Cited is not used: what 21,143 AI citations reveal about GEO

Being cited by an answer engine and feeding its answer are two different outcomes. A preprint posted to arXiv on 29 April 2026 puts numbers on the gap, across 602 controlled prompts and 21,143 citations: ChatGPT keeps 6.88 sources per prompt on average, Perplexity 16.35. Yet among pages that were successfully fetched, the mean influence of a source cited by ChatGPT reaches 0.2713, against 0.0646 for Perplexity. A ratio above four.

Put differently: the engine that cites the most is the one that uses each source the least. A mention counter therefore measures exposure. It says nothing about contribution. That is precisely the metric most GEO dashboards sold today display in large type, and often on its own.

What is citation absorption?

Citation absorption is the share a cited source takes in the final text of a generated answer: contributed language, figures, structure or supporting evidence. It differs from citation selection, which is only the fact that an engine triggered a search and kept the source in its list. A page can sit in the sources of an answer without having supplied a sentence. Another can carry three paragraphs.

The authors measure absorption with a composite score computed on the cited page. It combines five signals observable in the answer: how often the source is referenced, where it first appears, the share of answer paragraphs it covers, TF-IDF similarity between page and answer, and bigram and trigram overlap. Similarity carries the heaviest weight (0.25), first position the lightest (0.15).

This is a reconstruction from visible text, not a trace of model attention. The authors say so plainly, and draw a methodological rule from it: the components of the score cannot serve as explanatory variables for the score. That kind of discipline is rare in commercial GEO literature.

How many sources does each engine cite, and how deeply?

All three platforms trigger a search for almost every prompt: 98.64% for ChatGPT, 99.67% for Google AI Overview, 100% for Perplexity. Whether engines go looking for sources is a settled question. What separates them is how many sources they keep and what they do with them.

Source: arXiv:2604.25707v2, search-layer and absorption-layer summary tables. Influence is computed only on pages whose content could be fetched (overall fetch success: 76.44%).

The gap widens further on complex prompts. On multi-constraint tasks, ChatGPT drops to 3.4 sources on average while Perplexity climbs to 17.7. Two opposite strategies: one compresses and synthesises internally from a narrow base, the other decomposes the task and widens retrieval.

For a brand, the consequence is concrete. Presence in Perplexity looks like a slot in a source directory. Presence in ChatGPT looks closer to an editorial contribution. Both have value, but they are not driven by the same content or tracked with the same indicator.

Does Q&A formatting actually improve AI citation?

No, not on its own. Pages in question-and-answer format score a mean influence of 0.0947, against 0.1005 for pages that do not use the format. A relative gap of -5.74%, against the FAQ. It is the only content marker in the study pointing the wrong way.

Do not over-read it. The measure compares means across a corpus where many Q&A pages are thin support pages, while many non-Q&A pages are long explainers, official documents or encyclopedic articles. The study does not prove FAQs hurt. It shows that surface packaging is not enough, and that value comes from the evidence inside the page, not from question marks in the headings.

This result lands on us directly. Our own articles end with an FAQ block, this one included. What the number changes in our method: the FAQ stays a coverage tool — it catches phrasing variants of a single intent and feeds FAQPage markup — but it stops being presented as an absorption lever. An FAQ carrying no dated figure, no clean definition and no comparison criterion is decoration.

Which content formats do engines absorb most?

Pages carrying reusable units of evidence are markedly more absorbed than the rest. The study compares mean influence for pages holding a given marker against pages without it.

Source: arXiv:2604.25707v2, evidence-genre table. Unadjusted comparison of means — no control for length, domain type or editorial quality.

The ranking by semantic role tells the same story from another angle. A source used as a definition reaches 0.1531 mean influence, a source used as a comparison 0.1524, a source used as a plain reference 0.0529. Close to a threefold spread between the extremes. Supplying a fact beats supplying a link.

What does a heavily absorbed page look like?

The top-quartile profile, set against the bottom quartile, is stark: 1,943 words against 170, 10.6 headings against 0.85, 47 paragraphs against 8, list density multiplied by nine, and semantic similarity with the generated answer of 0.570 against 0.247.

The easy reading would be "write long." It would be wrong. The strongest correlation reported in the study is not length: it is the page's relevance as scored by an LLM (r = 0.4322), followed by embedding similarity between page and answer (0.3561). Length helps only when it carries structure and evidence. A 3,000-word page made of navigation, repetition and self-promotion offers nothing to extract.

The authors turn this into their central hypothesis, which they call the evidence container. A page becomes useful to a generative engine when it can be broken into self-contained, on-topic information units: bounded topical scope, headings that mirror the likely sub-questions, one paragraph per claim, and extractable material in the form of definitions, figures, comparisons and steps.

A long, modular page gives the engine several chances to map an intent to a specific segment. A long, shapeless page gives it none.

Why is Wikipedia absorbed while the press is only cited?

News sits heavily in the selected sources — it accounts for 31.17% of ChatGPT's citations — yet it lands at the bottom of the absorption ranking. News_media pages cap at 0.0726 mean influence, while encyclopedia pages reach 0.2144.

Source: arXiv:2604.25707v2, influence by domain type. The authors flag that the website-type taxonomy contains noisy values.

The reading the authors propose is plausible and useful: the engine reaches for a news outlet to date and anchor a fact, then for an explanatory page to build the body of its answer. Two needs, two roles, two kinds of source. The commercial citation volume — close to 11,800 in the sample, at 0.1028 influence — also shows that company pages are not shut out of the game. They are simply rarely written to hold the explanatory role.

Does domain authority still matter?

Yes, and it decides upstream. Official, news and vertical sources concentrate between 79.12% and 87.52% of citations depending on the platform. The median domain rating of selected sources sits between 526 and 592. Sources identified as US-based account for 82.70% to 86.76% of the sample, English-language sources for 82.90% to 95.07%.

For an agency in Bangkok, a law firm in Geneva or a Luxembourg trust company, this is the most important figure in the study, and the least pleasant. Editorial quality separates sources that are already eligible. It makes nobody eligible. A local, French-speaking or Thai player starts with a structural handicap at the entrance to the candidate pool, however sharp its paragraphs.

So the lever is not editorial first. It is building a recognisable entity corroborated elsewhere: consistent mentions, presence in indexable third-party sources, proprietary data nobody else publishes. An original figure is the only asset that forces attribution: when three sources state the same number, the model attributes by domain preference, and the smaller player loses.

What this study does not support

• No causal link. The authors explicitly reserve prescriptions such as "adding a comparison will increase your citations" for future controlled intervention experiments.
• The influence score is a proxy built from visible text. A different weighting scheme would produce different numbers — the authors put exactly that at the top of their own robustness plan.
• 602 researcher-designed prompts are not a sample of real traffic. External validity is bounded to that distribution.
• The language contrast covers Chinese and English. Nothing in this data says anything about French, Thai or any other language. Language effects also vary by platform, which rules out any global rule.
• Absorption is computed only on successfully fetched pages, 76.44% of the total. If unfetched pages differ systematically, the means are conditional on that selection.
• It is a snapshot. Generative AI products change retrieval components, citation interfaces and underlying models within months.

What we are changing in the KAIS method

1. Two columns instead of one in the Barometer. A selection rate — does the brand appear among sources — and a depth-of-use indicator — does the answer reuse its wording, its figures, its structure. The two move independently.
2. One objective per engine. Perplexity is worked for coverage, ChatGPT for contribution. Folding both into a single score erases the information.
3. Sections written as units of evidence. A clean definition opening the section, a dated figure with its source, a table when the intent is comparative. No argument chopped into bullets.
4. The FAQ drops a rank. It serves phrasing coverage and markup. It no longer serves as a depth argument.
5. Segmentation by prompt family. A page that wins on "what is" queries does not win on comparison queries. We measure them separately.

The question to ask a GEO provider in 2026 is no longer how many citations you got. It is: what did the engine take — and from whom.

Sources and references

• Zhang K., He X., Yao J. (2026). From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms. arXiv:2604.25707v2, 29 April 2026 — sole source of every figure quoted in this article.
• geo-citation-lab — public dataset and analysis pipeline
• Aggarwal P. et al. (2024). GEO: Generative Engine Optimization. arXiv:2311.09735, KDD 2024 — the paper that coined the term.