If ChatGPT cites your page, did it actually use it?

A ChatGPT citation means your page was selected as a source, which is not the same as the answer having been built from it. The engine can list your URL under an answer that borrowed almost nothing from you, and it does that routinely.

The distinction has a name now. Zhang, He and Yao propose separating citation selection from citation absorption: selection is whether a platform triggers a search and chooses your page, absorption is how much content, evidence or fact the written answer actually takes from it. Their measurement covers 602 controlled prompts, 21,143 valid search-layer citations, 23,745 citation-level feature records and 18,151 successfully fetched pages across 72 extracted features.

Mean fetched-page influence in that study came out at 0.2713 for ChatGPT, 0.0646 for Perplexity and 0.0584 for Google AI Overview and Gemini. Read those as relative weights rather than as percentages of anything: a cited page on ChatGPT carried roughly four times as much of the answer as a cited page on Perplexity or Google. Hold the figures at arm's length for the usual reasons, since this is one preprint analysing a single public dataset and nobody has replicated it.

What follows is a way of reading your own reporting. A dashboard showing citations across four engines is counting events that are not equivalent, and a rise driven by Perplexity is a smaller thing than the same rise on ChatGPT. Our page on what counts as a good AI citation rate covers the denominators; this page is about what each citation is worth once you have one.

What is the difference between being cited and being used?

Being cited is a slot in a source list. Being used is your language, your numbers or your framing turning up in the sentences the assistant writes. The two come apart because an engine fetches far more than it needs and then writes from a subset of what it fetched.

Three things explain why the gap is so wide. Engines retrieve defensively, pulling in more sources than the answer requires. Attribution is generated alongside the text rather than derived from it, so a URL can be listed without having contributed. And an answer assembled from several sources will lean heavily on one or two of them, which leaves the rest as decoration. OpenAI's own developer documentation makes the gap explicit: a sources field returns every URL retrieved during a search, inline citations show "only the most relevant references", and "the number of sources is often greater than the number of citations" (OpenAI, web search guide, read 21 October 2026). That field, and Anthropic's equivalent, are compared in our guide to what ChatGPT and Claude actually pass to a search index.

The consequence for anyone paying for this work is that two different problems look identical in a report. A page that never appears has a retrieval problem, and the fix is access, coverage and alignment with the question. A page that appears constantly and never shapes the answer has a substance problem, and the fix is on the page itself. Treating the second as though it were the first produces another round of restructuring that changes nothing.

Separating them is also the honest limit of what any page-level score can tell you. Chen and Koudas, across 158,099 answers and 468,037 citations, found that which pages an engine's own retrieval returned explained about 32% of the variation between pages against about 2.8% for page properties. So most of whether you get cited is decided before your page is compared with anything, and what your page controls is mainly what happens once it is in the room. Our methodology sets out the four pipeline stages a page has to clear and the seven dimensions we score.

Which engines cite the most sources, and which lean on them hardest?

Perplexity cites the most sources and leans on each one least. ChatGPT cites the fewest and leans hardest. The table gives the measured figures from the absorption study and what each row means if that engine is where your citations are coming from.

Engine	Mean citations per prompt	Mean fetched-page influence	What a citation there is worth
ChatGPT	6.88	0.2713	The most valuable slot measured. Fewer sources, each carrying much more of the written answer
Google AI Overview and Gemini	12.06	0.0584	Broad and shallow. Appearing is common, shaping the answer is not
Perplexity	16.35	0.0646	The easiest citation to win and the thinnest slice. Good for presence, weak as evidence of influence

All three figures come from the same 2026 study of 602 prompts and 18,151 fetched pages, so they are comparable with each other even if none is confirmed by a second measurement. The median influence values ran lower than the means on every engine, which tells you the distribution is skewed: most cited pages contribute little and a few contribute a lot.

Practically, this changes where to look first rather than which engine to chase. If your citations concentrate on Perplexity, presence is not the problem and substance might be. Our page on which AI engine to optimise for first covers the prioritisation, and why you show up in ChatGPT but not Google AI Overviews covers the case where the engines disagree about you entirely.

Does question and answer formatting make your page easier to absorb?

Question and answer formatting does not make a page easier to absorb, and this is the most inconvenient finding in the study for the whole AEO category, ourselves included. The measured difference runs slightly the wrong way.

In the absorption study, Q&A pages averaged 0.0947 mean influence against 0.1005 for pages that were not structured as questions and answers, a relative difference of about -5.7%. The authors read this as question and answer packaging being insufficient on its own without evidence density and semantic fit. Taken with the schema result from Ahrefs' matched test of 1,885 pages that added JSON-LD between August 2025 and March 2026, which found changes of +2.2% on ChatGPT and +2.4% on AI Mode that were statistically indistinguishable from zero, the pattern across both is that page furniture is worth less than this category sells it for.

Question headings still earn their place, on narrower and better ground. A heading written as the question a buyer asks is an alignment signal: it matches the wording the query arrives in, which is a retrieval effect rather than an extraction one, and it is the one page-level property the within-site study found separating pages on the same domain. So keep the question headings and stop expecting the format itself to do the work.

What this rules out is the common shortcut, which is reformatting an existing thin page into questions and declaring it optimised. Adding a question mark to a heading above a paragraph that asserts something unverifiable changes the shape of the page and nothing else. Our page on whether schema markup helps AI citations makes the same argument about structured data.

What kind of page carries the most of an answer?

Pages dense in checkable evidence carry the most of an answer. The split in the data is wide enough to be worth designing around, and it does not track the things most content briefs specify.

Encyclopedia pages in the absorption study averaged 0.2144 mean influence across 527 observations, against 0.0726 for news media pages across 1,546. That is close to a factor of three between two kinds of page that are both well written and both regularly cited. The authors describe high-influence pages as longer, more structured, semantically aligned and richer in extractable evidence, naming definitions, numerical facts, comparisons and procedural steps as the features that matter.

Read the encyclopedia result as a description of a property rather than a genre to imitate. What an encyclopedia page has is a high density of statements an answer can borrow and stand behind: a definition stated outright, a figure with its source attached, a comparison that resolves. What a news page has is narrative, context and quotes, which read well and transfer badly. A B2B page can have the first without pretending to be a reference work.

Our own category research is built on that principle, which is why it exists in the form it does. Our study of who AI engines recommend for pharmaceutical process skids is mostly tables of which firms were named for which query, because a table of findings is the kind of passage an answer can lift and attribute, and almost nobody in that sector had published one.

What should you change on a page that gets cited but never quoted?

Add checkable substance rather than restructuring the page again. A page that already gets cited has cleared retrieval, which is the expensive stage, so another round of headings and summary boxes is work on the part that is already working.

Four edits map onto the features the absorption measurements associate with high influence. Put a number next to each claim that needs one, with its source in the same paragraph, because a figure an answer cannot attribute is a figure it will not use. Define your terms outright in a sentence that would survive being quoted alone. Give the comparison the question implies, usually as a table. And where the question is procedural, write the steps rather than describing the approach.

What to stop doing is as useful. Remove the sentences that assert something unverifiable, because they are the ones an engine cannot stand behind and they dilute the passages around them. Our AI citation checklist covers the full list; the short version is that every paragraph should contain something an answer could borrow and defend.

Then measure the right thing. Instead of counting citations, read the answers: when your page is cited, is any of your language or any of your numbers in the text above the source list? That question is answerable by hand in twenty minutes a month, and it tells you more than a share-of-voice chart. A free visibility assessment reports which of your pages are cited and on which questions, which is the list this exercise needs.

What else do people ask about AI citations and whether a page was used?

The five questions below come up when a client reads their list of citations for the first time and wants to know what each one is worth.