 of 7 cited in an AI answer" beside it, which looked like a contradiction and was not: one component rendered both columns and announced the same word for each. The Indexed column says indexed now. No number changed — the figures underneath were right the whole time.

Review before publishing now covers every page the engine writes for your site

With the switch on (Settings → Review before publishing), a page the engine writes to close a gap is finished and held on that gap under Needs you until you press Close it now, and you are told when one is waiting. Turn the switch off and anything held goes out on the next hourly check. Nothing is written twice while a page waits. The switch stays off by default — we publish and report — and this switch does not hold the Docs, Sheets and decks that go to your Google Drive. (Since 25 September each project can hold those too — see the entry at the top.)

The shape of every page is decided before it is written

Before anything is written for a topic, two or three of your own AI providers say what kind of page the question needs and which things it must contain — once per topic, on your credit. The page is written to that shape and checked for each element before it goes out. This now runs on every project. A topic reached before its shape is decided is asked about and written on the next run, and one your providers cannot agree on is written as a guide.

Answerability grading is now a setting

The nightly grade — two or three of your providers asked whether each page we published answers its question — runs on your own AI credit, so it runs only where you switch it on: Settings → Answerability grading. The card says when it is off. The free nightly read that confirms a verified passage is still on the page runs for everyone, and the shape decision above is the check that comes first.

Writers work only from your own material, so nothing about your products is invented

The crawl now follows the links on your home page as well as your sitemap, and keeps each page’s videos, tags and text. Every writer is given your own pages and videos on the subject it is writing about, and may say what your business does only from them, from your line, or from a source it has read. A topic that names one of your products which nothing of yours explains is not written until you say what it is — a line, a link to the page or video that shows it, or both — and the page you point at is read in full. Figures nobody supplied, table rows that cannot be supported and sentences that only send the reader elsewhere are cut before publishing. Anything published before these rules can be withdrawn: never assembled, counted or reused again, and its Drive files moved to the bin.

Re-align brings a published page into line with what you say a product is

On the project page, Re-align lists the pages we published that name one of your products. Give each product a line, and a link to its page or video if you have one; press Re-align and the page is rewritten so that what it says about the product agrees with your line and your own pages, keeping its structure, with any sentence that defined the product without your say-so named and dropped. Nothing replaces the live page until you approve the rewrite. On WordPress, a post you set back to draft is updated and stays a draft; you publish it there.

A small note on the dashboard says what needs you and what is new

Beside "Add a site" there is now a note with two counts: what needs you, and what is new since you last looked. Press it and the lines open over the page — a post to confirm, a product to explain, a Google account to connect, pages cited for the first time — each linking to where it is done. The cross, Escape or a click elsewhere closes it; "Got it" clears the new items. It never opens by itself and nothing on the page moves.

We check with you before a post on your site counts as your evidence

Your own posts count as evidence of what your business does; ours never do. A post the crawl cannot place — no record of ours, none of our markup — is listed on the project page with its own title and opening words, and you mark it yours or not, once. Until you answer it counts as yours, so what our writers say about your products may rest on it. A page that is really your front page under another address is never listed.

Drive checks say once when a file is gone, and a ping is one record per address

A file in Drive’s bin, or one Drive says no longer exists, is reported once, as one line, and never checked, shared or pinged again. A repair says which of three things happened — deleted, in the bin, or still there with sharing we could not restore. A ping is one record per address, refreshed each time it is sent. The weekly report counts real problems, not repeats.

Work you have finished stops counting as work waiting

A page you have already re-aligned is marked done and folded away instead of being counted among the pages ready to re-align — and you can still rewrite it from there when your description of the product changes. A topic that belongs somewhere we cannot publish is asked about once, naming every venue suggested, rather than once per venue. A row the engine gave up on can now be marked done or not doing, not only tried again; a rewrite held for you to read leads the answerability card with Read, Copy and Discard; and the dashboard note’s links stay on their card while the page finishes loading. What is left under Needs you is the work itself.

Every visibility run re-checks the topics we worked on, then rotates through the rest

A run asks a few questions about your brand and then as many of your topics as its prompt count allows. It now starts with the topics we published for and have not re-checked yet, so the work is measured rather than assumed, and then takes the topics asked about longest ago, so none sits at the bottom of the list forever. The number of questions and what they cost on your key are unchanged. Because two runs can now ask about different topics, the trend panel, the Momentum figure and the client report say so where they compare two runs.

The weekly email now leads with published, crawled, indexed, cited

It opens with those four, per site, from the same rows the project cards read, with the week’s strongest evidence as the headline: a citation beats indexing, indexing beats crawling, crawling beats publishing. The health line names what kind of thing needs you, or says nothing does. Withdrawn objects count for nothing anywhere.

Citations we could never have counted are now counted — and your account has been caught up

We publish to Google Docs, Sheets, Slides, Blogger and your entity hub, and then measured citations with a test that could not see any of them. Three separate faults. We compared the exact address we stored against the engine’s wording, and no engine writes a Doc link the way Google hands it back. We only ever read the body of an answer, never the source list an engine declares beside it. And Gemini names every source as a temporary Google redirect carrying only the bare site, so a post of yours read as "blogspot.com" and could never be traced back to the post. All three are fixed: Google files are recognised by their file id however the link is written, declared sources are read alongside the text, and Gemini’s links are followed once to the page they actually name. Google’s AI Overview citing a page we published for you is counted too, and shown as its own state rather than "does not name you". None of this is a new measurement — it re-reads answers you already paid for. Every account is swept nightly at 01:30 so the backlog clears on its own, and the direction is upward. An over-count would be exactly as dishonest as the under-count it replaces, so the catch-up uses the live measurement’s own matching rather than a looser rule written for a backfill.

Your AI Overview readings are now taken from your own country

An AI Overview depends on where the search is made. Ours passed no location at all, so every reading for every account was taken from the United Kingdom — which for a business outside the UK is not a weaker reading, it is a reading of somebody else’s search, and nothing on screen said so. Readings now use the campaign’s location, then the project’s, and only then the old default, kept deliberately so your existing history stays comparable rather than moving under you. Every reading records the vantage it was actually taken from. We will not guess one: a business type or a domain ending is not a place, and where two projects disagree we ask rather than choose.

Every file we create in your Drive is public from the moment it is made

Some of what the engine built for you was left private. A hub document, a data sheet and a deck were created for every keyword cluster and none of them was ever shared or published — so a stranger following a link got a request-access screen, and you got the email asking for it. That is worse than a broken link: a file nobody can open is a file nobody can cite, which is the entire reason we published it. Both create paths now share and publish, and the rule is enforced rather than remembered — a build-time check reads our own source and fails if any code path creates a Doc, Sheet or Slides without also sharing it, so the next one somebody writes cannot quietly repeat this. We also stopped recording a share that did not happen: where the account that owns a file is not the one being asked to share it, Drive refuses, and that refusal is now written down instead of being logged as a success. The existing files are being repaired in batches, asking Drive about each one rather than trusting our own records — believing those records is what let this run for months.

Whether an answer is still open, or somebody already holds it

"Topics with little to no competition" was the one thing we said that nothing measured. It does now, from readings you already have: where no source recurs across checks, the answer is still open; where one or two recur, it is settling; where three or more recur, an engine has an answer it returns to. Under three readings it says nothing at all rather than calling something open — that is the one way this feature could push somebody into publishing on a guess. Where an answer is held, you get the sites holding it, with how many checks out of how many, and two moves offered without either being chosen: drop the topic, or write against the sources that hold it. It never claims a rewrite will work. "Publishing more of the same is the one approach these readings show has not moved it" is a statement about what has already happened, which is the strongest thing the data supports. The Evidence Engine now works open answers first and held ones last.

Topics that are really one topic can now be grouped

Two phrasings of the same question were two topics — each measured separately against your topic limit, each built for separately. The project page now asks "Which of these are one topic?" using data rather than a model: the pages Google ranks for each phrase, and the sources its overview draws on. Where those overlap enough the proposal shows its working, because the shared pages are the reason, and says how many slots accepting it frees. One press sweeps the whole list instead of proposing one at a time, capped at twelve — a review nobody finishes is a review that did not happen — and whatever it could not reach is named rather than dropped. Children are kept, never merged: a child keeps its measurement and its history, stops counting against your limit, shows as a chip inside its parent, and can be detached again. The engine now works a cluster in one go, parent then children, instead of leaving a child three cycles behind.

What kind of page the overview reached for, not just which sites

A source list tells you which sites Google’s overview drew on. It could not tell you what shape of page it wanted, which is the half that decides what to write next. Each watched search now reports the formats it keeps citing — video, listicle, article, product page, documentation, community — counted per reading with the denominator, so three videos in one reading is one "wants a video". It is rules rather than a model, and "not classified" is a real answer instead of a guess. Measured across the 3,696 citations we already held before this shipped: article 32%, video 19%, listicle 17%, not classified 16%, product page 6%. This is a different question from the source labels you set yourself — those record a site’s relationship to your business, this records the shape of the page.

Where a topic sits in the buying journey, and the engine works the nearer ones first

Every topic now carries a stage, read from its own words by a rule you can see and correct: a question word means learning, a price word means buying, a comparison word means comparing, and anything else stays unset rather than being guessed. Click the chip on the row to change it. The Evidence Engine bands its queue by it — open answers first, then the commercially nearer question, then impact, then each cluster worked together — so work that is closer to an enquiry is not sitting behind a definition article. Setup also now proposes who you serve, read only from the words your own pages use after "for", each with the page it was found on. It never infers an audience from your business type: a shop does not thereby serve "shoppers", and silence is a real answer.

Search Console belongs to your site, not to a campaign

Your Search Console data was reachable only through a campaign, so dropping an old campaign took the site’s keyword history with it — and four campaigns on one site showed you the same table four times. The project page now has its own Search Console section and its own keyword table: every keyword the site’s campaigns track, one row each, paused and finished campaigns included, with its own Pull from GSC. Campaigns still own the keywords; nothing moved. We also stopped asking Google for whichever property a connection happened to have selected, and ask for your site’s own property instead — so a site you genuinely hold no longer reads "no property for this site" while the panel beside it shows that site’s clicks. And a reading pulled from Search Console now counts the same as a nightly rank check: it can mark a target met, where before it wrote the number and left the target unlit.

Disconnecting a Google account no longer erases what it published

The dialog said your campaigns would be "paused until reconnected". What it actually did was delete every record of what that account had published for you. On one account a database error stopped it; elsewhere it went through. Disconnecting now does what it said: campaigns pause and keep their link, the account goes inactive and hands back its tokens, and nothing is deleted. Reconnect the same account and it is found by email and given a fresh token. Connections offers Reconnect rather than leaving you to set it up from scratch.

The decks we build now end on the sources they rest on

Of the four kinds of slide deck the engine builds, only one ever asked its writer for outside sources — and it was not the one that gets published, shared and linked to. The other three were asked for "five slides, three bullets each" and a line pointing back at your own site, so there was no third-party reference anywhere in the request and therefore none in the deck. All four now ask the same way and close on a Sources slide. Nothing is invented to fill it: a line without a title or a real address is dropped, every address is fetched before it is kept, and a deck with no surviving sources simply gets no Sources slide rather than a padded one.

A second note on a file now says something the first one could not

The engine leaves a comment on the Docs, Sheets and decks it creates for you, a while after making them. It was picking one of seven fixed lines at random, every one of them pointing at your homepage — so a file with five comments carried five rewordings of "more details at your website". A comment now cites a page we actually published for that business, by name and address, chosen from what exists at the time. As the engine publishes more, a later note has something to say that an earlier one could not, which is the only good reason to leave a second one. It cites a page only where that page is plausibly about the same subject: a note on a document about satellite broadband pointing at a page about fishing gear is noise on your own Drive file, so with no overlap it falls back to the plain line. Your business name is excluded from that judgement, because it appears in every title we write and would otherwise make everything look related to everything.

Every publishing destination now shows its own health

The Publishing tab held one test result between all your destinations, so a destination nobody had ever tried looked identical to one that works — and testing one and seeing green read as a clean bill of health for the rest. It only ever meant "this was the last one tested". Each destination now carries its own line: never published, in amber, because never-tried is worth acting on rather than being the resting appearance of a healthy connection; when it last published, read from the server so it survives a reload; and "accepted, no page" in amber rather than green, because a receiver that answers successfully and names no page has not published anything you can open. Two more travel with it: a webhook can now send the header name your own receiver asks for, as well as the two we always sent, and a receiver that has seen a post before can update it instead of creating a second copy.

Connect a site and the work already done is offered, not left behind

An account that published to Google properties for months before connecting a CMS has subjects with evidence and no page of its own. When a publishing destination appears, that backlog is now offered — oldest source material first, because the age of the material is the reason it is worth having. It is an offer: nothing publishes on its own. The rate is deliberate. Where you have a Cloudflare zone connected, the crawl rate we observe governs it, since publishing faster than your site is read buys nothing; without one, five a day or 5% of your known pages, capped at 25. And a page the engine finished and parked because your site was refusing it now goes out on the next hourly check once the destination answers again — using the page it already wrote, rather than paying to write it a second time.

The engine keeps looking for topics even when nothing names you

Topic discovery mined your last visibility run for material, and collected only answers that mentioned your business or captured a competitor. An account nothing named produced no material, so discovery returned nothing — the engine broadened fastest for whoever needed it least and stopped dead for whoever needed it most, which is the inverse of what it is for. It now falls back to your own steer and your seeds, and the notification says which of those a topic came from. With neither, zero is still the honest answer: inventing a direction would expand your account toward somewhere nobody chose. Discovery also reads your own Search Console demand now — the strongest searches people really ran on your site that you are not yet tracking go in as candidates, with no model involved, because a search somebody typed is something we can point at.

Nothing in the product can post to Reddit any more

The Reddit screen had a Post button, and a scheduler behind it that published through a connected account. Both are gone — not hidden, removed, along with the job, the endpoints and the stored queue. A reply in somebody else’s community is a human act: the account earns its standing, and a comment that reads as posted by software is the one that gets downvoted and removed, which is the opposite of what the feature is chasing. Everything that never needed an account stays: find the thread, read it, draft the comment, copy it, post it yourself, and paste your comment’s address back to close the task.

The video switch now tells you why it produced nothing

Fourteen campaigns had the explainer-video switch on and had never produced a video. Every reason it declined was written to a log nobody reads, at a level production does not record — and a decline creates nothing, so no screen had a row to show. Each refusal now says itself three ways: in the log, on the run, and, where you can act on it, as a notification naming the fix. It also reaches the queue at all now. A campaign in Maintenance with video on gets one video a week, taken FIRST in that run rather than last in a budget it never reached; weekly, because each upload spends a fixed share of a daily quota shared across the platform. And the narration voice belongs to the project rather than the account, so several campaigns on one domain sound like the same business. It sits in project settings beside the writing voice — two different settings, now labelled so nobody changes the wrong one.

Which AI crawlers came for the pages we published

Published, crawled, retrieved, cited — and only the last of those had ever been watched. We could fetch your pages as GPTBot; we could not tell you whether GPTBot had. Connect the Cloudflare zone that serves your site, with a read-only token you create and can revoke, and every night we read which bots fetched which of the pages we placed there: the page, what crawled it and how often, whether an assistant fetched it live for somebody mid-answer, and when it was last seen — with "nothing yet" against the pages nothing has come for. That is the point: never fetched and fetched-then-ignored are two different failures with different fixes, and they used to look identical. The limits are stated rather than buried. Your site has to sit behind Cloudflare for there to be an edge to read, it covers the pages we published on that host rather than your whole site, and the edge keeps about eight days. A fetch is not a citation and is never shown as one. Growth and up.

Three things the reports were saying that were not true

The outcomes report said "citation is only counted after 30 days", and dropped every asset younger than that from both sides of the ratio — so a document published on the 1st and cited on the 6th simply vanished. A citation is an observed event with a date on it; nothing about our arithmetic makes it uncountable. Those are counted and shown now, kept out of the rate because a numerator cannot live outside its own denominator, and the sentence states OUR rule about when we will draw a conclusion rather than a claim about how long Google takes. Separately, "1 of 1 indexed" sat beside a published count of 19, with nothing saying the other 18 had never been checked — they are checked now, 200 a night, and every branch says what it excluded. And the two reports that seem to disagree — one counting answers that NAME you, the other counting answers that cite a page we published — now each name the other and say which way the difference runs. No number changed; the screens were simply silent about measuring different things.

Tell us what the sites AI cites actually are — and see how often each kind names you

A citation is a page the engine used; it is not a mention of you. Measured across every account on 1 September: only 37% of third-party citations sat in an answer that named the business at all. But that number hides an enormous spread. A page that STOCKS a business named it 89% of the time; a competitor’s guide named it 1% of the time. Both counted as "a third-party citation", and nothing told them apart — because whether a site is your stockist or your rival is not something we can read off its name, and guessing it from how often it names you would be circular. So we now ask. Under "Where AI’s answers come from" on the Visibility Audit, each cited source has a label: stockist, competitor, directory, publisher, community, reference. We pre-fill the ones a known list can tell (a review site, a forum, a site on your competitor list) and mark them as our guess; the rest are yours. The naming rate is then reported per kind, each as a count with its denominator. When sites AI keeps citing have no label, your dashboard’s "Only you can do this" band says so.

Pages we place on your WordPress site can now reach Bing — and so ChatGPT search — the day they go live

We have always told the search engines about every page we publish for you through IndexNow, the "this page is new" notice that Bing, Yandex and others accept. For pages on your OWN site, those notices were being rejected: the engines only take instructions about a host from someone who can put a small key file on it, and we had no way to put one on yours. Now we can, for WordPress: on Publishing → WordPress, press "Set up IndexNow" beside your site. We place the key file through your media library, fetch it back to prove your site serves it, and from then on every page we place there is accepted. It matters because ChatGPT search and Copilot retrieve from Bing’s index. For any other host, "Show the key file" gives you the file to place by hand; the Help article "IndexNow" explains the rest, including what Google does instead. Since 2 October 2026 each site has its own key, the button is “Place it through WordPress” on the site’s IndexNow card, and the card shows what IndexNow itself answered.

A failed AI engine no longer reads as your brand going unmentioned

Your AI visibility headline — the share of prompts where an engine named you — used to count every call that FAILED as a prompt where you were not mentioned. So when one engine had an outage or hit its quota for a whole run, your rate was dragged down by answers that were never given. One account was shown 45% for a run in which nine of ten real answers named it. Runs from 1 September 2026 onward exclude failed calls from the headline, the same way the per-engine and per-surface breakdowns always did; failed calls are still counted and shown as failures, they just no longer masquerade as absences. Older runs keep the number you were shown until we deliberately recompute them, and if we do, the movement will be upward and you will read about it here first, not discover it on a chart. Update, 3 September 2026: we have now recomputed them. 22 runs across the accounts had a failed engine counted as absences and were understated; their stored rates now exclude the failed calls, the same rule every run since 1 September has used. Runs where every call failed stay at zero, because they measured nothing. Nothing else about those runs changed, and the answers themselves are untouched.

Your own AI can now read your account — safely

If you use Claude or any AI that speaks MCP, you can now point it at your Ranking Factory account and ask it questions we never built screens for — "which topics am I measured on but never named for?", "what is still waiting on me?", "who do the engines cite instead of me?". In Settings → Agent access you create a token: name it, choose exactly what it may read (visibility, gaps, published evidence), copy it once, and hand it to your AI with the address shown on that screen. It is read-only by construction — nothing an agent can call changes anything — and the token opens nothing else: it is a different kind of credential from your login, each blind to the other’s doors. Revoke it any time; revocation is immediate. Every number an agent reads is the same number your screens show, sample sizes included. The technical description lives at /.well-known/auth.md.

Dismissing a gap now sticks — and the score stops counting things that didn’t happen

Two corrections to the Evidence Engine’s bookkeeping, both in your favour even where a number moves down. First: dismissing a gap quietly failed to stick — the next scan re-raised the identical gap, so "no" meant "ask me again next cycle". Now it sticks: a dismissed gap is never re-suggested, we ask (optionally) why you dismissed it so we can learn the difference between "not relevant to my business" and "not right now", and dismissed items sit in a collapsed "Set aside" section with a Reopen button instead of vanishing. Second: when the engine tried three times to publish something and could not, it recorded that gap as CLOSED — which inflated Entity Completeness, the number that is supposed to mean "AI can now see this". Those gaps are now marked honestly as given up, they appear under "Set aside" too, and reopening one gives it a fresh set of attempts. If your completeness percentage dips slightly after this update, that is the correction landing: the score stopped counting work that never happened.

When publishing to your site fails, we now diagnose it properly — and never pile up copies

A real case taught us three lessons at once. A customer’s WordPress had a plugin that crashed AFTER saving each post, so every publish "failed" with a server error while the article actually went live — and each retry added another copy. Meanwhile the dashboard card blamed an unreachable server (it had misread a post ID in the error as a status code), and its link dropped you on the general connections page to hunt for the problem yourself. All three are fixed. The card now recognises a site that answered-but-crashed, explains why your connection can still test as verified while publishing fails (the test signs in and reads; it is creating the post that crashes), and names the plugin your own site blamed. Its link now lands on the exact connection that is failing, highlighted. And before publishing anything to WordPress we now check whether that article already exists on your site — so a retry returns the page that is already there instead of creating a duplicate, and a copy sitting in your review queue blocks a second one rather than being written again.

Everything we built for a subject now becomes one page on your own site — and we found the ones that never did

The Evidence Engine builds several things for a subject — a guide, a specification, a checklist, a data table, a deck — publishes them to Google properties, and the page on YOUR domain is the placement that matters most. It turns out that last step had largely never happened — even our own site had none; its first went up on 5 September 2026. The Evidence Engine now shows you that list, per project, under "Built, but not yet on your site". Each row says how many pieces are ready and publishes in one click, assembled from what we already wrote — no new AI spend, and no second article competing with the first. If you already wrote that page yourself, paste its address and we record it as yours and stop offering it. Two things it deliberately will NOT do: publish them all at once, and offer a subject your own site already covers, because a page of ours competing with a page of yours for the same search helps nobody.

We can now read back what we published for you months ago

Assembling that page needs the words we wrote at the time, and for most of your history we had not kept a copy — only a link to the Google Doc, Sheet or deck. The documents themselves are still there, so we now read them back from your own Google account each night, a few hundred at a time. Nothing is regenerated and nothing is charged: the words were written once and paid for once. As they arrive, older subjects move from "cannot be assembled" onto the list you can act on, so that number falls on its own over about a week. Anything since deleted or unshared in Drive stays where it is, and says so.

Pages now open with the answer, and carry their questions as text

A study of 304,805 cited URLs scored content on traits with schema deliberately excluded, and found the correlation sits in the words a reader can see — a clear answer at +32.8%, useful questions and answers at +25.5%. We lean hard on markup, so the finding is not that markup is wrong; it is that it is not sufficient. Everything the engine writes now opens by answering the question rather than describing what the page is about, and carries real questions and answers as visible text rather than only as hidden markup. Where a model still opens with "In this guide we look at…", that sentence is removed before publishing rather than paid for twice.

A gap now closes on the thing it actually asked for

Each gap names the kind of evidence it wants — a guide, a checklist, a specification. When one of them was published, every open gap on that subject closed, whatever it had asked for. So a checklist could be recorded as done because a specification existed, and the number meant to mean "AI can now see this" was counting work nobody had done. Gaps now close only on the object they requested. You may see some reopen: that is the correction, not a fault, and it means those pieces genuinely were not built.

"What works" is now "What’s working" — and it opens with what we did for you

The report answered one question: how much of what we published got indexed and cited. Both of those have to wait — indexation is not counted for fourteen days, citation for thirty — so a young account could only ever read as a column of apologies while dozens of things had in fact been published for it. It now opens with the work itself, which needs no waiting period: the subject we found a gap on, the pieces we built to answer it, the article we pulled them together into on your own site, and the spreadsheet we fed it back into — each with the dates it happened on. The measured tables are unchanged and still sit underneath, because whether it WORKED is their job and this section must not borrow their authority. It also stops crediting the wrong thing: work from a recipe run is listed under that recipe, work from the Evidence Engine is listed by what it built (it does not use recipes), and your scheduled posting is counted separately from both — a rate that mixes them cannot tell you which one to do more of.

When nothing has reached your own site, we now say why — and offer the thing that fixes it

"No page on your own site yet" was a nag, and an ambiguous one: it means opposite things depending on whether a page is on its way or nothing was ever connected to send one. The line now names the actual cause, read from your account rather than guessed. Nowhere connected to publish to, and it sends you to choose. Articles written and sitting in your site’s review queue, with how many and how long they have waited. Publishing attempted and refused, with the reason — and where that reason is a wall of server output, we now pull out the part you can act on, such as which plugin on your site is crashing, rather than printing the stack trace. Only when everything is connected and working, and the page is genuinely next in line, do we offer to bring it forward: running that against a broken connection would just fail faster.

Your own blog posts now count as work — they were invisible before

The report read only the evidence ledger, so an account with thirty live articles on its own site was told nothing had reached it. Two things were wrong. Posts we publish through WordPress or your own CMS were never fed into the story at all. And where your site holds posts for review, we asked whether they had gone live using a web address we had guessed rather than the one your site actually assigned — so articles that were published, correct and readable were reported as still waiting. We now ask your site by title, which is the thing both ends agree on. Articles are named in the story rather than offered as a bare link, and where nothing is genuinely waiting the report stops saying that something is.

Wix sites can now be published to

Connections → Publishing takes a Wix site alongside WordPress, Ghost and webhooks, so the Evidence Engine can put gap-closing pages on a Wix blog. You paste your site ID and an API key from your own Wix account — no app to install and nothing to approve — and we handle the rest, including converting each article into the format Wix stores posts in, and finding the author from a post your blog already has so there is no ID to hunt down.

We measure Google's AI Overview now — as its own surface, never blended in

The generated answer above the blue links is where most people meet an AI answer at all, and we did not measure it. We do now. Choose the searches that matter to you and each is read from a LIVE Google results page: whether an overview appeared, whether it named you, and which sites it drew on instead. It is deliberately NOT part of your AI Visibility score — that score comes from asking the engines directly through their APIs, this is a live page with its own retrieval, and averaging the two would give you a number describing neither. You will find it under Reports & Tracking → AI Overview, next to Visibility, which is where the difference between them is easiest to see.

Which sites the overview keeps coming back to

Open any watched search and you get its source history: every site Google's overview has drawn on, in how many checks out of how many, marked as one it keeps returning to, one that comes and goes, or one seen only once. Your own domains are flagged. "You were not named this week" has no next step attached — "it has drawn on these four sites in every check for a month and you are not one of them" names the thing to go and be. Two rules keep it honest: nothing is called a site it keeps returning to on fewer than three readings, and the note always says that most of what moves in that list is the surface changing, not your site slipping — without putting a figure on how fast, because the published measurements disagree too sharply to quote one. Reading normal volatility as your own decline is the one mistake this feature could cause, so we say so every time.

A search Google answers without you can become work in one action

Where an overview appears and does not name you, the row now offers "Work on this" — it hands that search to the Evidence Engine as a topic on that project, through exactly the same rules the competitor-gap rows use. Often it will come back "already a topic in this project's plan", and that is a real answer rather than a dead button: it means the engine is already building for it, so the problem is not a missing topic — the evidence exists and is not being chosen. That is when the source list above it becomes the thing to study.

What you watch is a budget you control, and you choose how often

These checks run on your own DataForSEO account, which is how the plan price stays where it is. How many searches you can watch comes out of a monthly spend — $10 on Starter, $40 on Growth and Agency, your own figure on Enterprise — and you can set it LOWER than your plan allows whenever you like. How often is yours too: weekly keeps the list long, daily gives each search a firmer rate on a shorter list, and the page shows both numbers before you choose rather than presenting one as correct. At these volumes DataForSEO's $50 minimum deposit is the real cost rather than the usage, and it lasts a long time.

A citation we cannot trace no longer counts as independent coverage

Every source an engine cites is sorted into your own site, a property published for you, or genuinely independent — and that last figure decides whether a sentiment reading means anything or is your own marketing read back to you. Gemini returns its citations as opaque redirect links with the real site hidden in a separate field, and we were reading the link. That attributed every Gemini citation to Google and counted it as independent — including citations of your OWN pages. Fixed, and hardened beyond Gemini: anything we cannot trace to a site is now counted and shown as "not attributable" rather than quietly improving the one number this whole classification exists to keep honest. Since September we go one step further and follow each of those redirect links to the page it names, so a Gemini citation records the page rather than only the site.

Reddit threads can now find you — daily, by email, to whoever answers them

Thread discovery used to run only when you pressed the button. Now each project can turn it automatic: every morning the engine searches for live Reddit discussions already asking what you answer, opens each as a task with reply guidance, and emails the list to an address you choose — yours, or a VA’s, since a person writes the replies either way. We never post for you — as of September there is no posting route left in the product at all, so it cannot happen by accident. It runs on your own Serper key, sends nothing on an empty day, and every email carries a one-click turn-off link that needs no sign-in — so whoever receives it can also stop it. Project page → Announcements.

Gaps closed on Google properties now become pages on your site too

When a gap is closed with a Google Doc or a cloud page and you have a site connected (WordPress, Wix, Ghost or a webhook CMS), the engine now follows up with the same content as a page on YOUR domain — the stronger placement. It reuses the content it already wrote, so there is no second AI spend, and the page’s machine-readable provenance names the original object as its source. Spaced out rather than dumped at once, held for approval if you have review turned on, and the claim checks run again before anything lands on your domain. Projects without a connected site are untouched.

We now publish without asking — and tell you what we did

Until now, new public posts (LinkedIn, your own blog) waited in a review queue unless a campaign was set to a mode called AutoPilot that most people had never heard of. Many accounts had posts sitting there that nobody knew were waiting. That is backwards for a product whose promise is "you do not need to do this or that — we will." So the default has flipped: we publish, and your weekly report now opens with exactly what went out, every item with its link. If you would rather approve each post yourself, turn on "Review before publishing" in Settings — and when it is on, we email you the moment something is waiting, so nothing sits unseen. Either way the checks that are about correctness rather than taste still apply: no unsourced "best in the county" claims, no links a stranger cannot open, nothing written without real grounding from your site. If you have posts in a queue today, they will go out on the next nightly pass unless you turn review on first.

Safe fixes run themselves at night — the button is gone

"Fix All Safe Issues" asked you to press a button for work the nightly pass was already doing unattended; its label told you things were waiting on you when they were not. It is gone. Schema sync, entity injection, cluster building and the other low-risk maintenance run every night at 03:30, and your weekly report counts them. A small "run them now" remains on the campaign plan for anyone who does not want to wait until tonight — an option, not an obligation.

We read your site and tell you what we think you do

Setting up used to ask you to type keywords on your first day — an expert decision, before you had seen anything the product does. Now we read your site and propose it: what kind of business this looks like, the places you serve, and the topics you cover. Every topic says which page we found it on, so you can check us rather than take our word. Correct it, remove what is wrong, add what we missed — what you leave is what we work on. Nothing is saved until you confirm, and if we could not read your site (a browser-only site, or one that blocked us) we say exactly that instead of inventing a list that looks plausible.

Your first measurement now happens AFTER you have told us what you do

It used to run first, on our guess at your business, which meant every later comparison was against a starting line drawn in the wrong place. It now runs on the topics you confirmed, and asks two questions per topic instead of a fixed six. That also fixes something we were not being straight about: six questions carry a 95% confidence range of roughly ±35–40 points, which is a range where 30% and 65% are the same reading — we were showing that as where you stood. Under 20 questions it now calls itself a first look rather than a measurement, and tells you how many more topics would make it a baseline you can track against. The number of questions is shown before the run spends anything on your key.

Several branches or divisions, set up in one pass

A business with eight locations meant eight separate setups — eight crawls of the same site, eight copies of the same description. Now they go in together on the project, and each can become its own campaign in one action. A branch and a division are kept firmly apart, which matters more than it sounds: a division like "Commercial" is not a place, and folding it into a location search produces "best plumber in Commercial" — a phrase nobody types, that we would then track and charge you to check every night. Divisions are never searched as places, and the screen says so at the point you choose. Agency and Enterprise.

We check a link opens before announcing it

Google Business Profile posts went out without anyone checking the page they pointed at was readable by a stranger. LinkedIn has checked since a post in August sent people to a permission error; Business Profile did not — and it is the worst place for it, since the post sits under your business name in the map panel with a Learn more button, in front of people looking you up. It now checks first and holds the post back rather than publishing a dead end.

Pages published to your own site now get announced

The announcement layer was pointed at the wrong thing: it announced Google Docs, while the page published to your own site — the one a reader should actually land on — went out silently. It now announces the real page, using the summary already on it rather than writing new words about it.

Keywords a competitor holds and you do not

A report (Reports & Tracking, advanced mode) built from rank checks you are already paying for, so it costs no extra searches. It lists the keywords where a rival appears on page one and you do not, weakest rival position first, since those are the winnable ones. It states its sample size, and warns when it could not recognise any of your own domains — because if it cannot, you appear in your own competitor list.

Failures you can fix now reach you

Three more failures now send you a message instead of only a line in a log nobody reads: Search Console refusing a property you do not own, a saved connection that can no longer be decrypted, and a rank-tracking key out of credits. Each one says which screen fixes it. Failures that are ours to fix are never addressed to you.

A weekly check on your saved connections

Stored passwords and tokens can become unreadable — usually saved before an encryption change, and they never recover on their own. Rather than finding out one at a time when something fails at 3am, we now check them all weekly and tell you the whole list at once, with the screen that fixes each. Four broken cloud keys is one afternoon’s work, not four separate problems.

Two quiet faults, fixed

A campaign with no Google account connected used to fail with a meaningless error five times a night; it now says plainly that no account is connected and carries on with everything that does not need one. And a settings page could save to one row while the job read another — so a change appeared to take and then never happened. That affected email settings, blog scheduling and Google Sites credentials.

See how AI describes you, not just whether it mentions you

A visibility run now shows the words the engines keep using about your business, with how many answers used each one. This is a different question from "were you named" and from "was the tone positive": a business can be highly visible and described in terms that lose it every serious enquiry, and a mention count would show that as a rising line. We report the words as found and never score them — whether "automated" or "budget" or "boutique" helps or hurts is your judgement about your market, not ours. Costs nothing extra: it reads answers the run already collected.

A shorter dashboard

Projects are now a single row you scroll across, and the status bands scroll inside their own boxes rather than stretching the page. With a few projects and a busy engine the dashboard had grown to several screens, and your projects — usually the reason you opened it — sat below the fold. The row is the same height whether you have three projects or thirty.

We now tell you when the "reputation" we measured is your own marketing

Every source an AI engine cites is now sorted into three: your own site, properties published on your behalf (Docs, Blogger, cloud pages), and genuinely independent coverage. When nothing independent was cited, the run says so — because a sentiment score built only from your own material is that material read back to you, not a reputation. Ambiguous sources are counted as yours rather than independent, so uncertainty never flatters the number. Earned coverage is the one thing this product cannot publish for you, and now you can see exactly how much of it you have.

The visibility score shows what it is made of

It was always five weighted parts — citation 40%, crawlability 20%, machine legibility 20%, crawler permissions 15%, llms.txt 5% — and you only ever saw the total. Now you see each part with its weight, and which one is costing you the most points, which is not the same as the lowest number: zero on llms.txt costs 5 points, while 40 on citation costs 24. Anything we could not measure is left out of the total rather than scored zero, and labelled as such.

Published pages now carry provenance an AI can read

Pages the engine publishes to your own site now include a machine-readable record: when the page was drafted, when it was published, and a linked list of the sources behind it. An AI system reading the page can follow that back to the original rather than taking our word for it — a sentence saying "published in August" proves nothing to a machine. We credit your business, never a person who did not write it: attaching a human byline to generated text is exactly the kind of thing this product refuses to do. One exception, named rather than covered: Wix stores posts as structured content with nowhere to put a block like this, so a Wix page carries the same sources as visible links instead. Coming next: the same block on Google-property and cloud assets, and sources attached claim by claim.

We will not publish "we are the best" without a source

Comparative and superlative claims about your business — best in the county, leading supplier, faster than any competitor — are now caught before publishing. If the model writes one we ask it again, naming the phrase; if it still will not comply we cut those sentences and publish the rest, so one bad line does not cost you the page or delay it to the next cycle. Two reasons: comparative advertising is regulated in most markets, so an unsupported superlative is your liability that we created. And it is the pattern AI systems discount, so it costs the citation it was reaching for. A claim WITH a source — "rated best by the Yorkshire Post", with the link — passes untouched.

See what each kind of work actually achieves

A new report (Reports & Tracking, advanced mode) showing indexation and citation rates for everything published for you — so an approach can be proved or retired on evidence. It is built to be able to tell you bad news: work published too recently to have had a chance is excluded rather than counted as a failure, anything we never checked is left out rather than counted as a miss, and below ten mature items you get counts instead of a percentage. On a new account it will mostly say "not enough yet", which is the point.

Each project now says where it publishes

Publishing connections belong to your account, and you may have several — but nothing recorded which one a project used, so we took the first WordPress site we found. That meant one project could publish to another business's blog. We now only publish to a connection on the project's own domain, or one you pick: Project → Where this project publishes. If your blog is on a different address from the project (a client site, a shared agency blog), choose it there — until you do, those pages are built and handed to you rather than published somewhere plausible. The screen also says where each project is publishing right now, before you change anything.

Announcements are now off until you choose them

Google Business Profile posts used to go out through any connected account on your workspace, without anyone picking it. They now need the same per-campaign opt-in LinkedIn has had, so connecting an account for one purpose no longer volunteers it for another. If your posts have stopped, that is this: Campaign → SmartStack → Settings, tick the surface you want. Each card states the volume before you choose.

Engines now search the web when they can — and it costs a little more

A visibility check used to ask some engines what they already knew, which cannot include anything you published this week. So you were paying for a number that could not move: publish all you like, a trained-from-memory answer will not notice for months. Gemini and OpenAI now look things up through their own web search as they answer — and since 7 September so does Claude — each reporting which pages it read. That means a real web search per prompt, which your provider charges for on top of the usual tokens, and runs take longer — a few pence more on your own key, in exchange for a measurement that can actually respond to your work. Perplexity always searched, so nothing changes there; engines that cannot search are unchanged and are labelled as answering from training rather than quietly presented as fresh. Expect your Gemini figures to shift when this lands: it is now measuring something different, so it starts a new trend line instead of rewriting the old one.

See which engines can respond to your work — before you run

Choosing engines now tells you how many of them search the web. Pick three that answer only from training and the numbers are guaranteed to stay flat no matter what you publish, which looks exactly like failure. Both kinds are worth measuring; you should just know which you picked.

We say where each number came from

Every visibility figure is measured by calling an engine's API — which is repeatable and can carry a confidence interval, but is not the same as Google AI Overviews, AI Mode or the ChatGPT app. Runs now state that plainly instead of leaving you to assume.

A dashboard that says what is running, and what needs you

Three bands instead of a checklist: what is happening without you, what only you can do, and what you could unlock. The first band only ever claims what is genuinely running on your account — if your plan runs the engine on demand, it says so rather than implying a schedule. No progress bar, no dismiss: an account with nothing outstanding now gets told exactly that.

Reddit threads worth answering

Finds live discussions already asking what you answer, and opens each as a task with guidance — no link and no business name to begin with. It needs a Serper key, and we tell you what that costs rather than letting you find out.

Your topic list evolves, and you can steer it

The keywords you typed at setup are now a living list on the project page: what you seeded, what we found, how often each was measured and named, and controls to remove, add or mark a focus topic. Nothing is hidden — including how many topics each run measures, since that spends your own AI credits.

A wasted AI key stops being wasted

A provider that has hit its daily quota, or a model the provider has retired, is now skipped for the rest of the day instead of being retried hundreds of times. You get told once, with the actual fix — and choosing a current model lifts it immediately rather than making you wait out a cooldown.

Links that go nowhere are no longer published

A cloud connection missing its region or endpoint could produce an address that does not resolve. Now, if we cannot build a real URL for something, we say the page is not publishable instead of publishing a link that looks fine and goes nowhere.

Page-by-page analysis

The site check tells you whether AI can read your site; this tells you which PAGES are weak and what would fix each one. It also spots problems appearing on most of your pages and names them as one template fix rather than thirty separate ones. Growth and up.

Write in your own language

Set a content language for your account, or per project, and generated content follows it. Change it later and we ask which existing projects and campaigns should move too, rather than silently rewriting them.

We spot content trapped in images

A page whose services, hours or prices only exist inside a graphic looks almost empty to AI. The site check now says so in those terms, instead of telling you the page is thin when the writing already exists — as a JPEG.

Scan several services at once

Gap scanning took one service at a time. Agency plans now scan 10 at once and Enterprise 25, with anything over the limit named rather than silently dropped.

Three steps, not six

Onboarding is now check your site, add an AI engine, start closing gaps — and the dashboard tells you what to do next from what we actually found, instead of asking you to pick a recipe on day one.

Dig deeper from any finding

Every site-check finding links straight through to the audit behind it, so a one-line summary is never the end of the trail.

Ask each prompt more than once

AI answers are not deterministic. You can now ask each prompt 3 or 5 times and see which ones the engines have not made their mind up about — reported as counts, never as a percentage that hides the sample size.

Publish to your own site — beyond WordPress

New Webhook and Ghost connectors in Connections → Publishing let the Evidence Engine publish gap-closing pages to any headless CMS, Zapier/Make, or a Ghost blog.

Doc Chain

A recipe that builds a sequence of Google Docs, each linking on to the next, with the chain ending at your site anchored with your business name. It needs at least two Docs to form a chain and skips rather than leaving a dangling link. Each link is a comment on its Doc, not a link in the Doc’s text (said in its guide since 7 October 2026).

Real jobs you've done

Add your genuine past work on a project, and case-study content is written from those rather than invented. If there is nothing real to draw on, it holds back instead of making something up.

Clearer errors when something fails

A campaign that produces nothing now tells you why, naming only the parts it actually uses — and if your AI provider is out of credit, it says so rather than sending you to check settings that were fine.

The Evidence Engine now reads your real website

It crawls your sitemap to learn what your site already publishes, so it only flags evidence you're genuinely missing — not things you already have. Run "Analyse my site" on the Evidence Engine to see it.

Gap content is grounded in your actual business

Generated FAQs, guides and case studies now draw on your real site and services (no invented facts), and each asset links to the specific page that targets the topic — not just your homepage.

Close gallery gaps with your own images

A gallery only counts as evidence if the photos are genuinely yours. Paste image links, upload files, or point at a Google Photos album — never stock or AI.

Weekly news & tips email

A branded weekly email of actionable tips plus your latest posts. You preview exactly what goes out and send it with one click — nothing sends automatically.

Explainer video for any blog post

Generate a narrated video from a post and embed it at the top. The narration voice is yours to choose — the free on-server voices need no key and sound synthetic, and Google’s cost a few pence a video on your own key. Videos now go to your connected YouTube channel and the post embeds them from there, rather than serving a file from our server.