An AI engine can list your page as a source and take nothing from it.
It can also build half its answer on your page and credit someone else. Most LLM SEO reporting cannot tell these two cases apart, because it counts citations and stops there.
The 2026 critical survey of generative engine optimisation research calls that count one component of seven, and the component that tells you whether you shaped the answer is a different one: absorption.
This article explains the difference, shows what it looks like in a real answer, and gives you a way to check each of the seven dimensions without buying a tool.
What the visibility vector says a citation is
A citation is an explicit selection of your page as a source, and nothing more. That is the survey's own wording. Olivier Martinez's critical survey of 45 GEO studies (Sciences Po, July 2026) breaks "visibility" in a generative engine into seven measurable parts: discoverability, context exposure, citation, prominence, absorption, fidelity and behavioural outcome.
For each one the paper states what the metric identifies and what it does not establish. For citation, the row in its taxonomy table reads: identifies "explicit selection as a source", does not establish "exact support or substantive contribution". Absorption is defined as the source's "effective contribution to the facts, language, or structure of the answer".
"The same source may be absent, cited without being used, paraphrased without a link, or decisive in shaping the structure of an answer." Olivier Martinez, A Critical Survey of Generative Engine Optimization
Those four states are the whole argument. Two of them (cited without being used, paraphrased without a link) break the assumption behind every citation-share dashboard, which is that a link in the sources panel means your content did some work.
How often a citation is decorative
Nobody has a clean figure for how often a cited page is ignored, but the evidence that citations and support come apart is old and consistent. The survey reports Liu et al's 2023 audit of four generative search engines, in which only 51.5% of answer sentences were fully supported by their citations, and 74.5% of citations supported the claim they were attached to. A quarter of citations pointed at pages that did not say what the answer said.
The engines have moved on since 2023, but the newer figures still show daylight. The survey cites Xu et al's 2026 study, which classed about 11% of 98,020 atomic claims as insufficiently supported by the sources the engines gave for them. The survey's section heading on this is blunt: "Citation implies neither credibility nor support."
There is a commercial version of the same point. The survey warns that without a link from citations to value, GEO risks optimising decorative citations. A page that is listed and never read is decorative. Your monthly report counts it as a win.
What the absorption study found about the engines
ChatGPT cites fewer pages and relies on each one more heavily; Google and Perplexity spread their citations thin. That comes from the one study so far that has tried to score absorption at scale, by Zhang, He and Yao, published in April 2026 as From Citation Selection to Citation Absorption.
They ran 602 controlled prompts through ChatGPT, Google AI Overviews and Perplexity, collected 21,143 citations and fetched 18,151 of the cited pages. For each fetched page they computed an influence score built from how many times the page was referenced, where it first appeared, how many answer paragraphs it covered, and how closely the answer's wording matched the page's.
The per-engine results explain why one citation is not worth the same everywhere:
| Engine | Citations per answer (mean) | Influence per cited page (mean) |
|---|---|---|
| ChatGPT | 6.88 | 0.2713 |
| Google AI Overviews | 12.06 | 0.0584 |
| Perplexity | 16.35 | 0.0646 |
Source: Zhang, He and Yao, 2026, pages 8 to 11. Influence is the paper's composite proxy, not a measure of model attention.
A Perplexity citation is one of sixteen or seventeen, and the median influence per page there is 0.0333, half the mean, which means most cited pages contribute almost nothing and a few carry the answer.
A ChatGPT citation is one of about seven and the typical page in that list has done real work. The same "we were cited" line in a report describes two very different outcomes.
The authors are careful about the limits. The score is, in their words, a weighted proxy for how deeply a cited page appears to shape a generated answer, and the Martinez survey adds that a true measure of absorption would need a matched ablation with and without the source.
Treat the numbers as a strong signal about engine behaviour and a weak one about any single page.
A worked example: cited and ignored, or absorbed
Here is what the gap looks like in an actual answer. The prompt, pages and figures below are invented for the example; the pattern is the one the two studies describe.
The prompt is "How often does a landlord in England need an EICR and what does one cost for a two-bed flat?" Two electrical contractors in Northampton have pages that get retrieved.
Page A is 1,400 words. It opens with three paragraphs on why electrical safety matters, says the cost "depends on a number of factors including the size and age of the property", and asks the reader to call for a quote. The five-year rule is in the ninth paragraph, inside a sentence about "ensuring full compliance with the relevant regulations".
Page B is 500 words. Its second sentence is "A landlord EICR for a two-bedroom flat costs £160 to £210 with us, including the certificate." Its third is "Landlords in England must have one at least every five years, and give tenants a copy within 28 days." A short table below lists prices by property size.
The answer says: "Landlords in England must have an EICR at least every five years and give the tenant a copy within 28 days. For a two-bedroom flat, expect to pay around £160 to £210." Both pages appear in the sources panel.
Page A was cited. Page B was absorbed. Every fact, the figure, the deadline and the sentence shape came from B. A's citation is there because A was retrieved and ranked, and the engine attached it to the compliance point.
If A's owner runs a citation tracker, this month shows a hit. Ask the question the survey proposes, which claim or decision would have changed in its absence, and the answer for A is none.
What makes a page absorbable
Pages that state short, attributable facts get absorbed; pages that talk around them get listed. The Zhang study compared influence scores for pages with and without particular content features, across all three engines:
- Numbers and statistics: 0.1171 against 0.0725 without them, a 61.55% higher influence score.
- Definition markers (a sentence that says plainly what something is): 57.33% higher.
- Comparison content (this against that, with the differences stated): 55.28% higher.
- How-to content with ordered steps: 41.20% higher.
- Question-and-answer formatting: 5.74% lower. Writing "Q: How often is an EICR needed?" above your answer did not help.
List density told the same story. Pages in the top quartile of influence had a list density of 0.428; the bottom quartile had 0.048.
The authors do not claim causation, and neither should you. A page with a price table is also more likely to be the page that actually answers the question, and the engine may be rewarding the answer, not the markup.
The Martinez survey reaches a similar conclusion by a different route. Statistics, definitions, comparisons, prices, dates and references have a plausible advantage because they are units the generator can select and attribute. A fact the engine can lift out in one piece, with your name on it, is easier to use than an argument spread over three paragraphs.
Two cautions from the same paper. Authoritative tone scored weak and unstable, so sounding confident is not the same as being quotable. And the survey warns that adding a fabricated statistic may increase reuse while degrading quality, which is a polite way of saying that invented numbers get absorbed too. The fidelity dimension exists partly to catch that.
Google's own guidance lands in the same place without the data. Its AI optimisation guide, updated July 2026, asks for content that "provides unique expert or experienced takes that go beyond common knowledge", says you "don't need to create new machine readable files, AI text files, markup, or Markdown", and that "there's no special schema.org markup you need to add".
A real price and a real rule, stated once and clearly, is what both the regulator-grade study and the search engine are asking for.
The seven dimensions, and how to check each one
You can check every dimension except absorption with a spreadsheet and an afternoon, and you can get a usable proxy for absorption by reading. The table below is the survey's vector turned into questions.
| Dimension | What it means | How to check it |
|---|---|---|
| Discoverability | The engine retrieves your page for the prompt at all | Run each prompt 7 to 8 times and record whether your domain appears anywhere in the sources, including below the fold |
| Context exposure | How much of your page the engine gets, and how high it ranks among retrieved pages | Note your position in the sources list and whether the engine quotes a passage or just a title |
| Citation | Your page is named or linked as a source | Count the runs in which you are cited, per engine, and report ChatGPT and Perplexity separately |
| Prominence | Where the citation sits and how often it recurs in the answer | Record whether you are the first citation, and how many sentences carry your marker |
| Absorption | Your facts, wording or structure appear in the answer body | Copy the answer, highlight every sentence that could only have come from your page, and count them |
| Fidelity | What the engine attributes to you is what you actually said | Read each attributed claim against your page and mark it supported, distorted or absent |
| Outcome | Someone clicks, enquires or buys | Filter GA4 referrals by chatgpt.com, perplexity.ai and gemini.google.com, and ask new enquiries where they found you |
The 7 to 8 repetitions figure is the starting point the survey takes from Schulte et al, who found that 57.8% of ChatGPT repetitions did not trigger web search at all. Those null runs count as results. A prompt that never triggers search cannot cite anyone, and that is useful to know before you rewrite a page for it.
The absorption check is manual and it is the one that matters. If you find nothing to mark in the answers where you are cited, your citations are decorative.
If you find your sentences in answers where you are not cited, you have the "paraphrased without a link" case, and the fix is to make the fact more attributable (a named figure, a dated claim, your business name next to it), not to chase the link.
Why this matters more in the UK than the dashboards suggest
The AI answers most UK adults see are Google's AI Overviews, and Google is the engine that spreads citations thinnest. Ofcom's December 2025 report on how the UK goes online found that about 30% of searches now show AI Overviews and 53% of adults report often seeing them.
ChatGPT took 1.8 billion UK visits in the first eight months of 2025, up from 368 million a year earlier.
Put those together with the Zhang figures. On the surface most UK users see, a citation is one of twelve and carries a mean influence of 0.0584. On the surface growing fastest, it is one of seven and carries 0.2713.
A report that adds the two up as "citations this month" is adding a decorative mention to a load-bearing one. For a UK firm, the honest number is absorbed sentences per engine, with the ChatGPT column likely to be the one that moves enquiries. For legal practices and others where a single new client is worth a lot, that one column can justify the whole exercise.
Our own LLM optimisation work reports citation and absorption as separate columns for that reason, with the number of runs behind each figure. That takes longer to produce than a share-of-voice score, and it is the only version of the number that tells you whether to rewrite the page or leave it alone.
The decision rule is simple. Before you celebrate a citation, read the answer and find your sentence. If it is not there, you were listed, not used, and the page needs a fact the engine can take.
If you want the seven checks run against your own prompts and pages, with ChatGPT, Perplexity and Google reported separately, we do this for clinics and other healthcare providers as readily as for any service firm. Talk to us about AI visibility.