How to tell what AI search actually shows about your business

18 September 2026 · Chris Dolan

The short version

You can see three separate things about how AI search treats your business: which sites AI answers cite when asked about your market, how often Google's own AI surfaces showed your pages, and which AI crawlers fetched the pages you publish. Each is measured on its own and reported with the size of the sample behind it, because an average of different instruments describes none of them.

The second of those arrived this week. Google counts how often its AI Overviews and AI Mode showed a link to your pages, but it offers that count only as a download from Search Console, with no way for software to fetch it. So the product now takes that download once a week and charts it. It is Google's count, not ours, and it cannot be tied to a search phrase, because the file contains none. Expect days of zero: on our own two sites, 33 and 58 of the 92 days exported carried any AI impression at all, and the screen shows the zero days rather than hiding them.

Two other changes make the numbers easier to trust. Every site that AI answers cite can now be listed on one screen, Sites AI cites under Reports & Tracking, and labelled by what it is to you: a stockist, a competitor, a directory or review site, a community, a publisher. Your own other websites can be labelled "Ours", so they stop counting as outsiders who happen to mention you. And the competitor report, which is built from Google's first page, now says plainly that ranking there and being cited by an AI answer are different questions.

A small note on the dashboard shows whether the services behind these numbers are working: green when they are, red when one needs you, and grey when the check itself could not run. Your AI keys are checked with each provider directly, and because most providers publish no balance, the note says when it cannot confirm credit rather than showing a tick. A run that did less because a key had stopped working should not be read as a real fall.

What no screen shows yet is how many visits, enquiries or sales an AI answer sent you. We do not claim it.

AI assistants reviewing the product in September said it could not see the retrieval step, treated AI Overviews as a simple yes or no, offered no access for AI agents and covered forums, video and review sites poorly. None of those was true.

The technical detail

Every week an AI assistant with web search reviews the product and lists what it lacks, and we check each claim against the code. Five September claims about measurement were wrong; two were fair and are now fixed.

"No native agent or MCP posture yet" (review of 14 September)

What is true. A read-only MCP server has answered at therankingfactory.com/api/mcp since 28 August. Its public server card lists ten tools across four read scopes: visibility, gaps, evidence and entity. A customer creates a token in Settings → Agent access, chooses its scopes and can revoke it at any time; the token opens that endpoint and nothing else. Since 18 September, on the Agency and Enterprise plans, an assistant such as Claude can also connect by signing in: the customer approves it on a consent screen, chooses whether it may read one project or all of them, and can disconnect it from the same settings page. A connection made that way reads the same endpoint and nothing else.

How to check. Fetch therankingfactory.com/.well-known/mcp/server-card.json and therankingfactory.com/.well-known/auth.md.

What the review got right. No agent can publish, approve or change anything through the server. That is deliberate for now.

"Does not measure whether a source was retrieved" (review of 17 September)

What is true. Where a customer connects the Cloudflare zone that serves their site, with a read-only token they create, the product reads each night which crawlers fetched which of the pages it published on that host. Sixteen crawlers are named. Three of them, ChatGPT-User, Claude-User and Perplexity-User, fetch a page live to answer somebody's question and count as the Retrieved step; the others count as Crawled. A page nothing has fetched shows "nothing yet". Citation is measured separately, from the answers engines give.

Limits. The site must sit behind Cloudflare, only pages we published on that host are covered, the edge keeps about eight days, and a crawler's name can be forged. Whether an engine read a page and chose not to use it cannot be observed from outside the engine. It is on Growth and above.

How to check. The roadmap entry "Which AI crawlers came for the pages we published", and the list headed "What these numbers do not claim" at therankingfactory.com/how-we-measure.

"AI Overviews are modelled as citation and presence, not patterns" (review of 17 September)

What is true. Google's AI Overview is measured as its own surface, from a live results page for the searches a customer chose, and never averaged into the AI Visibility Score. Each watched search keeps a source history: which sites the overview drew on, in how many readings out of how many, marked as one it keeps returning to, one that comes and goes, or one seen once, with nothing given the first label on fewer than three readings. Each reading also records what kind of page was cited (video, listicle, article, product page, community page and others), counted per reading with its denominator. From those readings a search is called open, settling or held, and the Evidence Engine works open answers first.

How to check. Section 7 of therankingfactory.com/how-we-measure, and the roadmap entries "Which sites the AI Overview keeps coming back to" and "Whether an answer is still open, or somebody already holds it".

"No visibility vector separating discoverability, citation, absorption and outcome" (review of 17 September)

What is true. The AI Visibility page shows the score under "What this score is made of": cited by AI engines 40%, AI crawlers can read the page 20%, machine legibility 20%, crawler permissions 15%, llms.txt 5%. A part that could not be measured is marked "not measured" and left out rather than scored zero. Each engine's answers are counted three ways over the questions asked: named the business, cited it, recommended it. Answers from a model's memory are broken out from answers grounded in a live search. AI Overview readings, Google's own AI impressions and crawler visits are separate instruments and never enter the score. For each site, the weekly email leads with four separate counts: published, crawled, indexed and cited.

What the review got right. How much of an answer's wording came from your pages is not measured, and nor is business outcome. A classifier that could pick AI assistants out of referral traffic exists and is tested, but no traffic is fed to it, so no figure is reported.

How to check. The breakdown panel on the AI Visibility page, and therankingfactory.com/how-we-measure.

"Limited coverage of forums, video and review sites" (review of 17 September)

What is true. Every site that AI answers cite, across an account's projects, can be listed on one screen with its citation count, how many of those answers named the business, and its kind: stockist or partner, competitor, directory or review site, publisher, community, reference, or Ours. A kind guessed from the domain name by a rule is shown as a guess until the customer confirms it, and sites labelled Ours are left out of every third-party rate. The AI Overview history counts video and community pages among the kinds of page cited. Reddit threads worth joining are suggested, and a person writes and posts every reply: the product contains no code that posts to Reddit.

What our own data shows. Across the 20 most recent visibility runs on our own account (31 August to 17 September, seven sites, six of them in marketing software), communities plus directory and review sites made up 87 of 1,749 third-party citations, and reddit.com was not among the 100 most-cited sites. That is one account in one market, which is why each customer sees their own mix rather than a general rule.

How to check. Reports & Tracking → Sites AI cites, the roadmap entry "Every site AI cited, in one screen you can label", and therankingfactory.com/llms.txt, which states that the product does not post to Reddit.

Two fair findings, now fixed

Both reviews asked for Google's generative AI performance data. It has no API: the Search Analytics API accepts no AI search type or AI search appearance, which we checked four times between 21 August and 17 September. Since 17 September the product takes the weekly Search Console export instead (Reports & Tracking → Google's AI impressions, in Advanced mode). It refuses the ordinary performance export by name, since that file carries clicks, CTR and position and would give a confident wrong figure. A day present in two uploads is replaced rather than added. The figures tie to a page, never to a prompt or topic, and are never combined with anything measured by asking engines.

The reviews also said the competitor report was built from page-one rankings while AI answers cite other kinds of source. That was fair. Since 17 September the report says it is a ranking view of a citation question, that AI answers routinely cite pages that do not hold page one, and where to see who is actually quoted.

Where two of the wrong claims came from

The claims about agent access and about governance of machine-written text came from the description of the product we give that weekly review, which had not caught up with those features. It was corrected on 14 September, with tests so the statements cannot be dropped again. An assistant's picture of a business is only as current as the descriptions it reads, and that holds for yours as much as ours.