Everything it measures, writes, publishes and proves

Site check

One address in, one prioritised list out. It runs the SEO and GEO audits together and folds in your AI-visibility results, then orders everything by what to act on first: blockers where AI cannot read you at all, services AI is not naming you for, being named but not cited, machine legibility, then emerging standards. Every finding says what was found, why it matters, what to do, and who can do it — the engine, a snippet you paste, or a developer. It checks the on-page basics too — title and description length, one H1, a viewport that fits a phone, a canonical tag, Open Graph and profile links a visitor can see — and ends with what it does not measure: Core Web Vitals, how the page behaves on a phone, and links from other sites, shown as not measured rather than passed. It also reads X-Robots-Tag headers, times each AI agent’s fetch, and checks that http and www redirect to one address, that headings are in order, that images carry alt text and that the page states its language.

Plans: Every plan. Needs nothing but a web address.

AI Visibility Tracker

Asks ChatGPT, Gemini, Perplexity, Claude, Grok and DeepSeek the questions your buyers ask, on your own keys. Named, cited and recommended are three separate signals, each with a 95% confidence range and its sample size, and since Nov 3, 2026 they are the headline; site checks sit in their own Search eligibility panel, never blended in. Answers from live search and answers from training are reported apart, and every run records its engine set so a trend never joins runs measured on different engines. Ask each prompt 3 or 5 times and stability is reported separately from rate, with the exact prompts that disagreed. Before you run, it says how many of your chosen engines can search the web at all. Each run also says whose sources the answers used — your site, published for you, or independent — and a page we published for you is never counted as independent, wherever it sits (a video or page made with the one-off YouTube Video or Cloud Stack tools before Oct 7, 2026 was not recorded, so it is judged by its site). Models are listed cheapest first with a cost per answer, and one click puts your chosen engines on every watchlist; the run form says how firm a rate from that many prompts will be. The trend is an interactive line by date in plain engine names, breaking only where an engine was added or dropped, and a gap’s before-reading is never a score or a point on it. After a site fix, “Re-read my site and re-score” re-scores the latest run from the site as it is now, with that run’s own answers: no AI asked, nothing spent. The score is 40% how the answers treated you and 60% how readable your site is to AI, worked through on How we measure. Each project is measured once a week by default once it qualifies: we email you what it will ask and about what it will cost before anything is spent, the first run is no sooner than a day after that email, and “Don’t measure this project” switches it off.

Plans: Every plan, on your own AI keys. Gemini needs billing on to search Google live; with it on, the first 5,000 searches a month are free.

Measured on our own account: Gemini’s live search cited therankingfactory.com in 32% of its answers on Aug 15, 2026 (6 of 19) and 74% on Aug 20 (14 of 19) — the same model (Gemini 3.7 Flash with Google Search), the same 25-question set and the same counting rule: our address or a link to us in the answer. 16 of the 19 questions carried our name.

Tracked prompts

Up to 20 exact questions per project, asked on your visibility engines daily, weekly or monthly, with every answer kept: whether you were named and where, whether your site was cited, and what the answer cited. One question, answer by answer, beside the score rather than in it. Questions you type under “What do customers ask you?” when you set a project up are saved as weekly tracked prompts.

Plans: Every plan, on your own AI keys. The screen says about how many answers a month a prompt costs before you add it.

The ChatGPT and Gemini apps themselves

Your tracked prompts asked of the ChatGPT app and the Gemini app, not only the vendors’ APIs: DataForSEO asks the apps through their own web pages, from your project’s location — ChatGPT from its country, Gemini from the place itself — and with no location set nothing is asked. Each app is its own line: named and cited, counted with the number of answers, and for ChatGPT the pages it looked at and the searches it ran. Never part of the visibility score, and an answer that asks a question back is counted apart rather than as “not named”.

Plans: Every plan, switched on per project, on your own DataForSEO credit: about $0.0012 a question, outside the AI Overview budget, with each app’s monthly estimate shown before you switch it on.

Google AI Mode, for your tracked prompts

Your tracked prompts asked of Google AI Mode itself, through DataForSEO, from your project’s location — with no location set nothing is asked — on each prompt’s own schedule. Its own line: named and cited, counted with the number of answers, and the pages it cited. Never part of the visibility score or of any AI Overview figure; a reading where Google gave no AI Mode answer is counted apart. A page of yours it cites is recorded as cited in Google AI Mode.

Plans: Every plan, switched on per project, on your own DataForSEO credit: about $0.0012 a question, no budget and outside the AI Overview budget, with the monthly estimate shown beside the switch.

Google AI Overview tracking

Whether Google’s AI Overview names you for the searches you choose to watch, read from a live Google results page rather than a model API, from the place you set for the project (the United States where none is set, and each search says so). Presence and citation are reported apart, with the sources the overview drew on instead and what kind of page each was. Checked weekly, or daily if you prefer — an overview is generated and two looks on the same day can differ, so you get counts across several readings with the sample size, never a percentage from one. Open any search for its source history and hand it to the Evidence Engine as work in one action. Kept as its own surface and never folded into the AI Visibility score. Only a place on DataForSEO’s own list is used; nothing is guessed.

Plans: Every plan, on your own DataForSEO credit: a default monthly budget of $10 on Starter and $40 on Growth and Agency, Enterprise sets its own, each lowerable. One ceiling of 500 watched searches per account is ours and says so.

Google’s own AI impressions

How often Google’s AI surfaces showed your pages, imported from the Search Console generative report because no API exposes it — not the Search Analytics API, not the BigQuery export. Uploaded weekly, reported apart from everything measured by asking engines questions, and tied to pages rather than prompts: the export carries no search terms, and the screen says so before you upload. It is charted by day in the Activity report and on What’s working, with each page Google showed marked as one our checks saw cited, or not.

Sites AI cites

Every site an AI answer cited for your topics, across all your projects, with links to the exact pages. Sources the engine declared and pages extracted from answer prose are kept apart; your own pages and third-party pages are kept apart; and you label what kind of site each one is — stockist, competitor, neighbouring field, directory, publisher, community, reference, or not related — because only you know. The naming rate is then reported per kind as counts with denominators, and a rule’s guess is shown as a guess. Where an engine hands back the sentence it quoted on a route we read — today, Claude asked directly on your own Anthropic key — that sentence is kept with the citation and your own AI can read the newest one per site through agent access; from Claude through OpenRouter and the other engines we keep only the page’s address, so most citations carry no quote. Below it, an outreach list ranks other people’s pages by how often the answers that cited them recommended you or a competitor, with a spreadsheet download.

Measured on our own account: Our 12 latest runs (Aug 15 – Oct 6, 2026, 27 subjects): therankingfactory.com is the most-cited of 588 sites, with 971 citations, and most of the questions carried our name. Each site is labelled by kind, so you can see who the engines lean on for a topic and where you already appear.

Ask your data

A box on each project’s Overview answers a question from that project’s stored readings. Read only, at most 6 lookups a question, and every number and link in the answer is checked against what the lookups returned. A drafted message is text for you, never sent.

Plans: Every plan, on your own AI key. The key and model are named before you ask.

llms.txt and llms-full.txt for your site

Your entity’s llms.txt in the standard layout, listing the pages we published on your site, and llms-full.txt with their full text, to download and place at your root. With our WordPress plugin connected, or on a cloud site with a root of its own, they are served or uploaded with no press, never over a file that is not ours. Google says Search ignores these files, and no crawler we know of says it reads them.

Answer accuracy against your own facts

What AI answers of the last 30 days say about your prices, phone, founding year and address, checked against the facts you stored - only the ones you stored. Each difference shows what was said, on which engines, and the page each answer cited, read again: 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 comparing several businesses, a figure counts as yours only in your own sentence, table row or column.

Plans: Every plan. No AI is asked anything.

AI citation share, per engine

On the Activity report, a ring for each AI engine your runs measured: how the sites its answers cited were shared between you, your three most-cited competitors, your other competitors and every other site, each slice a count with its whole - “14 of 166 citations” - beside how many answers cited each site, with its 95% range. Engines are never added together, and with three answers with sources or fewer a ring shows counts only. Put it in the report’s PDF or send a frozen share link a client opens without signing in; both name only your watchlist’s competitors.

Plans: Every plan; share links on Starter and above. It reads the runs you already have: no AI is asked anything.

A topic list that stays alive

The topics you confirm at setup become a living list: what you seeded, what the engine found in recent AI answers, how often each was measured and named, a buying-journey stage that orders the queue, and clusters so two phrasings of one question are not measured twice. A topic measured repeatedly and never named is suggested for retirement, and retired or taken off measuring only when you accept; new ones are found each cycle, up to 250 main topics per project. Whether an answer is still open or already held by a few recurring sources is judged from readings, never under three.

Searches that could become topics

The searches Google and Bing showed your site for, as one list on the project page: each engine’s impressions and clicks side by side and never added, the stage the words point at - learning, comparing or buying, marked inferred - and what each already is on your site. Add any as a topic, say Not this one and it is never offered or added again, or track a question in its own words. Once ten searches that are not your name each reach 20 impressions in one engine, topic discovery adds up to five a pass on its own, up to 60 main topics.

Plans: Every plan. The list asks no AI; Track this question shows its cost on your own AI key first. Discovery adds on its own on the Evidence Engine’s schedule (Growth and above) or when you press Run now.

Competitor gaps you can act on in one click

Searches where a rival sits on page one and you do not appear, built from rank checks already running so it costs no extra searches — weakest rival first, with the sample size stated. Each one can be handed to the Evidence Engine: it becomes a topic on that project, is scanned for what is missing, and the page is written on your own site. Where a search is really a longer way of asking something you already cover, file it under that topic instead. Each row links the winning rival page and says what kind of page it is; “Not for me” sets one aside; “Sort these for me” asks your own AI once to mark each as take it, fold it under a topic you have, or the same as one. Rank checks, Reddit and news look-ups use your Serper key first and carry on through your own DataForSEO account when it runs out.

GEO & Agent Readiness Audit

AI-crawler access probed as the actual bots — GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and the assistants ChatGPT-User, Claude-User and Perplexity-User — because robots.txt states intent, not behaviour; schema inventory; content structure; and a JavaScript-rendering check that catches content Google can read and AI crawlers cannot. Agent readiness is shown as its own part, for information and not scored since Nov 3, 2026: Content Signals in robots.txt, RFC 8288 Link headers, llms.txt, the .well-known API catalog, an MCP server card and markdown-for-agents, none of which a major engine has confirmed it reads. Branded PDF export. The robots.txt it hands you allows every answer and search crawler, Google-Extended included, and the audit reads the same list.

SEO Audit

Technical audit — metadata, headings, schema, links, redirects, sitemap and robots — with competitor analysis, a prioritised action plan and a branded PDF report. A title over 60 characters, a description over 160, more than one H1, a viewport that does not fit a phone, partial Open Graph and no visible profile link each cost stated points. The site check sits above it; this goes deeper on its own part.

Page-by-page analysis

Which of your pages is weak, and what would fix each. Template problems lead, because a finding on thirty pages is one theme edit; a page we could not read gets no score rather than a zero; and where your site is bigger than the scan, the report says so rather than presenting part of it as the whole.

Plans: Growth 15 pages a scan, Agency 40, Enterprise 100 — every page is a real fetch against your own server.

Evidence Engine

The self-running loop. It finds where AI engines and Google are missing you, decides what evidence the question needs, writes it, publishes it, and checks on later runs whether anything changed. Each cycle suggests retiring topics measured without ever being named — you decide — mines recent AI answers for topics you are emerging for and ones competitors own, then finds gaps against the list it just refreshed. A gap names the object it wants and closes only when that object is published. Pages for your own site wait for your approval on new accounts until you switch review off; it reports what it did, item by item, with links. Every writer is now told that each section must name its own subject, so a passage quoted on its own still makes sense; reference tables and calculators open with a one-line summary. It never sends a second piece on the same subject to the same place within 30 days (7 to 365 in Settings).

Plans: Growth every 6 hours, Agency every 4, Enterprise every 3. Starter runs it on demand.

The shape is decided before a word is written

For every topic, two or three of your own AI providers, from different companies, are asked what kind of page the question needs — a guide, a comparison, a how-to, a specification — and which concrete elements it must carry. The majority answer is stored as a contract; the page is written to it and checked element by element, each one quoted from the page, with one redraft for anything missing. Every topic competes on your own site unless you switch it to "Place it elsewhere": where the panel thinks the topic belongs to another kind of source, the page is still written in the shape that wins and the places it named are advice, and the panel is told which of the cited pages are your competitors’ own. Two or three calls per topic, once, on your keys. The panel also looks at what AI answers cite for the topic today: a Sheet or a deck is made only where that kind of file is being cited, a subject your site already covers gets no second page, and pages are filed under categories your site already has.

Written only from your own material

The crawl follows the links on your home page as well as the sitemap and keeps each page’s videos, tags and text. Every writer in the Evidence Engine 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 own words, 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, or both — and the page you point at is read in full. Figures nobody supplied are cut before anything is published, and so is a comparative or superlative claim about you that no source supports on a page for your own site.

Sources shown, not just named

For every subject, the pages that AI answers and search results cite are found and each address checked; a page whose site refuses AI use in its robots.txt is left out. The part of the three strongest that is about the subject is put in front of the writer, chosen by relevance, so it cites what a source says rather than what a model remembers about it. Before publishing, any figure in a sentence that names a source is looked for in that source’s own words and cut if absent; a sentence that reproduces a source rather than reporting it is cut; a source listed but never used is dropped. Every piece ends on the third-party sources it rests on, each fetched before it was kept.

Answerability

Whether a page actually answers the question it is shown for, judged the way an engine reads it rather than by word count. Before it goes out the page is checked against its shape contract, quoting the sentence that carries each element. Afterwards, a real Search Console query can be graded by two or three of your providers with a verdict and the grader tally; a page with no query is asked about its own subject and carries no verdict, only what the models say is missing. Pages that miss elements are rebuilt in shape rather than rewritten from scratch - and a page an AI engine already cites waits for your approval before its rebuild replaces anything - and the scores are readable by your own AI through agent access.

Plans: The contract check runs on every plan. Nightly grading of published pages is a setting, off by default, on your own AI credit.

Re-align published pages

Rewrites a page already published so that what it says about your product agrees with how you describe it — your one line, and your own pages and videos about it. The page keeps its structure and links, the sentences that defined your product without your say-so are named and dropped, and nothing replaces the live page until you have read the rewrite and approved it. On WordPress, a post you set back to draft is updated and stays a draft until you publish it.

A page follows its parts

A page on your own site is assembled from the Docs, Sheets and decks built for its subject. A part that arrives after the page — usually the deck — is now added and the live page updated in place, in the parts’ own words, with no new writing. It refuses to update over a change you made on your site, over a page that was rewritten as one text (or while a rewrite of it waits for you), or without you on an account that reviews first — there the update waits for your approval. A page an AI engine is citing is never rewritten until you have approved the change, whatever the setting, and after you approve an update the next new material waits for you again. The page gets a date it never had: when it was last modified.

Entity hub

One entity per domain: the organisation, its identifiers, its claims and the sources that cite it. It is written to a Google Doc and a Google Sheet that are kept in step when the entity changes, generates the schema.org JSON-LD your pages declare, drafts a Wikidata submission for you to file (kept on your account, with your Q-number and whether you have submitted it), and scores the eight places engines reconcile an identity against — LinkedIn, Crunchbase, OpenCorporates, G2, Yelp, Trustpilot, Product Hunt and Wikipedia, the last checked live against Wikipedia itself. Your canonical page is published to your own site as a dated AboutPage, carrying the facts you keep under your own names.

YouTube videos

A short narrated explainer — script from your material, a spoken voiceover, on-brand slides — uploaded unlisted to your own YouTube channel for you to review and make public, and embedded from there. Counted as published for you — in a run’s sources and on What is working — never as independent, and a Gemini link that never resolved to the video counts as one we could not trace; someone else’s video about you is still independent.

Measured on our own account: 14 videos from our own channel, published Sep 10–21, 2026, have been cited in AI answers since Sep 19: 7 first by Google’s AI Overview, 6 by an engine we reach through OpenRouter, 1 by Gemini (read Oct 6, 2026).

Reddit threads worth answering

Finds live discussions already asking what you answer and opens each as a task with guidance on replying. It cannot post: an automated reply is exactly what those communities remove, and the citation is only worth having because a person earned it. You reply, paste the link back, and that closes the task.

Conversations to join

Forum threads on the questions where AI answers did not name you - Reddit, Quora, Stack Exchange and community forums, and an X thread only where no forum thread passed. Each is checked to take replies, to be from the last 6 months, and - by your own AI - to be someone choosing, comparing or asking for a recommendation for what you sell; one line says what was left out and why. Below them, the pages AI answers cited that you could pitch, with “It’s a competitor” for a publisher a rival owns. We post nothing and contact nobody.

Plans: Every plan. Searches on your Serper or DataForSEO key (ours on Refresh where you have neither) and judging on your own AI key, its cost shown before Refresh.

New versions of the videos that are working

A video of yours that an AI answer cited, a Google AI Overview carried or Google’s page one shows is listed, unlisted ones first, with a brief for your video maker: the narration word for word, the key points, why it matters, a suggested title and description, and the page it will go on. Add the link of the public video they make and we watch it for citations as your own work; the original stays up as it is.

Your own site first

Connect WordPress, Shopify, Wix, Ghost or any webhook and a project publishes to a destination on its own domain first — never to whichever connection happened to be first, and never to another site’s. The Docs, Sheets and decks become the parts its pages are assembled from. Where no destination applies, the page is written and handed to you rather than published somewhere plausible. Every project opens with your own site: the subjects with a page, how many were asked about and cited in the last 30 days, and how far each page has got.

Webhooks the standard way, and updates in place

Every webhook message is signed the Standard Webhooks way — webhook-id, webhook-timestamp and an HMAC-SHA256 webhook-signature over the body — typed post.created or post.updated, and keyed on an externalId that never changes, so a resend cannot duplicate a post. A destination marked “Accepts updates” receives the update when a page changes; one that cannot take updates gets “Send again” and the HTML to copy beside every page. A scheduled post that times out or meets a 408, 429 or 5xx is tried again, five attempts over about two and a half hours, under the same webhook-id. The format is documented for your developer, with Node and PHP examples.

Your own facts, where Google reads

Docs, Sheets, Slides and Blogger under your own Google account, and Business Profile posts where you opt in: your own material, published where Google’s engines read it. Each run’s source count files a cited Doc, Sheet or deck, or a post on a blogspot.com blog, as published for you, never as independent. Every file is shared and published to the web as it is made, unless its project holds Google documents for your review — then each waits unshared in your own Drive until you publish it. A build-time check fails our own build if any code path creates one without a sharing step. Published work is read back out of Google nightly so pages can be assembled from the words as written. And you can take a campaign’s Drive files down from Google again — one file, a chosen set or the whole campaign — each unpublished, its own link sharing removed and moved to Drive’s bin; a file something else of yours still uses is left as it is, and named.

Announcements you choose

LinkedIn and Google Business Profile announcements of published work are off until you turn them on, per project, and each states its volume before you choose. Connecting an account for one purpose never volunteers it for another.

IndexNow

Each site has an IndexNow key of its own — made for it, or the one your SEO plugin already serves. Once the site serves the key file, the pages we place there are announced to Bing and the indexes that read it, including the one behind ChatGPT search. “Check now” fetches the file from your site and then asks IndexNow itself, so the card shows what IndexNow answered rather than what we assume: a refusal is shown in its own words and not repeated for a day. The key file can be downloaded to place at the root of any site, or placed through a WordPress connection’s media library — a folder, which IndexNow may not accept for your posts, and the card says so.

Provenance and references

Pages published to your site carry schema.org JSON-LD saying when they were drafted and published and which sources they rest on. On your site the record credits your organisation as the author, never a person; dates never observed are omitted, and ClaimReview is never emitted. Numbered references and a Sources list appear in the page itself. Blogger posts and hosted cloud pages carry the same record, with the person you set as the project’s author as their author when you set one (a blog or cloud storage served from a domain of yours is your own site, so there it is your organisation), most of them with dates and publisher but no source list, because they are not written from fetched sources. Wix stores posts as structured content with nowhere to put the block, so a Wix page carries the same sources as visible links.

Review before publishing

On by default for new accounts; switch it off if you would rather pages go straight out. With it on, every page written for your own site — by the Evidence Engine or on a schedule — waits for your approval, and you are told when one is waiting. Google Docs, Sheets and Slides have a switch of their own, per project: with it on, each is written into your own Drive but not shared or counted until you publish it.

Crawler reality

Your pages are fetched as the real bots, and a throttle or timeout is reported as inconclusive, never as a block. Install our WordPress plugin, or connect the Cloudflare zone that serves your site with a read-only token you create and can revoke, and we read which bots fetched which of the pages we published there, how often, and when last, with “nothing yet” against the pages nothing has come for. Never fetched and fetched-then-ignored are two failures with different fixes, and they used to look identical. Beside each page’s crawl stage you also see whether Google has it indexed, and whether OpenAI’s or Claude’s search returned it for an answer that did not cite it — looked at, not cited. Connect Bing Webmaster Tools, read only, and each page we published on a site Bing has verified shows Bing’s index status too, on its own line.

Plans: Crawler probes and the WordPress plugin’s crawler log on every plan. The Cloudflare read from Growth. So is the detailed crawler feed: your press makes a Logpush job in your own Cloudflare account, which Cloudflare may bill.

Measured on our own account: Nine AI or search crawlers, by the names they give — ClaudeBot, Googlebot, Meta-ExternalAgent, OAI-SearchBot, Amazonbot, Bingbot, GPTBot, PerplexityBot and ChatGPT-User — are recorded fetching pages on our own site in our Cloudflare zone log, and ClaudeBot fetched all 30 live pages our Evidence Engine added between Sep 7 and Oct 5, 2026 (read Oct 6, 2026).

Citations observed, never promised

A later AI answer — a visibility run’s, a tracked prompt’s on your own engines, or a Google AI Overview reading — that cites a published item is recorded against it, with the engine and the date. Google files are matched on their file id however the link was written, declared source lists are read alongside the answer text, and Gemini’s temporary redirect links are followed once to the page they name. “Never cited” is an honest value, and a nightly catch-up re-reads stored answers whenever matching improves. Each gap whose evidence was published shows “Cited in X of N runs that asked about it since publishing”, with the engines — an association, not proof the page caused it.

Measured on our own account: Google’s AI Overview was the first to cite 7 of the videos on our own channel that AI answers have cited, published Sep 10–21, 2026 and cited since Sep 19 (read Oct 6, 2026).

Reviewed on, earned

Pages come due for review every 3, 6 or 12 months, as you set, and the ones due are listed for you. A “Reviewed on” date is recorded only after you confirm the page’s links, figures and dates — never by re-dating text that has not changed. Once a review is confirmed, the page’s structured data names your project’s Content Author as its reviewer, for as long as its text is the text that was reviewed.

Plans: Every plan.

A before and after, if you want one

A gap found by the scan has no reading from before its page existed, so a later run can show a level but not a change. Switch on before-readings and each new gap gets one small visibility check on its subject before anything is published — its two standard questions, asked of your watchlist’s engines — so the note later reads before and after, each with the number of answers behind it. Switch on the passage check as well and, before a page for your own site goes out, your cheapest engine reads each section on its own and says whether it stands alone, and any that does not is flagged in the gap’s notes rather than rewritten. Neither holds anything up.

Plans: Both off by default, asked as Turn on or Pass when you set up a project, on your own AI key: before-readings up to 8 calls with 4 engines, at most 5 a day per project; the passage check one call per section, at most 12 a page.

What is working

The outcomes report leads with the work done, then what happened to it — published, indexed, cited — per site, from the same records the project cards read. A blocker names its cause. Recipe outcomes are measured the same way: items too young to have had a chance are excluded, anything never checked is left out rather than counted as a miss, and below ten mature items you get counts instead of a percentage. By ownership, a piece is on your own site, on a platform we publish to, or not attributable where we cannot place it — never under independent, because every row is our own work. The weekly email opens the same way, and adds what was crawled where we can see your site’s crawler log (your Cloudflare zone or our WordPress plugin). Upload Bing Webmaster Tools’ AI Performance export and the report shows Microsoft’s own count of citations of your pages in its AI answers (Copilot, Bing’s AI summaries and partner integrations, combined) on a card of its own beside Google’s — counts only, never a rate.

Entity Health you can see

Five gauges per project: how much of the expected evidence exists, how many of the gaps found are closed, the score from the latest visibility run, how often that run’s answers cited you, and the change since the run before. A gauge with nothing behind it shows a dash and says what is missing, never a zero.

What needs you, with a way to end each item

Site edits, placements, replies and products only you can explain arrive as a list per project. Each item can be marked done with the address of what you published, declined for good, or retried where the engine gave up. A dashboard note lists what is new and what needs you, each line linking to where it is done, and a health note says whether the outside services your account runs on — AI providers, the results provider — are working, with grey rather than green when the check itself could not be reached.

Recommendations from a nightly review

Every night, with no AI, each project is read for what is worth changing: tracked keywords that match no topic and never rank, topics a watchlist reaches too rarely, a key out of credit, a retired model or a cheaper one of the same maker and tier, a crawler robots.txt blocks, pages held for a week, short topics worth asking as a customer would. Each is a title and one sentence with the figures under Details, on the project page’s To do list and on the page it is about, and nothing changes until you press a button. An AI review on your own key - on request, or weekly if you switch it on - orders the three to do first and adds no figure of its own.

Plans: Every plan. The nightly review is free; the AI review states its key, model and cost before it runs.

MCP server

Mint a scoped key in Settings → Advanced → Agent access — named, shown once, revocable — and point any MCP client at it: Claude Code, Claude Desktop, the Claude API, or your own tooling. Or have Claude connect by signing in instead: add a custom connector at https://therankingfactory.com/api/mcp, sign in, and choose whether it reads every project or one; a connection given one project is matched to it exactly and reads nothing else. Fourteen read tools across five scopes you grant separately: visibility runs and per-topic measurement, the sites engines cited, the gap list, the published-evidence ledger and a topic’s evidence pool, answerability, what is waiting for your approval, and the entity with its claims and sources — always on, for either kind of credential. Every rate arrives with its sample size; every answer says which project it covers; a project that does not exist is an error naming the ones that do. It only reads unless you tick “Let it make changes” when you make the key or the connection, which is off by default: with it, six more tools correct the entity, say what a product of yours is, ask the engine to work a gap now, approve a page or a held Google document that is waiting for you, spread your held Google documents over the coming days, and start a visibility run — through the same checks as the app, one project at a time, audited and capped at 30 writes an hour, with every attempt listed in Settings. A second server, https://therankingfactory.com/api/mcp/check, needs no account: its one tool runs the free site check.

Plans: Agent keys and Claude sign-in are both on every plan, Starter included, and in the trial.