Product roadmap
AI search is changing faster than any twelve-month plan survives. So we publish direction, not dates: what has already shipped, what we are building now, and what we are still working out.
Every week the platform audits its own capabilities against how AI search is evolving. This page is what that review keeps telling us.
Actively being built.
Why AI picked who it picked
Partly here. Now that engines are asked to search, Gemini, OpenAI, Perplexity and — since 7 September — Claude hand back the pages they actually read, so the sources behind an answer are the pages themselves rather than links guessed at from the prose. Gemini was the exception until 11 September: it names each source only as a temporary Google redirect carrying the bare site, so its citations recorded a domain and never a page. Those links are now followed once, as the answer arrives, to the page they name. What is still to come is the rest: the exact wording lifted, and that history kept per page over time — so you get "change this sentence on that page", not "your score moved". Claude already returns the quoted sentence and we do not yet store it; nothing anywhere can tell you which sentence on which page was used.
Published work that can be checked
Partly here. Pages the engine publishes to your own site now carry machine-readable provenance — when the page was drafted, when it was published, and a linked list of the sources behind it, in the form an AI system reads rather than a sentence it has to take on trust. We credit the business, never a person who did not write it. Three things are still to come: the same block on the Google-property and cloud assets, the same block on Wix (whose structured post format has nowhere to carry it, so those pages rely on their visible source links), and sources attached claim by claim rather than page by page.
Content built for retrieval
Partly here. AI systems answer from passages, not pages. Every guide, checklist, FAQ, article and own-site page the engine writes is now told to open with the answer, in one or two sentences a reader could quote on their own, and to carry real questions each answered so it reads away from the rest of the page — and it is enforced rather than requested: a model that still opens with "In this guide we look at…" has that sentence cut before publishing. Three things are still to come. The data tables and calculators we publish as Sheets get no such instruction. The sections in between are passages in shape but are not asked to stand alone. On domains enrolled in it, the shape a page must take is now decided before it is written, and each element it promised is checked by quote against the published page. What is still missing is the harder claim: nothing yet splits a finished page the way a retrieval system would and checks that a piece still answers on its own.
The Evidence Engine loop, and its last link
Partly here. Finding what AI search is missing about you, generating the evidence and publishing it all run on their own — the scan hourly, the closing cycle every three to six hours depending on plan, on Growth and above; below that it runs when you press the button. The queue is now ordered by whether the answer is still open, then by how close the question is to buying, then by impact, with each cluster of related topics worked together. The last link is what is being built. Every closed gap is re-measured on later visibility runs, but a before-and-after needs a before, and only a gap that came from a visibility run has one — the gaps that dominate real accounts come from the completeness scan instead, so for those we can tell you the level now and not that it moved. Whether an engine goes on to cite the page is a separate and slower measurement again. We do now observe those citations — since the counting faults were fixed on 11 September, evidence the engine published has been credited as cited and shows on the report — but observing a citation is not the same as attributing it to the gap that produced it, and that attribution is still the missing half.
A triple store of your facts that stays current
Partly here. Your entity — the business, its identifiers, the sources that have cited it — is written out as subject, predicate and object rows into a Google Sheet, and rewritten whenever you edit the entity or the engine publishes something new, so a Google Site with that Sheet embedded shows the change without anyone touching it. Three things are still to come, and this page used to claim all three. The Sheet has to be linked to the entity before any of that fires, and today that only happens if you load an Entity Factory profile when you run Entity Stack. Nothing propagates past that one Sheet: every Doc, deck and page carries the facts baked in as text, deliberately, because an embedded frame contributes no readable words to the page holding it. And your Wikidata identity is a batch we draft for you to submit — we neither create nor maintain the item, and until this is finished it is remembered in your browser rather than on your account.
Committed direction, in design.
A managed signals layer
Today we generate the files — a full robots.txt with an AI-crawler policy, a starter llms.txt, the schema block — hand them to you to install, then check them every hour and tell you what is missing. The missing verb is deploy: putting them on your own host and keeping them right, instead of checking them and leaving you to it. Two things go with it. We audit sites for a Content-Signal line and generate none, which we should not be doing. And nothing advertises your resources over a standards-based Link header on your site — we set one on our own responses, and that is as far as it goes.
Wider first-party publishing
The engine publishes to WordPress, Ghost, Wix or any webhook today, and where a CMS is connected a subject closed on a Google property gets a follow-on page on your own domain. So the Google stack is still the default output and your site the second step. Next is turning that around: your own domain as the first placement rather than the follow-on, and more direct connectors — Squarespace, Shopify and Webflow have none and can only be reached through the generic webhook.
Earned-surface coverage
Community discussion is the one we work: we find Reddit threads already asking what you answer and hand each to you as a task, because an automated reply is what those communities remove. Review platforms and independent coverage we do not work at all — we can tell you an engine cited G2 or a trade publisher, and that is the whole of it. Next is working those surfaces rather than only counting them when a citation happens to land on one. Video sits outside this: the explainer we generate goes unlisted to your own channel, which is first-party by design and not coverage anyone earned.
Attribution to real outcomes
Connecting AI visibility to sessions and conversions that arrive from AI assistants, so improvement is measured in business terms. None of it exists yet: we read no referrer, store no visit, and connect to no analytics account. When it does, it will be reported as a floor and never as a total, because several assistants send no referrer at all.
Map Pack position, measured rather than implied
The local pack is the one local surface we do NOT measure, and the local-SEO page implied otherwise until 7 September 2026. The honest position today: we build the local entity signals - GBP posts, geo-targeted assets, consistent details across the properties Google reads - and those are inputs to local ranking, but we have never reported your Map Pack position and the detector we had could not fire. Our search provider does not return the local block on the endpoint we use; it lives on a separate places lookup, one extra call per local-intent search. So this is a cost decision rather than a research problem, and the shape it would take is the shape everything else here takes: your position recorded over several checks with the sample size attached, never a single reading presented as a rank. Wanted by enough local businesses that it is on this list rather than buried; say so if it is what you would upgrade for.
AI access as a deliberate policy
Retrieval and training are different permissions, and today most sites grant or refuse them as one undifferentiated list. We already treat them apart everywhere we read: the robots.txt we hand you allows the search crawlers and blocks the training-only ones by default, and the two kinds of coverage are scored separately. What is missing is the surface where you make that choice yourself, per crawler and per purpose — until then it is our opinion in a template, which is still a default you inherited rather than a policy you set.
Researching. Genuinely uncertain — included for transparency.
Being callable, not just citable
If assistants start completing tasks rather than answering questions, brands may need to expose structured, machine-callable facts and actions. We are prototyping rather than betting the product on it.
Drift detection
Noticing when an engine changes how it sources answers and adapting automatically, instead of repeating a tactic that quietly stopped working.
Answer-accuracy monitoring
Alerting when an AI states something wrong about you — outdated pricing, discontinued products, plain invention — and helping correct the record.
Broader knowledge-graph presence
Extending entity work past Wikidata into the wider set of sources engines reconcile identity against, with alerts when your record is edited.
Nothing above is a dated commitment. Items move between columns as the engines change, and some “Exploring” work will be dropped if the industry moves elsewhere — we would rather tell you that than publish a roadmap we quietly abandon.
Live in the product today — most recent first.
On Agency and Enterprise, Claude — and any AI client that supports MCP sign-in — connects to your account without a token to copy. Add a custom connector with your account’s MCP address, sign in, and choose what it may read: every project, or one. A connection given one project reads that project and no other. It is matched to that site exactly, not to anything whose name contains it, and work the account never attributed to a project is kept back from it, with a note saying how much. Underneath it is OAuth with PKCE: access that lasts an hour, refresh tokens that change on every use — present an old one and the connection ends, because only a copy would do that — and a list in Settings → Agent access where you can end any connection at once. Tokens minted in Settings still work, on every plan with agent access.
A page on your own site is assembled from the Docs, Sheets and decks built for its subject — and until now that happened once. A part that arrived after its page, usually the deck, never reached a reader: of 32 decks in five days, 3 were on a subject that had a page. Now a late part is added and the live page updated in place, in the parts’ own words with no new writing. Three refusals make that safe: never over a change you made on your site, because every sentence we published must still be there and nothing you added may be in the article (on WordPress, anywhere in the post); never over a page that was rewritten as one text; and never without you on an account that reviews first — there the update waits on the project page for your approval. It also gives the page a date it never had: when it was last modified.
We already found the pages AI answers and search results cite on a subject, and checked every address still loaded — then handed the writer a list of titles. A writer given a title can only say about that page what it already believed, and the claim check then cuts the sentence when the belief is wrong, so the correct claim never gets made at all. Measured on one subject: eight relevant pages in the pool, and the two pieces published about it cited none of them. The part of the three strongest that is about the subject now goes in front of the writer, chosen by relevance rather than by the order they were found; every other source is still offered by name and address, which costs nothing. A sentence that reproduces a source rather than reporting it is cut before publishing — a quotation is the point of the feature, the same words with the marks removed are not. This is how everything is written now.
Everything else here asks an engine a question and reads the answer. This is the one figure Google reports rather than we measure — and there is no API for it at all, which we established four separate times before building anything: the Search Analytics API accepts no AI type and no AI search appearance, and the BigQuery export does not carry it either. So the report is exported from Search Console and uploaded, weekly, and the two measurements are never averaged into one number. It carries no search terms, so a figure ties to a page and never to a prompt, a topic or a gap, and the screen says so before you upload rather than after you go looking for the join. Most days on a real property read zero, and a zero is a reading rather than a gap.
An account runs on outside services — AI providers, a results provider, an AI Overview provider — and until now the first sign that one had stopped working was a run that quietly did less. Your AI keys are now checked with the providers directly, so a revoked or disabled key is caught before anything tries to use it rather than after a run has failed. Where a balance can be read it is read, free, and you are warned while credit remains rather than once it has gone. Where it cannot, the note says so instead of showing a tick it cannot justify: most providers publish no balance anybody can ask for, and pretending otherwise would be the comfortable lie. A refusal is never reported as an empty account unless the provider said so — a rate limit is not depletion. It is always on the dashboard: green when everything checked is healthy, red when something needs you, and grey rather than green when the check itself could not be reached — “we could not look” must never read as “all is well”.
The dashboard counted the sites AI cited across a whole account, while the only place you could label one showed a single project at a time — so you could label everything in front of you and watch the number barely move. They are now in one place, searchable and filterable, with the ones still waiting on you marked. Two things came out of building it. A guess from a rule is shown as a guess on the row rather than hidden behind a label that looks like your answer. And “ours” is a label now: an account’s other sites were being counted as third parties, which quietly inflated the rate at which stockists were said to name the business — on our own data, 117 of 134 citations, most of them us naming ourselves.
Re-align rewrites a page already published so that what it says about a product agrees with the owner’s own description of it, keeps the page’s structure, and stops at review. No line from the owner, no rewrite. Alongside it, review-before-publishing became a setting that covers every page the engine writes for your own site rather than some of them, and the nightly answerability grade became opt-in because it spends your AI credit — the card says when it is off rather than promising a sweep that is not running. The Docs, Sheets and decks that go to Drive are still not held: they are supporting signals, not posts on your site.
We publish to Google Docs, Sheets, Slides, Blogger and an entity hub, and then measured citations with a test that could not see any of them. We compared the exact address we stored against an engine's wording, and no engine writes a Doc link the way Google hands it back. We read only the body of an answer, never the source list the engine declares beside it. And Gemini names each source as a temporary Google redirect carrying the bare site, so a customer's own post read as "blogspot.com". All three are fixed — Google files matched on their file id however the link is written, declared sources read alongside the text, redirect links followed once to the page they name — and an AI Overview citing a page we published is counted as its own state rather than as absence. It is not a new measurement: it re-reads answers already collected and paid for, every account nightly. The numbers move upward, which is why the catch-up uses the live path's own matching rather than a looser rule written for a backfill: an over-count is exactly as dishonest as the under-count it replaces.
An AI Overview depends on where the search is made, and ours passed no location at all — so every reading for every account was taken from the United Kingdom. For a business outside it that 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 previous default, kept deliberately so existing history stays comparable rather than moving underneath it. Each reading records the vantage it was actually taken from, so it is reproducible. It refuses to guess one: a business type or a domain ending is not a place, and where two projects disagree it asks rather than choosing.
Some of what the engine built was left private — a hub document, a data sheet and a deck per keyword cluster, none of them shared or published — so a stranger following a link met a request-access screen and the customer got the email. A file nobody can open is a file nobody can cite, which is the whole reason it was published. Both create paths were fixed, but the durable half is a source rule rather than two fixes: a build-time check reads our own code and fails if any path creates a Doc, Sheet or Slides without also sharing it, because the same class of fault — a rule applied to one path and not its siblings — has been found here repeatedly. A share that Drive refuses is no longer recorded as one, and the backlog is repaired by asking Drive about each file rather than believing our own ledger, which is what let this run unnoticed.
The one half of the proposition nothing measured. From readings already taken: no source recurring means the answer is not yet anybody's, one or two means a set is forming, three or more means the engine has an answer it returns to. Under three readings it classifies nothing — calling an unmeasured topic open is the single way this could push somebody into publishing on a guess. Where an answer is held it names the sites holding it with their denominators, and offers two moves without choosing either, because which is right depends on how much that ground is wanted. It never claims a rewrite would succeed; "publishing more of the same has not moved it" is a statement about the past, which is the strongest thing the readings support. The Evidence Engine works open answers first and held ones last.
Two phrasings of one question were two topics, each measured separately and each counting against the limit. Grouping is proposed from data rather than a model — the pages Google ranks for each phrase, and the sources its overview draws on — and the proposal shows its working, because the shared pages are the reason. One press sweeps the whole list rather than proposing one at a time, capped at twelve, since 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 rather than merged: a child keeps its measurement and history, stops counting against the ceiling, and can be detached. The engine then works a cluster together, parent then children, instead of leaving a child several cycles behind.
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 the same table four times. The project now carries its own Search Console section and its own keyword table — every keyword the site's campaigns track, one row each, with its own import — while campaigns still own the keywords. We also stopped asking Google for whichever property a connection happened to have selected and ask for the site's own property, so a site genuinely held no longer reads "no property for this site" beside a panel showing that site's clicks. A reading pulled from Search Console now counts the same as a nightly rank check and can mark a target met.
The dialog promised campaigns "paused until reconnected"; what it did was delete every record of what that account had published. On one account a database error stopped it, which is the only reason it was found. Disconnecting now does what it said — campaigns pause and keep their link, the account goes inactive and hands back its tokens, nothing is deleted — and reconnecting the same account finds it by email and takes a fresh token.
Published → Crawled → Retrieved → Cited, and only the last step had ever been watched: we could fetch your pages as GPTBot, and we could not tell you whether GPTBot had. Connect the Cloudflare zone that serves your site — your zone, a read-only token you create and can revoke — and every night we read which bots fetched which of the pages we published there. You get the page, what crawled it and how often ("3× GPTBot, ClaudeBot"), 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 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 pages we put on Google Docs or Blogger sit on somebody else's edge and are counted apart rather than reported as uncrawled. A fetch is not a citation and is never shown as one. Growth and up.
A real link to a real standards body with an invented percentage attached is fabricated evidence wearing a verified address, and the fetch-before-publish gate cannot catch it — that proves an address answers, not that the page says what we said it says. Every writer now asks for outside sources by name, and before publishing, any figure in a sentence that names a source is looked for in that source's own words: where none of the sources that sentence names carries it, the sentence is cut, the way an unsourced superlative is cut. A source listed but never used by the text is dropped from the list rather than left there to pad it. A source we could not read cannot be checked, and it says exactly that instead of passing.
A stockist names you in nearly nine answers in ten and a competitor's guide in one in a hundred, and nothing told those apart. The tempting fix was to infer the kind from the naming rate, which is a circle, so instead you label each cited source yourself — stockist, competitor, directory, publisher, community, reference. We pre-fill only what a domain name can honestly carry, shown as a guess in amber, and never guess stockist or publisher. The naming rate then appears per kind as counts with their denominators, on the page and to your own agent.
Bing — and so ChatGPT search and Copilot — only accepts a "this page is new" notice when a key file is served from the site it is about, and ours were being rejected with a 4xx nobody read. We now place that key file on your WordPress site through its own media library, fetch it back over HTTP to prove the host really serves it, and record where it lives; from then on every page we place there is submitted with it. If the file lands on a CDN host rather than yours, that fails and says so, because a key served from somewhere else does not count. For any host we cannot write to, the key and the file name are on screen to place by hand.
Your account answers an AI agent directly now. Mint a token in Settings → Agent access, point Claude — or anything else that speaks MCP — at it, and ask in plain language: what are my open gaps, which topics am I measured on but never named for, who is getting cited instead of me. It covers your visibility runs, per-topic measurement, the sources engines actually cited, your gap list and the evidence you have published. It reads and never writes, and the token opens that one door and nothing else — it is not a login, so it cannot be used against the rest of the app, and you can revoke it whenever you like. Two things travel with every answer on purpose. The sample size comes attached to every rate, so an agent quoting "45% cited" has to quote what it was measured over. And every answer says which project it covers: ask about one client and you get that client, ask across the account and it tells you that is what you got — an account-wide total is not a per-client number. Ask about a client you do not have and it names the ones you do, rather than returning an empty list that reads like no work. Growth and up, and included in the trial.
The engine builds several objects for a subject — a guide, a specification, a checklist, a data table, a deck — and the page on your own domain is the placement that matters most. That last step had largely never happened: across live accounts, 493 of 505 subjects with published evidence had no page on the customer's own site. The Evidence Engine now lists them per project and publishes one in a click, assembled from what was already written rather than generated again, so it costs nothing and produces no second article competing with the first. It refuses two things on purpose: publishing them all at once, and offering a subject your own site already answers — a page of ours competing with a page of yours for the same search helps nobody. If you wrote that page yourself, paste its address and it is recorded as yours.
Assembling that page needs the words as written, and for most of the back catalogue only the link had been kept. The documents are still in your Google account, so they are now read back a few hundred a night — no regeneration and no AI spend, because the words were written once and paid for once. Older subjects move from "cannot be assembled" onto the actionable list as they arrive. Anything since deleted or unshared stays put and says so rather than being quietly rewritten.
A study of 304,805 cited URLs scored content traits with schema deliberately excluded and found the correlation sits in the visible text — a clear answer at +32.8%, useful Q&A at +25.5%. Not an argument against markup we already emit; an argument that markup is not sufficient. Generated evidence now opens by answering the question rather than describing the page, and carries real questions and answers as text rather than only as hidden markup. A model that still opens with "In this guide we look at…" has that sentence cut before publishing rather than a second generation bought to replace it.
Each gap names the evidence it wants. When one was published, every open gap on that subject closed regardless of what it had asked for — so a checklist could read as done because a specification existed, and the number meant to mean "AI can now see this" counted work nobody had done. Gaps now settle only against the object they requested. The visible consequence is that some reopen, which is the correction rather than a fault.
Open any watched search and you see every site Google's overview has drawn on, in how many checks out of how many, labelled as one it keeps returning to, one that comes and goes, or one seen once. Your own domains are marked. "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. Nothing is called a site it keeps returning to on fewer than three readings, and the churn caveat travels with the list — most of what moves there is the surface changing rather than your site, and we put no figure on how fast because the published measurements disagree.
Where Google shows an overview and does not name you, the row hands that search to the Evidence Engine as a topic — the same route the competitor-gap rows use, with the same rules. Where it is already a topic in that project, we say so instead of adding a second copy. That is a real answer rather than a dead button: it tells you the evidence exists and is not being chosen, which is a different problem from a missing topic, and the source list beside it is where you look next.
How many searches you can watch comes out of a monthly spend on your own DataForSEO credit — and you can set it lower than your plan allows. How often to check is yours too: weekly keeps the list long, daily gives each search a firmer rate on a shorter list, and both figures are shown before you choose. One ceiling is ours rather than your budget's, and it says so rather than sending you to raise a number that would change nothing.
The most common thing said about this product by AI systems reviewing it was that we did not measure the AI Overview, which is where most people meet an AI answer at all. We do now: the searches you choose are read from a live Google results page, and you see 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 engines directly; this is a live page with its own retrieval, and averaging the two would produce a number describing neither. Checked weekly, or daily if you choose — an AI Overview is generated and two looks on the same day can genuinely differ, which is an argument for more readings rather than fewer. What daily costs is list length, not accuracy. Either way you get counts across several checks with the sample size, never one reading dressed as a fact.
The competitor report could always tell you a rival was on page one for something you were invisible for, and then it stopped — you wrote the keyword down on paper. Each one now hands straight to the Evidence Engine: it becomes a topic on that project, gets scanned for what is missing, and the page is written on your own site. And where a search is really a longer way of asking something you already cover, you can file it under that topic rather than starting a new one — a topic earns its own page, a question shapes an existing one, and we say which you are getting.
The SEO audit, the GEO audit and the visibility check all take the same input and used to hand back three reports in three vocabularies, leaving you to work out which to act on first. They now run together from one web address and reduce to a single prioritised list, with every finding linking through to the audit behind it so a one-line summary is never the end of the trail. It does not dead-end either: the result finishes with one onward step — set the site up as a project, or go to where the work happens if it already is one.
The site check reads one address and reports on the domain, so it could never answer "which of my pages is the problem". This is that half. Template problems lead, because a finding on thirty pages is one theme edit and burying it in thirty identical rows is the failure this exists to fix. A page we could not read gets no score rather than a zero — a zero reads as a terrible page instead of a missing measurement — and where your site is bigger than the scan, the report says so rather than presenting part of it as the whole. Growth and up: 15 pages a scan, 40 on Agency, 100 on Enterprise, because every page is a real fetch against your own server.
Past Wikidata there is a short list of places an engine checks to decide you are who you say you are — LinkedIn, Crunchbase, OpenCorporates, G2, Yelp, Trustpilot, Product Hunt, Wikipedia — and Entity Stack scores you out of those eight, with why each one matters and where to create the listing. Wikipedia is checked live against Wikipedia itself, and where you would not qualify yet we say so plainly rather than sending you to be reverted. The other seven are counted from the addresses you enter, so that score records what you told us rather than a page we fetched and read. Verifying them is work still to do.
Setup used to ask you to type keywords on your first day — an expert decision before you had seen anything the product does. Now we crawl your site and propose the business type, the places you serve and your topics, each proposal naming the page we found it on so you can check us. You correct it; what you leave is what we work on. A site we could not read is reported as exactly that rather than filled in with a plausible-looking guess.
It used to run first, on our guess at your business — so every later comparison was against a starting line drawn in the wrong place. It now runs on the topics you confirmed, with the question count shown before anything is spent on your key. Under twenty questions it calls itself a first look rather than a measurement, and says what would make it a baseline — because six questions carry a confidence range where 30% and 65% are the same reading, and we used to show that as where you stood.
The engine now publishes without waiting for approval, and the weekly report opens with what it did, item by item, with links. Prefer to approve first? Turn on review-before-publish and every page written for your own site waits for you, and you are told the moment one is. Either way the correctness guards still apply: nothing unsourced, nothing ungrounded, no link a stranger cannot open.
How we measure — the same prompts to every engine, 95% confidence ranges with their sample sizes, repeat sampling reported as stability rather than certainty, engine sets that never mix, and what is excluded and why — is now a page anyone can read, including what the numbers do NOT claim. Nothing on it is aspirational; every sentence describes something implemented and tested.
A business with eight locations used to mean eight separate setups. They now go in together on the project, each able to become its own campaign in one action. A division is kept firmly apart from a location: "Commercial" is not a place, and we will never track "best plumber in Commercial" — a phrase nobody types — at your expense. Agency and Enterprise.
Every source an engine cites is classified: your own site, properties published on your behalf, or genuinely independent coverage. When nothing independent was cited, the run says so plainly — sentiment measured only from your own material is that material read back to you, not a reputation. Ambiguity counts as yours, so uncertainty never flatters the score.
The recurring words engines use about your business, with how many answers used each. Reported as found and never scored — whether "budget" or "boutique" helps is your judgement about your market, not ours.
Built from rank checks you already pay for, so it costs no extra searches: where a rival sits on page one and you are absent, weakest rival first, with the sample size stated — and a warning when it could not recognise your own domains.
The way in is now: check your site, add an AI engine, let the engine find and close the gaps. Picking a recipe and aiming a push is no longer the first thing you meet. Nothing was taken away — if you have campaigns you still see them, they still run, and any account can switch the campaign and recipe screens back on in one click. What changed is which one is the front door.
Your publishing destination and your LinkedIn and Business Profile opt-ins are now properties of the project rather than of a recipe — so they are reachable whether or not you ever use campaigns. A campaign that made its own choice keeps it. This is what turns "your own site first, other properties supporting" from a claim into a setting.
Indexation and citation rates per recipe and per publishing surface, so a recipe can be proved or retired on evidence rather than opinion. Built to be able to tell us bad news: work published too recently to have had a chance is excluded rather than counted as failure, anything never checked is left out rather than counted as a miss, and below ten mature items you get counts instead of a percentage.
Pages we publish to your own site now include a machine-readable record — when the page was drafted, when it was published, and the linked sources behind it — so an AI system can verify it rather than take our word for it. WordPress, Ghost and webhook destinations carry it; Wix stores posts as structured content with no place to put it, so a Wix page carries the same sources as visible links instead. We credit your business, never a person who did not write it. And we will not publish "we are the best in the county" without a source: if the model writes one we ask again, then cut the sentence rather than publish a claim you cannot substantiate.
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. Gemini now searches with Google, and OpenAI looks things up through its web-search tool as it answers — and since 7 September so does Claude. Perplexity always searched. All four report back the pages they actually read. Gemini reports its sources as temporary Google redirect links naming only the site, so since 11 September those are followed to the page they name before anything is recorded. Where an engine cannot search, we label the answer as coming from training rather than presenting it as fresh.
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, in advance rather than afterwards.
These are measured by calling each engine’s API — repeatable, and able to carry a confidence interval — but that is not the same as Google AI Overviews, AI Mode or the ChatGPT app, which are separate products with their own retrieval. Runs now state that plainly instead of leaving you to assume we measured something we did not.
Publishing connections live on your account and you may have many. A project now publishes only to a connection on its own domain, or to one you chose for it — never to whichever happened to be connected first. Where none applies, the page is written and handed to you rather than published somewhere plausible.
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. We do not post for you: an automated reply is exactly what those communities remove, and the citation is only worth having because a person earned it.
The keywords you typed at setup became a living list: what you seeded, what we found, how often each was measured and named, and controls to add, remove or focus. It shows how many topics a run measures, because that spends your own AI credits.
Three bands — what is happening without you, what only you can do, what you could unlock — replacing a checklist that told a healthy account it was 40% finished forever. The first band only claims what is genuinely running on your plan.
LinkedIn and Google Business Profile announcements are off until you turn them on, and each states its volume before you choose. Connecting an account for one purpose should not volunteer it for another.
Asking a chat model what it already knows and asking a search-grounded one what it just found are two different questions on two different timescales. They are now measured and reported separately instead of averaged, each engine trends on its own line, and runs that did not record which they measured say so rather than guessing.
AI answers are not deterministic, so each prompt can be asked several times per engine. You see which prompts the engines disagree with themselves about — reported as counts, never as a percentage that hides how small the sample was.
Ask a question from any screen and get an answer drawn from the product documentation, with links to where it came from. Where the docs are silent it says so rather than inventing a settings page — and it cannot run anything for you, by design.
AI answers vary run to run, so the Mentioned, Recommended and Cited rates carry a 95% confidence range and the sample size they were measured over, on the run and per engine — and those three are measured as three separate signals rather than one blurred number. Share of voice, sentiment, the trend line and the PDF exports do not carry a range yet: where you see a bare percentage, it has not been given one, and that is a gap rather than a claim of precision.
Checks how well your site answers to AI crawlers and agents — which named bots your robots.txt lets in, the structured context and schema on the page, and the machine-readable signals an agent looks for: Content Signals in robots.txt, an RFC 8288 Link header, /llms.txt, an MCP server card and the other .well-known files. Agent readiness is scored on its own so unproven extras cannot inflate the headline, and the PDF carries copy-paste fixes for what it found. The report carries our branding: white-labelling reaches the client report and the weekly email, not this one yet.
The engine mines your recent AI answers for topics you are being named alongside, and ones a competitor is cited for and you are not, adds up to five per cycle and shows each as "Found by us" with controls to remove or focus it. It runs on the schedule on Growth and above, spends your own AI key, and needs a completed visibility run to learn from — so a new or quiet account sees nothing until it has been measured, which is the account that would benefit most and is the part we still owe you.
WordPress, Wix, Ghost, LinkedIn, Blogger and any webhook-compatible or headless CMS, alongside the Google property stack and your own cloud-hosted pages on your own domain.
Brand presence, share of voice, sentiment and per-engine citation rates across ChatGPT, Gemini, Perplexity, Claude, Grok and DeepSeek — whichever you connect with your own keys, up to three in a run, each engine on its own trend line. Four of them search the web as they answer. Grok and DeepSeek answer from what they were trained on, so nothing you publish can move their numbers, and the engine picker says which is which before you spend anything — three flat lines chosen by accident look exactly like failure.
Check your own site — the site check needs nothing but a web address, and you will get the same measurements this roadmap is built around. Plans come with a 14-day trial on Growth.