ChatGPT cites a page when it can lift one complete, self-contained answer out of it — answer in the first screen, scope stated, claims dated, sources named. Structure gets a page considered; it cannot guarantee a citation. Citations are unstable: the same question returns different answers across sessions, so a single sighting proves nothing. Sample the same question three times before treating a citation as real.
Checked- The seven structural signals we check before calling a page citable
- A 40-minute pass you can run on one page yourself, without a tool
- What a single citation sighting does and does not prove
- A promise of being cited, or a rank inside a generative answer
- Google web ranking factors — those are a separate surface
- Prompt engineering for the assistant you are measuring
Why the answer is the unit, not the page
Page analytics describe a page. A generative answer is assembled per session, from whatever the assistant retrieved that time, and it can mention a brand without citing it or cite a page without naming the brand. Impressions and average position do not model that, which is why a page can look flat in search console and still be doing work in answers — or the reverse.
What travels instead is the answer: one question, sampled repeatedly, with the cited URLs kept. That is why the measurement unit here is a question and a dated sample, not a page view. Our definition of AI visibility goes further into what the sample has to contain before it means anything.
The practical consequence is uncomfortable: a page cannot be certified as “AI-ready”. It can only be tested against a question. So the checklist below is written as things you can see on the page, and the results are reported as a dated observation.
The seven signals
These are ordered by how often their absence kills a candidate page in our own audits. None of them is exotic; all of them are skipped in practice.
- Answer in the first screen. The conclusion arrives before any history, framing or “it depends”. A reader who stops after two sentences should still have the answer.
- Liftable sentences. Each claim survives being copied alone: it names its own subject and stays under roughly 30 words. Sentences that begin “This means…” do not survive.
- Dated claims. A visible checked date, in a machine-readable element, not just in a byline. Prices, counts and policies move; an undated number is a liability.
- Explicit scope. The page says what it does not cover. This is the signal most sites refuse to write, and the one that most changes whether a recommendation is safe.
- Named sources. Outbound links to the primary document, with the date each was read.
- FAQ that matches its markup. The questions a reader sees and the questions in the
FAQPageblock are the same words, in the same order of meaning. Mismatch means neither surface trusts it. - One question per page. Two questions produce a page that answers both halfway, and a half-answer is not quotable.
What structure cannot do
Structure is a precondition, not a lever. A well-structured page that says nothing new competes with a hundred others that say nothing new, and being first in that queue is not a strategy.
Three things stay out of reach no matter how clean the markup is. You cannot choose which paragraph gets quoted. You cannot make a citation stick across sessions — the same question can produce a different answer an hour later. And you cannot verify that the assistant read the sources you listed; the link proves you did the work, not that it was used.
How this page is built
This page follows its own checklist, which makes it a cheap test case. The answer block is first; the scope is stated as a pair rather than a footnote; every claim that can move carries a checked date; the FAQ below is generated from the same data as the FAQPage markup, so the two cannot drift apart; and the sources are listed with the date each was read.
What it does not claim is a result. We will re-check the same questions at day 60 and publish whatever comes back, including nothing. That commitment is the product: a dated sample and a source audit, not a rank.
Facts worth quoting
- A citation is a URL, not a ranking — a generated answer can cite a page while never naming the brand on it.
- Three samples is the floor — below that, a sighting is a coincidence rather than a pattern.
- FAQ text and FAQ markup must match — when they diverge, the page is a candidate for neither the rich result nor a clean quote.
- Dating a page costs nothing — and it lets an answer carry a checked date instead of a guess.
A 40-minute pass on one page
The pass, step by step
- 01Pick one question a buyer would actually type
Write it down verbatim. If the page answers two questions, split it before you start.
- 02Move the answer into the first 60 words
Delete any opening that describes the industry, the history or the writer.
- 03Cut the answer into liftable sentences
Under about 30 words each, subject named, no sentence that depends on the one before it.
- 04Add a checked date and a scope pair
A visible date in a time element, plus two short lists: what the page covers, what it does not.
- 05Make the FAQ and its markup identical
Generate both from the same data, or paste the visible questions into the JSON-LD and diff them.
- 06Sample the question three times, then at day 60
Keep the raw answers and the cited URLs. Compare the second sample to the first, not to a target.
What a checklist can and cannot verify
| Signal | What it looks like on the page | We can verify | We cannot |
|---|---|---|---|
| Answer first | Conclusion inside the first 60 words | Yes — by reading it | Which paragraph the assistant used |
| Liftable sentences | Claims that survive being copied | Yes | Which sentence made the shortlist |
| Dated claims | “Checked 31 Oct 2026” beside the number | Yes | Whether the answer carried the date |
| Explicit scope | A “does not cover” list | Yes | Whether scope changed the recommendation |
| Named sources | Outbound links with access dates | Yes | Whether those sources were read |
Questions
Does FAQ schema make ChatGPT cite my page?
No, not on its own. Schema removes ambiguity for crawlers and can earn a rich result on Google, but a citation depends on the answer being useful enough to lift. Pages with clean markup and a vague answer are not cited more often than pages with neither.
How many times should I sample the same question?
At least three, on the same model, the same day, with the same wording. Answers vary by session and by what was retrieved, so a single sighting is noise. Compare sample two and three to sample one, and record the cited URLs each time.
Can I check citations without a tool?
Yes. Ask the question, then ask the assistant to list the URLs it used, and write them down with the date. It is slow and manual, but it is the honest starting point, and it is what most paid tools automate rather than replace.
Does a page need to rank on Google to be cited?
No published requirement says so, and the two surfaces retrieve differently — a page can be cited without ranking, and ranking does not guarantee a citation. Treat them as separate measurements rather than one funnel.
Is llms.txt required for AI visibility?
No. It is a convention some sites adopt; there is no evidence it grants citations. It is cheap to add and harmless to skip. What it does and does not do is covered separately.
How long before a structural change shows up?
Unknown, and it varies by question. This is why the diagnostic re-runs the same questions at day 60 and reports the variance rather than a before/after number. Structural changes are cheap enough to make without waiting for proof.
Sources
- schema.org — FAQPage type definitionAccessed 2026-10-31
- Google Search Central — FAQ structured data documentationAccessed 2026-10-31
- llms.txt — the proposal as publishedAccessed 2026-10-31