Naming Your URL Is Not Reading Your Page

There are three separate things people collapse into the word “cited”, and the distinction decides whether your AI-visibility reporting means anything.

An engine can mention you, naming your company in prose. It can link you, printing your URL as a source. And it can read you, actually fetching the page and using what is on it. A dashboard sees the second one. Buyers are affected by the first. Only the third tells you the words you wrote did any work.

We built a small harness to pull those apart on our own pages, ran it once, and the three came apart faster than we expected.

What we ran

The mechanism is simple enough to reproduce. Instead of asking an engine a category question and hoping it reaches for us, we ask it a question that names one of our URLs directly. Naming the URL is what forces a live fetch rather than an answer assembled from whatever the model already carries about us — and the difference between those two is the entire point of the exercise.

The questions are written in an operator’s voice rather than a marketer’s, because a prompt that sounds like a press release gets answered like one. Targets come from our own analytics, highest-impression pages first, so the harness always points at whatever is currently getting attention rather than at whatever we feel like promoting.

The first live run was on 11 August 2026. ChatGPT and Claude each answered four questions, for eight answers in total. Perplexity was in a rate-limit cooldown and Google AI Overviews blocked the run outright, so neither produced a reading. We record those as no data rather than as zeroes, for the same reason we do it on our numbers: a refusal is missing information, and scoring it as absence quietly flatters whatever you measure next.

What came back

	ChatGPT	Claude
Questions answered	4	4
Answers showing clear evidence of reading	3	3
Answers that named the URL back	1	3
Shortest answer	Substantive	427 characters

Six of the eight answers contained material that could only have come from the page — specific claims, specific structure, in two cases a direct criticism of the argument. So the fetch mechanism works. That part we can state.

Everything interesting is in the disagreement between the two rows underneath.

Two failure modes that look identical from a dashboard

ChatGPT read three of its four pages without naming the URL back. It engaged with the content in detail and attributed none of it. To any monitoring tool counting citations, those three answers are a zero. In terms of what the model actually did with our page, they are the best result in the run.

Claude named the URL on three pages and barely engaged with any of them. Two of those answers were 427 and 458 characters — roughly a long paragraph each, for a question about a 1,400-word article. To the same monitoring tool, those three are the clean win.

The measurement and the reality are inverted. A tool that counts links would have scored Claude above ChatGPT on this run, and the transcripts say the opposite.

This is not a criticism of either engine; they are doing different things and neither owes us an attribution style. It is a criticism of counting links and calling it visibility, which is what most of this category sells, including the part of it that is us. Our own share-of-voice figure counts mentions, and this run is a standing reminder of what that figure cannot see.

What this does not show

It does not show that making an engine fetch your page improves your visibility. We want to be exact here, because this is the claim the technique invites and it is not one the data supports.

What is established is narrow: a question naming a URL causes a fetch, and the fetched content shows up in the answer. What is entirely unestablished is whether inducing fetches changes how often you are named in answers to questions that do not name you — which is the only thing that matters commercially. Nothing in this run bears on that, and we found no public evidence establishing it either way. On our evidence register it is graded unproven, alongside the industry’s time-to-results convention, and it will stay there until an audit diff moves in a way we can attribute.

The sample is also tiny. Eight answers, two engines, one day, one site — our own. Two of the four engines we normally sample produced nothing at all. This is a pilot we are publishing because the mechanism is interesting, not a result we are asking anyone to act on.

What would turn this into a finding

Three things, none of which we have yet.

Repetition. The same eight questions on several dates, because engine behaviour varies between runs and a single pass cannot separate a pattern from a mood. This is the same discipline that makes a share-of-voice number worth comparing to itself, and it is the first thing every shortcut in this field drops.

A control. Matched pages where we do not induce a fetch, measured on the same schedule, so any movement in citations can be attributed to the technique rather than to everything else happening at once.

An outcome measure that is not the technique itself. A fetch proving a fetch happened is circular. The only honest endpoint is the unprompted category question: does the engine name us when nobody has pointed it at our URL?

What we changed because of it

One unplanned use turned out to be the most valuable part of the run. Because the answers engage with the page’s argument, they function as an unusually direct critique from the position of a reader who has no interest in being polite.

ChatGPT’s verdict on one of our own articles was that it was “a real methodology wrapped in an agency pitch”. That was correct, it was the exact criticism our positioning work was trying to fix, and the post was rewritten in response — leading with the do-it-yourself method and moving the commercial framing to the end. That article is now how to track your brand’s visibility in ChatGPT, and it is a better page because a machine told us the truth about it.

We do not claim this improves rankings, citations, or anything else. We claim it is a cheap and unusually candid editing review, which is a smaller and more defensible thing to say.

Questions

Is this just prompting the model to read a page? Essentially, yes — that is the whole mechanism, and its simplicity is the point. The interesting result is not the technique, it is what the transcripts reveal about the gap between attribution and comprehension.

Could an engine fake evidence of reading? It could produce plausible-sounding material from training data rather than from the live page. We guard against this by asking about content specific enough that a model recalling the domain in general would get it wrong — recent figures, section structure, particular phrasings. It is a defence, not a proof, and on a sample of eight it is one of the weaker parts of the design.

Does this mean citation tracking is worthless? No. It means a citation count is one of three signals and the least informative of them on its own. Read it alongside mention rate and the cited-source map, which is what our methodology sets out and what how LLMs choose citations covers in more depth.

The bottom line

Mentioned, linked and read are three different events, and every reporting tool in this category — ours included — can see at most two of them. Before you accept a visibility number, ask which of the three it counts. If the answer is links, you are looking at Claude’s three attributions and not at ChatGPT’s three silent reads, and on this run that would have been exactly backwards.

We will run this again, publish the next one whichever way it goes, and keep the claim graded unproven until something other than our own enthusiasm moves it. The infrastructure behind it is described on our platform page.