Does AI search actually use the pages it cites?

What is the difference between being cited and being used?

Being cited means your page appears in the source list. Being used means its facts or wording shape the answer.

Zhang Kai, He Xinyue and Yao Jingang (opens in a new tab) call these two stages “citation selection” and “citation absorption.” They studied 602 prompts across ChatGPT, Google’s AI Overviews and Gemini, and Perplexity, covering 21,143 citations. For each cited page they built an influence score from 0 to 1. It rises when the answer refers to the page often and early, covers it across several paragraphs, and shares its wording.

The idea is not theirs alone. Martinez’s 2026 survey (opens in a new tab) of GEO research notes that a source may be “cited without being used, paraphrased without a link, or decisive in shaping the structure of an answer.” The reverse also happens: an answer can draw on pages it never lists, as Tannenbaum (opens in a new tab) points out.

Note the source. The absorption study comes from three independent researchers, uses a public dataset, and has not been peer reviewed.

Which AI search engines give each cited source more influence?

ChatGPT cites fewer pages but leans on each one more. Google and Perplexity spread their answers across many more sources.

The gap widened on hard questions. For prompts with several constraints, ChatGPT averaged just 3.4 citations, while Perplexity averaged 17.7. The authors read this as ChatGPT doing more of the work itself after picking fewer sources, but say that is an interpretation, not a proven cause.

A separate team reached a similar result by another method. Huang and colleagues (opens in a new tab) at the University of Illinois collected answers to 11,000 real search queries. For 1,100 queries per system, they checked which facts in each AI summary could be traced to each cited page. Google’s AI Overviews cited nearly twice as many sources as OpenAI’s search-enabled GPT, at 5.0 against 2.7. Yet they captured less of those sources’ content, scoring 0.447 against 0.500. More citations meant thinner use of each.

Are some cited sources only there for show?

Often, yes. Several audits found sources that are listed or cited but add little or nothing to the answer.

A 2024 audit by Narayanan Venkit and colleagues (opens in a new tab) tested You.com, BingChat and Perplexity on 303 queries. Of the sources Perplexity listed, 8% were never cited in the answer text; for BingChat it was 36%. Every engine included at least 30% of sources that backed no unique information in the answer. Users in their study called this “buffing”: an impression of thoroughness without substance.

The gap can also be selective. Huang and colleagues found that Google’s AI Overviews cite Reddit and Quora, but drew 22.1 percentage points less content from those forum sources than from others. Sources with a negative tone were drawn on 13.8 percentage points less in AI Overviews. The authors call this a “citation-synthesis gap”: the list looks diverse while the text relies on a narrower set.

The role a page plays matters too. In the absorption study, pages cited as a definition averaged an influence of 0.1531. Pages cited only as a reference averaged 0.0529, about a third as much.

Which cited pages get used the most?

Pages that supply a definition, a figure or a comparison get used most. Pages that are merely relevant get used least.

The absorption study shows a clear split between being picked and being used. News sites are cited often, yet news pages averaged an influence of 0.0726, against 0.2144 for encyclopedia pages. Pages cited as a direct source of facts averaged 0.1241; pages used only as background, 0.0775.

Format alone did not help. Pages written in a question-and-answer format scored 5.74% lower than other pages. We cover that result in our article on FAQ pages. Length helped when it carried substance: Huang and colleagues found short sources were drawn on 17.7 percentage points less in Google’s AI Overviews.

Does a cited page always back up what the answer says?

No. Some claims in AI answers are not supported by any page the answer cites.

In a 40-day audit of Google’s AI Overviews in spring 2026, Xu, Iqbal and Montgomery (opens in a new tab) checked 98,020 claims. They found 11.0% were not supported by the cited pages, mostly because no cited page mentioned the claim at all. We look at this problem, including claims wrongly pinned on a page, in our article on misattributed citations.

Why isn’t citation count enough to measure GEO?

Because a count treats a passing reference and the page that wrote half the answer as equal. It also rewards engines that cite many sources lightly.

Count citations on Perplexity and your page may appear often while shaping little. On ChatGPT a page may appear less often but carry far more of the answer. The absorption study’s authors warn that “a dashboard that counts citations alone will miss the central pattern of this dataset.” They suggest tracking five things side by side: how often you are selected, how many citations you get, how much your page shapes the answer, whether the answer is accurate, and how concentrated citations are.

We discuss how to set up those measures in our article on what to measure and in our guide to share of citations.

What should you do about it?

Measure whether AI answers use your pages, not only whether they list them. Then make your pages easier to use.

1. Read the answer, not just the source list. For your key buyer questions, check which of your facts, figures or phrasing appear in the text.
2. Report results by engine. A citation on ChatGPT and one on Perplexity are not worth the same.
3. Give engines something to take. Clear definitions, specific numbers and comparisons went with the most use.
4. Check accuracy. When your page is cited, confirm the claim next to it matches what your page says.
5. Do not chase citation counts alone. A rise in citations with no change in what answers say about you is a weak win.

If you want help building this kind of measurement, see our generative engine optimization service.

What does the research not tell us yet?

Researchers can describe how unevenly sources are used, but cannot yet see inside the engines. The main gaps:

• The influence score is a stand-in built from wording overlap and position. Martinez notes it “does not reveal an internal causal trace.” A true test would compare answers with and without a source.
• Measures of use only cover pages researchers could read. The absorption study fetched 76.44% of cited pages. In our reputation study, only 38.6% of cited pages could be read automatically.
• The studies are snapshots. Engines change quickly, and the absorption study lacks dates for each record.
• No study yet links being used, rather than merely cited, to clicks, trust or sales.

Frequently asked questions

What is citation absorption in AI search?

It is how much a cited page actually shapes an AI answer. Zhang, He and Yao coined the term to separate it from simply appearing in the source list.

Which AI search engine relies most on each source it cites?

ChatGPT, in the one study that measured it. Each fetched page ChatGPT cited scored 0.2713 on influence, against 0.0584 for Google and 0.0646 for Perplexity.

Is citation count a good GEO KPI?

Not on its own. Citations vary in how much they shape the answer, so pair the count with a check of what each answer actually says.

Does Perplexity use every source it lists?

Not always. In a 2024 audit, 8% of the sources Perplexity listed were not cited in its answer. In a later study, its cited pages averaged a low influence score of 0.0646.

Sources

• Zhang, He and Yao (2026), From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms (opens in a new tab), arXiv:2604.25707.
• Huang, Goyal, Saha and Chandrasekharan (2026), Answer Bubbles: Information Exposure in AI-Mediated Search (opens in a new tab), arXiv:2603.16138.
• Narayanan Venkit, Laban, Zhou, Mao and Wu (2024), Search Engines in an AI Era: The False Promise of Factual and Verifiable Source-Cited Responses (opens in a new tab), arXiv:2410.22349.
• Xu, Iqbal and Montgomery (2026), Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact (opens in a new tab), arXiv:2605.14021.
• Martinez, O. (2026), Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026) (opens in a new tab), arXiv:2607.14035.
• Tannenbaum (2026), Scoring With the Engine: Retrieval Exposure, Cross-Engine Divergence, and the Limits of Engine-Agnostic GEO Scores (opens in a new tab), arXiv:2609.22655.
• Underneath (2026), “Is this brand legit?” How AI assistants build a reputation