Answer citation checker
One buying question, put to four AI assistants with web search on. Every answer verbatim, the brands each named in order, and the pages the answers were built from.
Write it the way somebody choosing a vendor would — "best crm for a small sales team", not "is Attensira good". A question with a brand name in it measures nothing: the assistants will talk about the brand you named.
Four assistants, one question, every answer kept.
You get each answer in full, the brands each one named in the order it named them, and every page the answers were built from — grouped by site, so you can see which third-party pages decided the result.
It takes a few minutes, so we ask for an address and send you the link. The report itself is a public page you can share with anyone; nothing in it is behind the form.
What this check actually does
It puts your question to four assistants — ChatGPT, Claude, Perplexity and Gemini — once each, through their model APIs with web search enabled, and keeps everything that came back.
- The answers, verbatim. Every reply in full, not a summary of it. Each number on the report is a fold over these, so you can check any claim against the text it came from.
- Who was named, in order. Brand names are read out of each answer's own prose, in the order they appear. That order is the assistant's ranking, and it is the thing a competitor board is actually made of.
- What the answers were built from. Every page each assistant cited, grouped by site, with how many assistants leaned on each one. A page three assistants used is a better afternoon's work than one cited three times by the same assistant.
Named is not cited, and the difference is the finding
An assistant can recommend a brand it never links, and it can cite a page and then recommend somebody else. Those are different facts with different fixes, so this report never merges them.
The second one is the more useful and the more commonly missed: if your own page was cited by an answer that went on to name a rival, the page was good enough to retrieve and not good enough to convince. That is a content problem with a known fix, rather than a visibility problem with none.
This is an API surface, and we say so on the page
We read model APIs with web search on. That is not the same surface as the consumer apps a person types into — different system prompts, different retrieval, sometimes a different model version — and a brand can swing sharply between the two on the same day, on the same question.
So the report labels the surface next to the numbers rather than in a footnote, and we do not average the two. If what you need is what the consumer products say, that is what the paid product reads; this is a real measurement of a real surface, and it is not that one.
One answer each, and why we say so rather than sampling louder
Each assistant is asked once. Running the same prompt repeatedly returns different citations, so a rate computed from a handful of answers is a snapshot rather than a settled fact — and a free tool cannot afford the depth that would make it one.
The honest response is to carry the number of answers everywhere the percentage goes, and to refuse to rank inside the noise. At four answers, a one-place difference in average position is not a finding, and the report says that on the page rather than leaving you to infer it. Sampling once and writing it up as though it were settled would be the same confident-font failure this product exists to argue against.
Why it asks for an address, and what it does not gate
This one costs us real money per run and takes minutes — several live web searches and an analysis pass. So we ask for a name and an address before starting one, and send you the link when it is done rather than making you sit on a loading screen.
Nothing in the result is behind that form. The report is a public page with its own permanent link, shareable with anyone, and it stays that way. The exchange is for the running of it, not the reading — every other tool in the directory answers immediately and asks for nothing at all.
The newsletter tick beside the address is separate and starts unticked. Asking for a report is not agreeing to hear from us again, and the two are separate lists you can leave separately.
Questions people ask about this check
Is this what ChatGPT told a real user?
No. It is what the model behind ChatGPT answered our request through its API, with web search on and a system prompt that mimics the product's answering style. That is a real answer from a real surface, and it is not a recording of somebody's session. Nobody outside OpenAI can show you that, and any tool claiming to is claiming something it cannot do.
Why did my brand not appear at all?
Sometimes because it was not named, and sometimes because the question was not one your buyers would ask. A question with no commercial intent — "what is a CRM" — returns an explainer that names no vendors and measures nothing. Try the shape a buyer types: "best X for Y".
Can I run the same question again?
Yes, and after a week it re-runs and updates the same page at the same link. Inside that window you get the report we already have, immediately — which is deliberate: the answer to your question already exists, and re-running it would spend money to produce roughly the same page.
Who can see my report?
Anyone with the link, which is unguessable and only goes to you until you share it. The page shows the question and the answers; it never shows who asked for it. Your address is used to send you the link and is not on the page.
How are the findings written?
From the same rows you can see, under published instructions — the report links the exact version of each one it used. A finding that does not point at evidence on the page is dropped before the report is stored, so there is no commentary here that you cannot check.
Where to go next
- See the whole picture for your domain, not one questionThe free audit scopes eight buying questions for your domain and reads all four assistants on each of them.
- Check whether the assistants can reach your pages at allBeing cited starts with being fetchable. This makes one real request per crawler and reports what came back.
- Track these answers on a scheduleOne question once is a snapshot. Knowing when an answer changes is the thing a snapshot cannot give you.
- Measure how a page reads to an assistantIf your page was cited and the answer still recommended somebody else, how it is organised is where to look first.
- Read the instructions behind the findingsThe skills that write the findings are published, version by version, so you can check what we asked for.