What makes AI engines cite a page

Generative engine optimization (GEO) promises to make a brand’s pages show up as sources in AI answers. The advice sold under that name ranges from rewriting copy with more statistics to adding FAQ schema, and most of it rests on a handful of lab studies. We built an evidence base of 68 studies: the 45 studies in the Martinez (2026) critical survey, plus 23 studies published after its cutoff or missing from it. Each study was coded for the stage of the answer pipeline it measures, the engines it covers, and its strength of evidence on a five-level scale. We then split every study into lever-level findings (61 findings across 19 levers) and counted their direction instead of averaging effect sizes, because the tasks and metrics differ too much to pool.

Findings. Two levers have consistent support from live-engine data. The first is being retrieved at all: in the largest within-site study, whether the engine’s own search returned a page explained 32% of the variation in citations, against 2.8% for every on-page feature combined. The second is how closely a page’s title and best paragraph match the question (+14.1% and +15.5% citations per standard deviation, within the same site). Mentions on third-party pages and the model’s prior knowledge of a brand drive whether the brand is named in the answer. FAQ blocks, schema markup, summary boxes and author boxes look positive across sites and turn null within a site. Generic rewrite recipes (“add statistics”, “sound more authoritative”) rarely survive retests, and in the one study that included a real retrieval stage, they made pages less likely to be found.

Contribution. A ranked, source-linked map of what the current evidence supports, what it does not, and how strong each claim is. Only two of the 68 studies measure traffic in the field, and neither establishes a causal traffic effect. We also derive measurement rules for anyone reporting AI visibility: repeated runs, per-engine reporting, and a control group.

Our clients ask the same question every week: what should we change on our site so that ChatGPT cites us? The market answers it with checklists. Add an FAQ. Add schema. Add an llms.txt file. Rewrite the copy to include statistics and quotes. Most of these items trace back to one 2023 paper (Aggarwal et al., published at KDD 2024) that tested rewrites on a simulated engine with five documents already loaded in its context.

Since then, the research has moved fast. Between January and October 2026 alone, 46 of the studies in our base appeared. Several of them test the same tactics on live engines, with repeated runs, and some of them compare pages within the same site, which removes the most common confound in this field. We wanted one place that says, lever by lever, what that newer evidence supports. This page is that place, and we use it to decide what we recommend to clients.

The review complements our own experiments. In Experiments in AI distribution · 01 we showed that a fresh content domain could move ChatGPT from denying an app existed to citing our content in 42% of monitored answers. That was one subject on one domain. This review asks what holds across the wider literature.

Before you can ask what makes a page cited, you need a model of the pipeline that produces citations. Every study in our base measures one or more of these stages, and a tactic that helps at one stage can hurt at another.

The single most important point in this model is that stage 4 only sees what stage 2 returned. A page that the engine’s search does not return cannot be cited, whatever is on it. Studies that load documents straight into the model’s context skip stages 1 to 3, so they can only tell you how to win credit once you are already in the room. Our evidence levels (section 03) are built around this distinction.

Corpus

We started from the study matrix of Martinez (2026), a critical survey of 45 GEO studies with a cutoff of 14 July 2026. We added 23 studies: 11 published after that cutoff and 12 earlier studies the survey did not include, found through citation chasing from the newer papers. The final base holds 68 studies (1 from 2023, 5 from 2024, 16 from 2025, 46 from 2026). Of these, 22 are peer-reviewed, 3 are workshop papers and 43 are preprints. Four are theory, position or dataset papers with no usable effect estimate.

Evidence levels

Each study received one evidence level, following Martinez (2026, Appendix A). The level describes how close the study gets to real engines and real outcomes.

Nearly half the base (32 of 68) sits at level E. Those studies are useful for mechanisms. They cannot tell you whether a change gets a page found.

Lever findings and vote counting

We split each study into findings, one per lever it tests. A lever is something a publisher could change or something that predicts citation (19 levers, listed in section 04). Each finding received a direction: positive, positive but weak (fails a robustness, placebo or stability check), null, negative, mixed (depends on engine or stage), or risk (works now, creates quality or detection risk). This gave 61 findings.

We did not average percentage gains across studies. A 40% gain in “position-adjusted word count” on a simulated engine and a 15% gain in live citations per standard deviation are different quantities. We report the direction of each finding and the strongest evidence level behind it, and we name the study so you can check it.

Within-site versus across-site comparisons

Large, popular sites tend to use FAQ blocks, schema and author boxes, and they also get cited more. A comparison across sites will credit those features with an effect that belongs to the site. Where a study compares pages within the same site (S46), we give that result priority over cross-site associations.

Verification

Not every study was read with the same depth, and we say so. Six studies, all central to the conclusions, were read in full. For 45 studies we used the Martinez (2026) coding, in four cases checked against the abstract. Eight were read at abstract or conclusion level. Ten results are known only or partly through another paper’s report and are flagged as “to review” in our sheet; none of our conclusions rests on them alone. We then checked every headline number in this paper against the full text of its source (Appendix B). The full coding sheet is available on request.

The table counts the findings for each lever. “Best” is the strongest evidence level behind any finding for that lever. The last column is our reading of the evidence in November 2026.

4.1 Being retrieved comes first

Chen and Koudas (2026, S46) analyzed 158,099 answers and 468,037 citations from ChatGPT, Gemini and Perplexity over 90 days in two markets. Every page that was cited had first been retrieved by the engine’s search. Whether a page was retrieved explained 32% of the variation in citations between pages. All page features together (length, structure, markup, speed) explained 2.8%.

Rank inside the engine’s search results matters too. When S46 replayed its questions through the API, ChatGPT cited 43.8% of the pages its search ranked first, 12.2% of pages at ranks 6 to 10, and 3.5% at ranks 11 to 20. In the smaller B2B market, the rank-1 rate was 57.1%. An AirOps study of 16,851 queries run on the ChatGPT interface found a steeper curve: 58.4% at position 1 and 14.2% at position 10 (AO-a). The two sources differ in level, and both show the same steep drop.

The same pattern shows up at the brand level. In a panel of 34,960 unbranded prompts across 75 projects (Tannenbaum 2026, S47), a brand was named in 2.8% of GPT answers and 3.8% of Gemini answers when none of its own pages was among the sources. When one of its pages was cited, those numbers rose to 49.0% and 58.4%. Within the same prompt, the gain was about 40 to 49 points. S47 is observational and its author runs a vendor in this space, so we treat the size of the gap with care. The direction matches S46.

The engine’s search is a different index from Google’s. In S46, 97% of the pages ChatGPT retrieved were outside Google’s top 50 for the same query. For Google’s own surfaces the overlap is larger but still partial: 53% of AI Overview domains were outside the organic top 10 (Kirsten 2026, S30), and URL overlap between Google organic, AI Overviews and Gemini ranged from 0.11 to 0.18 Jaccard (Grossman 2026, S31). A good Google ranking is weak evidence of AI visibility.

4.2 The page that answers the question wins

Query-content match is the most consistent lever in the base: six positive findings out of seven, at every evidence level from fixed-context tests to live engines and Chinese-language engines. Within the same site, a one standard deviation increase in how closely the title matched the question went with 14.1% more citations, and the same increase for the best-matching paragraph went with 15.5% more (S46). Both results passed all five of the study’s robustness tests.

Controlled studies point the same way. In 252,000 paired trials across six models (Vishwakarma 2026, S37), four factors decided which of two documents was cited first: topic match, price, recency and position. Formatting was negligible. Models follow text that matches the question over text with human credibility cues such as references or a neutral tone (Wan 2024, S04). Adding content for the questions a page implies but does not answer gave up to +15.9% on objective visibility in a fixed-context test (S52).

The practical reading is that one page per main question, with the question in the title and a self-contained answer right under it, gives the engine’s ranker the closest possible match. A page that covers many questions matches none of them as well. A clear heading on a broader page serves close sub-questions.

4.3 Brand mentions come from other sites and from memory

A citation of your page and a mention of your brand are different outcomes. In S46, when ChatGPT named a brand that the question did not name, 46% of those mentions had no cited page naming the brand: the model named it from memory. Large brands benefit most. One vendor dataset found global brands in 73% of first-run answers, mid-market brands in 44% and niche brands in 11% (Kumar 2026, S58). Startups were recognized 99.4% of the time when named in the prompt, and surfaced in only 3.32% of discovery queries on ChatGPT (Sharma 2026, S40).

For brands the model does not already know, third-party pages carry the weight. In a niche B2B market, a third-party page that named the vendor multiplied the odds of a ChatGPT brand mention by 39.7 (S46; 4.2 in a consumer market). Across four Chinese engines and 160,860 citations, the number of different sources mentioning a brand was the strongest predictor of the brand reaching the answer: 2.1% with one source, 59.9% with eleven or more (Zhen 2026, S49). That result comes from Chinese-language engines only.

Most citations go to sites the brand does not own. In the S46 consumer market, the client’s own sites received 11.1% of citations, competitors 13.8%, and third-party sites 75.0%. In the niche B2B market, the client received 2.6%, competitors 46.7% and third parties 50.8%. A vendor dataset of 149,912 citations found 2.9% going to the brand’s own domain and 75.2% to other companies in the same space (S58). Third-party content made up about 69% of citations on Chinese engines (S49), and AI engines over-represent earned media compared with Google (Chen 2025a, S14).

Ranked “best-of” lists are the format to aim for. In S58, listicles were 35.7% of content citations and 21.0% of all citations, more than any other format. An AirOps study of 21,311 brand mentions found that nearly 90% of third-party mentions came from listicles, comparison pages and review roundups (AO-b). S46 adds a caveat: at equal rank, ChatGPT consults list pages often and cites them less (×0.54).

4.4 Dates and prices matter for time-sensitive questions

Recency had four positive findings. In S37, two anonymized documents were placed side by side and six models chose which to cite first. A 2026 timestamp beat a 2019 one in every model, with odds ratios from 14.4 (GPT-5 Nano) to over 10,000 (Gemini 2.5 Flash, Claude 3.5 Sonnet, GPT-5.2). A visible price beat no price in every model too, with odds ratios from 6.3 (GPT-5 Mini) to over 10,000. The authors list topic match, price, recency and list position as the four factors on which all six models agreed.

Live data is weaker. In S46 the within-site effect of page age was small (−5.1% per standard deviation) and passed only four of five tests, and the number of prices mentioned on a page had no effect. ChatGPT did add the current year to 23% of its searches, which favors pages that show it. On Chinese engines, the ages of cited pages fit a decay curve with a scale of 39 days for time-sensitive queries and 68 days for others (S49). That figure describes how old cited pages are. It does not track a page losing citations over time. We read recency and price as factors for time-sensitive and commercial queries, not universal signals.

4.5 Structure has a small effect on credit, none on admission

Formatting results are mixed. Headings, lists, tables and length were flat within site (S46) and negligible in paired trials (S37). The best causal evidence comes from Selvam and Ghosh (2026, S48), who replayed frozen agent transcripts and changed only how documents were rendered. Structured rendering added 0.50 citations per answer for the target document (95% interval +0.20 to +0.84), and the effect on whether it was cited at all (+4.5 points) was not significant. Structure concentrates credit on a page that is already in the context. It does not get the page into the context. One pipeline study found structural fields could help retrieval without helping citation (S18).

Lab work on structure gives modest numbers. Yu et al. (2026, S42) report a 17.3% citation gain from a full set of structural changes (45.0% to 52.8%), with paragraph-level changes accounting for about 40% of it. The same paper finds that chunks under 150 words fragment the text and reduce citation probability by 23%, and it recommends paragraphs of 150 to 300 words. In S46, headings went with 7.7% more citations in the niche B2B market (four of five tests) and had no effect in the consumer market. We found no study in which splitting text into short paragraphs doubled a page’s appearances in AI search.

4.6 Crawler access

Two results point in different directions. Grossman et al. (2026, S31) found that sites blocking Google’s AI crawler (Google-Extended) in robots.txt were significantly less likely to be retrieved by AI Overviews, even though AI Overviews could still access their content. Twenty-one major publishers that block Google-Extended, including the New York Times, the BBC and Reuters, were never cited by Gemini in that study. S46 found that robots.txt blocking, read at download time, does not keep pages out of answers: 7.1% of pages ChatGPT cited disallowed its search crawler (OAI-SearchBot), and 19.9% of pages Gemini cited disallowed Googlebot. Both results are associations. Within a site, pages whose main text needs JavaScript received 10.6% fewer Gemini citations (p = 0.008) and, not significantly, 6.8% fewer from ChatGPT (S46).

4.7 FAQ, schema and author boxes: a confound

This is the clearest case of a cross-site illusion. Across sites, pages with FAQ blocks, schema markup, summary boxes or author boxes were cited 7.5% to 20.9% more (S46), which matches the associations in cross-site audits such as the GEO-16 framework (S66). Within the same site, the effects shrank to FAQ −3.4%, schema −5.3%, summary +6.9% and author −4.6%, none significant. Popular sites use these elements and popular sites get cited. Within a site, the elements themselves add nothing measurable. We found no study linking llms.txt or Markdown copies of pages to citations.

4.8 Rewrite recipes do not travel

The original GEO paper (Aggarwal 2024, S01) reported that adding quotations raised a source’s position-adjusted word count from 19.3 to 27.2, about 41%, with statistics and cited sources also helping. That test placed the source in a fixed set of five documents. When C-SEO Bench retested common rewrite methods across six domains, only 3 of 54 method-domain cases were significantly positive, and none in question answering (Puerto 2025, S09). In e-commerce, 10 of 15 common heuristics were neutral or negative (Bagga 2025, S17).

Learned rewrites, optimized per engine or per document, do better in fixed-context tests (+36% average in AutoGEO, S12). The one study that ran them through a full retrieval pipeline found the opposite end to end: body-only rewrites cut top-20 retrieval by about 9%, top-10 after reranking by 16%, and final citation by 6% (Kim 2026, S18). A defense built for this purpose cut the success rate of GEO rewrites from 50.3% to 6.2% while keeping 94% of benign evidence (Li 2026, S55), and simulations show citation-seeking rewrites adding unsupported claims (S54). Gains also erode as competitors adopt the same method (S09, S51).

4.9 Manipulation works in the lab and is being closed

Hidden instructions aimed at the model have four positive findings. On controlled pages, preference manipulation raised a fictitious product’s recommendation rate from 34.0% to 59.4% (Nestaas 2025, S07). These are black-hat techniques, engines are adding defenses (S11, S55, S68), and classic keyword stuffing already lowers visibility (S01) or is filtered by the retrieval stack (S60). We list them for completeness and do not use them.

4.10 Traffic: no causal evidence yet

Only two studies measure traffic. A production test at Pinterest reported a 20% traffic lift against a control, without describing assignment or uncertainty (S41). A single-site study of ChatGPT referrals found that raw referral traffic grew 5.7×, most of it explained by the growth of ChatGPT itself (3.5×); the treated pages showed an extra 1.82× multiplier, and the placebo test did not rule out chance (p = 0.16) (Watanabe and Nakayashiki 2026, S39). Nobody has yet shown that a GEO program causes more traffic or leads.

Several studies in the base are about measurement itself, and their results set hard limits on what a single report can claim.

From these results we use the following rules in client reporting.

• Run each prompt 7 to 8 times per engine and report the average.
• Report over periods of two to four weeks, never single days.
• Report each engine on its own. ChatGPT, Gemini, Perplexity and AI Overviews cite different pages.
• Split branded prompts (“is [brand] good for X?”) from unbranded ones (“best tool for X”). Branded prompts name the brand almost every time and hide the real discovery problem.
• Report “brand named in the answer” and “client page cited” as two separate metrics.
• To claim an effect, compare changed pages or prompts with unchanged ones. AI referral traffic grows on its own as AI search grows (S39).

These recommendations follow the ranking of the evidence. The first three rest on live-engine data; the rest are lower-cost hygiene.

1. Cover the searches the engine runs. For each priority prompt, record which pages each engine cites today. List the short searches behind the prompt (“best X for Y”, “X vs Y”, “X price 2026”) and check that a page exists for each. Put brand, product and category names in titles and headings.
2. Answer one question per page or section. Use the question as the title or heading. Under it, write a two to three sentence answer that repeats the key words of the question and makes sense on its own.
3. Get named on the pages the engines already cite. List the third-party pages cited for your prompts: reviews, comparisons, “best-of” lists, media, forums. Aim to be listed and correctly described on many different sites. This matters most for brands the model does not already know.
4. Show dates and prices as text, and keep them current on time-sensitive pages.
5. Let AI crawlers read the page. The main answer should be in the HTML without JavaScript (within a site, JavaScript-only pages received 10.6% fewer Gemini citations in S46). Do not block the search crawlers of the engines you care about. Blocking Google-Extended went with fewer AI Overview retrievals in S31, so treat it as a trade-off the client chooses knowingly.
6. Use focused paragraphs under clear headings. One idea per paragraph, and a heading that says what the section answers. Very short fragments can hurt (S42). The effect on citation is small. Do not promise much from it.

And what we stop selling as AI-citation levers:

Heavy reliance on one study. S46 supplies the within-site results, the retrieval variance decomposition and several brand-mention figures. It is a preprint from October 2026, it measured API endpoints rather than consumer apps, and it covers English in the US. Its results are associations, strengthened by within-site design and robustness tests, not experiments.

Vendor authorship. Several live-engine studies (S47, S49, S50, S58, S63) come from companies that sell visibility tools. We kept them because they hold the largest live datasets, and we flag their weight where it matters.

Preprints and coding depth. 43 of 68 studies are preprints. Most were coded from the Martinez (2026) matrix rather than read in full. Ten results come only or partly from another paper’s report and were not checked.

API versus consumer apps. S46 queried GPT-5.4 nano with web search, Gemini 3.1 Flash Lite with Google Search grounding and Perplexity Sonar through their APIs, once a day from 24 June to 21 September 2026. Its own authors note that the APIs need not behave like the consumer apps. The AirOps position figures (section 4.1) come from the ChatGPT interface, and they are higher.

Industry reports. Two figures come from AirOps reports, which are vendor research and not peer-reviewed. We cite them because they match the direction of the academic data, and we do not count them in the lever table.

Vote counting. Counting directions ignores sample size and effect size. We mitigate this by showing the best evidence level next to each count and by leaning on level A to C results in the text.

Coverage and decay. Most studies cover English and the US in 2025 and 2026. Engines change their retrieval and citation behavior often, and the gain from any tactic shrinks as competitors adopt it. We plan to update this review as new studies appear.

The biggest gap in the literature is causal evidence on live engines. Our next studies use the same within-site, changed-versus-unchanged design on client and owned properties, with repeated runs per engine, to test the two levers the review ranks highest: coverage of the engine’s fan-out searches, and one-question-per-page rewrites. A separate study looks at Claude, where apps are suggested by the model rather than found through web search.

The studies that carry the conclusions above, with their design and main limitation.

This review grew out of an internal summary. We traced each number in that summary to its source and corrected it where the study says something different.

1. AO-aAirOps & Indig (2026). The Fan-Out Effect: What Happens Between a Query and a Citation. Industry report. www.airops.com/report/the-fan-out-effect-what-happens-between-a-query-and-a-citation
2. AO-bAirOps (2026). The 2026 State of AI Search: How Modern Brands Stay Visible. Industry report. www.airops.com/report/the-2026-state-of-ai-search
3. ·Martinez (2026). Critical survey of generative engine optimization research. arXiv 2607.14035. arxiv.org/abs/2607.14035
4. S01Aggarwal et al. (2024). GEO: Generative Engine Optimization. KDD 2024. arxiv.org/abs/2311.09735
5. S03Li & Sinnamon (2024). Generative AI Search Engines as Arbiters of Public Knowledge. Proc. ASIS&T. arxiv.org/abs/2405.14034
6. S04Wan et al. (2024). What Evidence Do Language Models Find Convincing? ACL 2024. aclanthology.org/2024.acl-long.403/
7. S07Nestaas et al. (2025). Adversarial Search Engine Optimization for Large Language Models. ICLR 2025. openreview.net/forum?id=hkdqxN3c7t
8. S09Puerto et al. (2025). C-SEO Bench: Does Conversational SEO Work? NeurIPS 2025 D&B. proceedings.neurips.cc/paper_files/paper/2025/hash/27aa3aeff0f8460a7b43d30fa6c5c032-Abstract-Datasets_and_Benchmarks_Track.html
9. S12Wu et al. (2026). What Generative Search Engines Like and How to Optimize Web Content Cooperatively. ICLR 2026. arxiv.org/abs/2510.11438
10. S14Chen et al. (2025). Generative Engine Optimization: How to Dominate AI Search. arxiv.org/abs/2509.08919
11. S17Bagga et al. (2025). E-GEO: A Testbed for Generative Engine Optimization in E-Commerce. arxiv.org/abs/2511.20867
12. S18Kim et al. (2026). SAGEO Arena: Evaluating Search-Augmented GEO. arxiv.org/abs/2602.12187
13. S30Kirsten et al. (2026). Characterizing Web Search in the Age of Generative AI. Findings of ACL 2026. arxiv.org/abs/2510.11560
14. S31Grossman et al. (2026). How Generative AI Disrupts Search: Google Search, Gemini, and AI Overviews. SIGIR 2026. arxiv.org/abs/2604.27790
15. S34Schulte et al. (2026). Don’t Measure Once: Measuring Visibility in AI Search. arxiv.org/abs/2604.07585
16. S37Vishwakarma et al. (2026). What Gets Cited: Competitive GEO in AI Answer Engines. SIGIR 2026. arxiv.org/abs/2605.25517
17. S39Watanabe & Nakayashiki (2026). Disentangling AEO from Platform Growth: ChatGPT Referral Traffic. arxiv.org/abs/2606.04362
18. S40Sharma (2026). The Discovery Gap: How Startups Vanish in LLM Discovery Queries. arxiv.org/abs/2601.00912
19. S42Yu et al. (2026). Structural Feature Engineering for GEO. arxiv.org/abs/2603.29979
20. S41Zhang et al. (2026). GEO: A VLM and Agent Framework for Pinterest Acquisition Growth. arxiv.org/abs/2602.02961
21. S46Chen & Koudas (2026). What Drives AI Citations? Content, Retrieval and Model Knowledge in GEO. arxiv.org/abs/2610.15033
22. S47Tannenbaum (2026). From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search. arxiv.org/abs/2609.23162
23. S48Selvam & Ghosh (2026). CITECHOICE: A Causal Audit of How Document Presentation Redistributes Citation Credit. arxiv.org/abs/2609.15164
24. S49Zhen et al. (2026). What Do Chinese-Language Generative Search Engines Cite and Surface? arxiv.org/abs/2607.15771
25. S50Sielinski (2026). From Stochastic to Stable: Rank Stability and Structural Sufficiency in AI Visibility Measurement. arxiv.org/abs/2607.10341
26. S51Sourirajan et al. (2026). Beyond the Vacuum: Competitor-Aware Strategy Selection for GEO. arxiv.org/abs/2608.27631
27. S52Ramakrishna & Andreopoulos (2026). Query Implied Generative Engine Optimization. arxiv.org/abs/2609.27845
28. S54Xu, Guo & Xiong (2026). Mechanism Design for Generative Engines: From Exploitation toward Win-Win. arxiv.org/abs/2608.11390
29. S55Li et al. (2026). When Optimization Becomes Manipulation: Defending against Malicious GEO. arxiv.org/abs/2609.02964
30. S58Kumar (2026). GEO at Scale: Measuring Brand Visibility Across AI Search Engines. arxiv.org/abs/2606.20065
31. S63Sielinski (2026). Quantifying Uncertainty in AI Visibility. arxiv.org/abs/2603.08924
32. S66GEO-16 framework (2025). AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO-16 Framework. arxiv.org/abs/2509.10762