An AI Citation Isn’t a Ranking: What a Page Must Survive to Get Cited

One groas post received 580 verified ChatGPT-User fetches in 30 days. That is not a ranking, and it is not proof of 580 citations. It is evidence of something more basic: when a live answer needs a page, the page has to be reachable before its words have any chance of appearing in the answer.

A prospect asked me last month how long it would take to rank number one in ChatGPT. I told him there is no fixed number one to rank for. Take the question I will use throughout this piece: What garage door opener should I buy for a cold climate? For a client’s page to earn a citation, a system has to find and fetch it, extract a useful passage, then choose to rely on that passage. If any step fails, the page does not make it into that answer.

I used to describe AI citations as rankings with new rules. I was wrong. Fetch, then extract, then trust is the more useful model. The link at the end is a possible outcome, not the starting point.

First, a page has to be available to fetch

A Google result gives you a list and a position on that list. ChatGPT and Perplexity produce answers that may include links to sources. Treating those links as positions hides the first practical question: could the system access the page when it needed it?

There are two ways a model might know about the garage door opener brand. It may have encountered the brand in training data, or it may retrieve information from the web while answering the question. Those are different paths. One comparison describes ChatGPT as relying more on training data unless it searches, while Perplexity searches the web by default. A brand can be familiar enough to get mentioned without the model fetching the client’s page or linking to it.

That distinction matters to an agency. You can work on a client’s broader reputation, but you cannot diagnose a missing page citation by staring at the answer alone. Start with whether the page could be retrieved.

A familiar brand is not necessarily a cited source

Training data is closer to memory than to an open browser tab. A model may have encountered a brand’s site, a review or a forum discussion before answering today’s question. That background can shape what it says. It does not mean it opened the brand’s current product page or checked a current specification.

For our cold-climate opener, the model might name a familiar manufacturer in a general answer about why cold weather affects equipment. The mention may be useful to the brand, but it tells the client little about whether their page is ready to support a specific buying answer. A mention without a fetch is not evidence that the page worked.

A live fetch gives the page a chance

When a system does retrieve pages, different bots can do different jobs. OpenAI distinguishes GPTBot for training, OAI-SearchBot for search discovery and ChatGPT-User for user-triggered page access. An agency that treats all three names as interchangeable can block the access it wants while trying to restrict the access it does not.

For the opener question, retrieval might bring back a page with an installation specification, a low-temperature rating or a price. Those details are only useful if the fetched page makes them readable. At this stage, I am not promising a citation. I am saying the client has cleared the first gate: the system can reach the evidence. Access is necessary; it is not sufficient.

Next, ask which questions call for retrieval

The same page does not face the same opportunity on every prompt. A broad question about why cold affects garage door openers can be answered from general knowledge. A request for current options, prices or specifications gives a system more reason to check pages. That is the practical difference between writing a general explainer and writing a page that answers a buyer’s specific question.

The two products also behave differently. In one comparison using 35 buyer-intent prompts, a brand appeared with a hyperlink in 61% of Perplexity answers and as a plain-text mention in 8% of ChatGPT answers. That is one test, not a rate I would promise a client. Its useful lesson is narrower: being named and being linked are different outcomes, and the answer system changes the odds.

Freshness can matter when the question asks for current information. The same comparison reports that half of Perplexity citations in its analysis were to pages under 13 weeks old. That does not mean an older page is useless or that adding a new date fixes weak content. It means a client answering a time-sensitive buying question should make the relevant, current facts easy to find.

So keep the two opener prompts separate. Why does cold affect openers? asks for an explanation. What opener should I buy for a cold climate in 2026? asks for a current buying answer. The second prompt gives a dated product page a clearer job to do. Write for the question that needs the page, not just the topic the page mentions.

Then, the page has to survive extraction and trust

Once a system tries to retrieve the opener page, the next checks are less glamorous than any AI SEO pitch deck. Can the relevant bot get through? Can the fetcher read the page without a login or a render it cannot use? Can it find the answer without excavating six screens of introduction?

Check access before rewriting copy. Review robots.txt, rendering and any CDN or firewall rules that might affect requests. Do not identify a bot by its user-agent string alone; that string can be spoofed. The crawler guidance describes IP verification using published bot lists. And do not assume robots.txt settles every kind of request: this crawler checklist distinguishes user-triggered fetches from ordinary crawler access and points to firewall controls.

After access comes extraction. If the opener page has a relevant temperature rating, state it plainly near the answer it supports. Put the price, specification or comparison in a passage a reader could understand on its own. A neat table can help when the client is comparing options; it cannot rescue a page that never says which option fits the cold-climate question.

Then comes selection. One account of Perplexity’s process describes an initial set of candidates narrowed to a smaller set of cited sources using factors including specificity, reliability and freshness. That is a reason to make a claim clear and supportable, not a guarantee that a well-formatted paragraph will win. A client saying its own opener works at minus 20 degrees gives the system one claim. Independent reviews or dealer pages that confirm the same true specification give it more to work with. A readable claim still has to be worth relying on.

That is where the plumbing analogy stops. Fetching is plumbing; deciding what to cite is not. A working pipe does not make every sentence credible, and an answer system does not owe a citation to every page it reads. The agency’s job is to remove avoidable failures at each gate, then measure what actually happened.

What 30 days of our crawler logs can—and cannot—show

Here is the number behind the opening. Over 30 days, a groas post about a Google Ads update received 580 verified ChatGPT-User fetches. A second related post received roughly 380. The page-by-page crawler teardown covers the underlying log review. These are requests to pages, not a count of published citations, distinct answers or readers who saw them. Server logs cannot tell us which passage, if any, appeared in the final answer.

Verification matters because a user-agent label is cheap to copy. We checked bot identities by IP against OpenAI’s published crawler files and Perplexity’s published bot list, rather than counting every request that announced itself as an AI bot. Unverified hits outnumbered verified ones by a wide margin. A dashboard that bundles those hits together can look busy while telling you very little about genuine bot access.

The two posts shared useful characteristics: each covered a dated Google Ads change soon after it happened, named the relevant setting, and put the explanation in short, answerable passages. Those details make them plausible candidates for retrieval on specific update questions. The logs support the narrower conclusion that the pages were fetched repeatedly. They do not isolate which detail caused a fetch, or prove that the pages were quoted afterward.

The robots.txt result deserves the same restraint. It shows how much access-checking appears in a server log; it does not establish the exact sequence of every answer-time request. Still, it gives an agency a sensible priority. If the relevant search crawler cannot reach a client’s pages, polishing FAQ wording is not the first fix. Use logs to diagnose access, not to manufacture a citation count.

Sell the work that gets a page through those gates

If you resell AI SEO, the defensible service is not a promise of position number one. It is work on fetchability, extractability and corroboration: checking bot access and rendering; publishing pages that answer specific buyer questions plainly; keeping time-sensitive facts clear and dated; and making sure claims can be supported beyond the client’s own site.

That order matters. Start with access, because an unreadable page cannot supply a passage. Move to extraction, because a reachable page can still bury the answer. Then examine trust, because a clear but unsupported claim may lose to a better-supported source. A white-label service description lists technical work, answer-focused content, schema and tracking among its components. Those components only make sense as a sequence when you can say what failure each one addresses.

I have sold versions of the other kind of deliverable: a visibility score with no log evidence, a screenshot of one answer, an audit that produces a handsome PDF but no clear next action. They feel like progress because someone can forward them. None, by itself, tells you whether a page became easier to fetch or read.

That does not make every un-fetched page a failure. A system may never need it for the questions asked. Nor does a fetch prove the work succeeded. The honest measurement is a chain of narrower observations: what you changed, whether the relevant bot could access the page, whether verified requests appeared, and whether you can separately observe a citation. Do not sell the last link in that chain on evidence from the first.

Make a partner show the mechanism, not a line chart

Ask a white-label AI SEO partner for 30 days of IP-verified fetches by page and the changes they made in response. Ask which bot requested the page, which access or content problem they found, and what they changed next. If they claim citations, ask how they observed citations separately from server requests. A screenshot of a single answer or an unlabeled bot-traffic chart does not answer those questions.

I run this work inside groas, where AI search fulfillment combines daily execution with a named strategist and logged changes. I would apply the same test to us. A fetch is evidence of access; a citation requires evidence of citation. A partner who understands that distinction has a better chance of fixing the right problem than one selling rankings language for a retrieval product.

Now take a new case: an agency has 12 HVAC clients who want to appear in Perplexity answers about service costs. Work through the model, rather than promising all 12 a position.

1. Fetch: Check access and rendering on each site for the relevant bots. Verify bot identity in the logs, and look for blocks or failed requests before changing page copy.
2. Extract: Give each cost question a clear answer. Put the relevant price range, service area and date where a fetcher can find them without guessing which part of a broad services page matters.
3. Trust: Check whether the figures are supportable and whether independent dealer or review pages corroborate the claims the client wants repeated.
4. Measure: Review verified fetches by page over 30 days. Treat pages with no requests as a question to investigate, not automatic proof of bad copy. Check any claimed citations separately.

The model does not guarantee that Perplexity will cite any of those pages. It tells the agency what to fix first and what the evidence can honestly support. That is the decision the numbers allow: sell reachable pages, liftable answers and supportable facts. Do not sell a ranking the system never assigned.