The GEO Claims Audit

The most repeated claim in GEO, and the one with the strongest evidence against it. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matched against control pages that never added it, and ran four separate statistical tests including a matched difference-in-differences analysis. No platform showed a meaningful citation increase. Google AI Overviews showed a 4.6% decline for treated pages — the only statistically significant result in the study.

Where the myth comes from: an earlier Ahrefs analysis of 6 million URLs found AI-cited pages roughly three times more likely to carry JSON-LD. That correlation is real. It is also explained by the fact that sites which implement schema tend to be well-maintained sites doing everything else right.

What is actually true: schema aids entity disambiguation, rich results, and knowledge graph association. Those are real benefits. Citation frequency is not one of them.

Google's Gary Illyes confirmed in July 2025 that Google does not support llms.txt and has no plans to. John Mueller compared it to the discredited keywords meta tag, and separately noted that server logs show AI crawlers do not even check for the file. As of Q1 2026 no major AI provider — OpenAI, Google, Anthropic, Meta, or Mistral — has publicly committed to reading it in production systems.

Adoption sits around 10% of domains in a study of 300,000 sites. One crawler analysis of over 500 million AI bot visits across 90 days found 408 requests targeting llms.txt directly.

What is actually true: llms.txt has emerging use in agentic and developer-tool contexts. It has no documented effect on search or answer-engine citation.

OpenAI documents these as separate crawlers with separate purposes. OAI-SearchBot governs whether a site can be surfaced in ChatGPT search results. GPTBot governs content that may be used to improve OpenAI's foundation models. ChatGPT-User is a user-triggered fetcher and controls neither.

Blocking OAI-SearchBot removes a site from ChatGPT search answers regardless of GPTBot status. This is the most consequential technical error we find in audits, because it is invisible — the site appears correctly configured to anyone who has not read the documentation.

Google-Extended is a control for certain Google AI training and grounding systems. Googlebot remains the relevant crawler for Google Search and its AI features. Google states that eligibility for AI Overviews and AI Mode requires only that a page be indexed and eligible to appear with a snippet, with no additional technical requirements.

Google removed FAQ rich results from Search on 7 May 2026, eliminating the traditional benefit. Google has never confirmed that FAQPage schema influences AI Overview or AI Mode citation, and the Ahrefs controlled study pooled FAQ schema with other types and found no citation effect overall.

Claims attributing specific percentage lifts to FAQ schema trace back to vendor experiments that changed page structure, content, and internal linking simultaneously.

No controlled study isolates page performance as a variable in AI citation selection. Performance affects Google indexing and page experience, which is a precondition for AI Overview eligibility — so an indirect path exists. No published evidence establishes a direct effect.

The foundational GEO paper — Aggarwal et al., arXiv, 16 November 2023, later accepted at KDD 2024 — tested nine content interventions against a benchmark of roughly 10,000 queries. Adding citations, quotations, and statistics, and improving fluency, produced the strongest visibility gains, reported at up to approximately 40% under test conditions.

The caveat that gets dropped: that figure is a relative, experiment-specific result on controlled rewrites, not a forecast for a commercial website. The paper reports that effectiveness varies by domain.

Ahrefs analysed 1.4 million ChatGPT prompts and found that pages covering 26–50% of ChatGPT's internally generated sub-queries were cited more often than pages covering 100% of them. Depth on a focused scope outperformed breadth.

The "ultimate guide" format that dominated traditional SEO appears to work against citation when query relevance is held constant.

In the same 1.4 million prompt analysis, pages with headings closely matching the query were cited 41% of the time, against 29% for weak matches. Cited pages showed a median title-to-prompt cosine similarity of 0.602 against 0.484 for uncited pages — rising to 0.656 when measured against the sub-queries ChatGPT generates internally.

Heading structure was the strongest on-page lever tested, ahead of word count, topical breadth, and body copy.

Freshness is plausible and is widely assumed. No controlled study isolates it from the content changes that usually accompany an update. A page that is republished is typically also rewritten, restructured, and re-linked.

The original GEO experiment found information-rich rewrites outperformed keyword-oriented ones. Nothing in the subsequent literature supports repetition as a citation strategy, and Ahrefs found semantic similarity — not term frequency — predicts selection.

E-E-A-T is not a ranking factor in Google Search. Google has stated repeatedly that there is no E-E-A-T score, that the concept comes from the Search Quality Rater Guidelines, and that rater assessments do not directly affect rankings. Raters evaluate results to test and refine ranking systems.

Extending an already-misunderstood Search concept to AI answer engines, where no operator has published anything comparable, compounds the error.

What is actually true: the underlying qualities — accuracy, transparent authorship, genuine expertise — correlate with things that do get measured. The acronym is not itself a lever.

In controlled testing of ChatGPT citation signals, domain authority and backlinks showed no positive correlation with citation, and were slightly inversely correlated. ChatGPT appeared to evaluate content on relevance and structure rather than on authority signals.

This is the finding most at odds with how the SEO industry has approached GEO.

Overlap studies consistently show partial correlation at best. seoClarity reported approximately 32% of AI Overview citations overlapping Google's top ten. Ahrefs reported approximately 37.9% across 863,000 keywords and roughly 4 million URLs.

For ChatGPT the gap is far wider: only 6.82% of ChatGPT search results appeared in Google's top ten for the same queries, and only 16.61% appeared anywhere in Google's organic results.

What is actually true: ranking correlates with citation and is worth having. It is neither necessary nor sufficient, and for ChatGPT specifically it explains very little.

This is the most misreported finding in AI search. The headline statistic — Reddit as the most-cited domain — is true for some engines and flatly false for others.

Analysis of OpenAI's search behaviour found Reddit pages appearing as candidate sources in 76% of ChatGPT searches, but only 0.61% were selected — 491,024 Reddit pages retrieved, 3,012 cited. A 99.39% rejection rate. For Claude, one analysis found zero Reddit citations across 139,601 grounding sources between May and July 2026; Reddit was not supplied as a candidate at all.

Gemini, Google AI Mode, Google AI Overviews, and Perplexity do cite Reddit heavily. Those four run live web retrieval.

What is actually true: a Reddit strategy targets four engines and is ignored by two, including the two that favour editorial and primary sources. Anyone quoting a single cross-engine Reddit percentage is averaging across systems that behave in opposite ways.

Wikipedia is among the most-cited domains, particularly for ChatGPT. But citation share moved sharply within weeks: ChatGPT's Wikipedia citation rate fell from roughly 55% of prompt responses to under 20% in mid-September, while remaining stable on other engines.

Being cited from Wikipedia is also not the same as a brand benefiting. The citation credits Wikipedia.

The direction is consistent across many datasets. Semrush reported 4.4x. Ahrefs found AI traffic at 0.5% of sessions driving 12.1% of signups. Shopify reported roughly 50% higher conversion on AI-referred sessions. Visibility Labs found a far smaller 1.3x across 94 ecommerce brands. BrightEdge found the opposite for Fortune 100 brands through August 2025.

Why it is still unsupported: every one of these is observational. The conversion event is defined differently in each — form fills, signups, purchases — so the multiples are not comparable. And the causal story runs backwards: AI users arrive later in their decision process, which the AI did not cause. Selection effect, not treatment effect.

What is actually true: AI-referred visitors behave differently and are worth tracking separately. No controlled study establishes that AI referral causes higher conversion.

Pew Research Center found users clicked a traditional result in 8% of visits where an AI Overview appeared against 15% where none did, across 68,879 searches from 900 US adults. Ahrefs found AI Overview presence associated with 58% lower average CTR for the top-ranking page. Seer Interactive reported CTR falling from 1.76% to 0.61% across its tracked queries.

Three independent datasets, three methodologies, same direction. This is the best-evidenced claim in the audit.

Semrush tracked AI Overviews on 6.49% of its keywords in January 2025, 24.61% in July 2025, and 15.69% in November 2025. Published figures elsewhere range from roughly 15% to 60%.

These are not contradictory. They measure different keyword panels, countries, devices, dates, and query mixes. Any single percentage presented without those parameters is not a measurement.

The number is real and comes from Aggarwal et al. It is the maximum relative improvement observed for the strongest intervention, on controlled rewrites, against a benchmark query set, in 2023. It is not an expected outcome for a commercial website, and the paper reports variation by domain.

Quoting it as a service benchmark misrepresents the source.

ChatGPT cited Reddit in close to 60% of prompt responses in early August before falling to around 10% by mid-September. Wikipedia fell from roughly 55% to under 20% in the same window on ChatGPT while holding steady on other engines.

Separately, a 2026 study found identical prompts producing distinct outputs roughly 25% of the time on GPT-4o-mini, and query wording accounting for approximately 26.5% of response variance.

What is actually true: citation position is a rate with an error range, measured over repeated runs. A single observation is not a state.

Yext explicitly documents that citations returned by APIs may differ from those shown in the consumer application. Most commercial tools collect via API because it is easier to automate.

The divergence between engines is also larger than most reporting suggests: one cross-platform analysis found only 11% of cited domains overlapping between ChatGPT and Perplexity for identical queries.

What is actually true: each tool reports visibility under its own prompt set, collection method, product mode, and citation definition. Two tools disagreeing does not mean one is wrong.

Ahrefs found that roughly half of URLs ChatGPT retrieves are cited, and in a separate cut, that 85% of retrieved pages went uncited. Retrieval and selection are distinct stages with distinct determinants.

A site can be perfectly crawlable, correctly indexed, retrieved for the right query, and still never appear in the answer.

ChatGPT generates internal sub-queries — fan-out — when answering. 95% of those fan-out queries had zero monthly search volume by conventional metrics, and 32.9% of cited pages appeared only in results for a fan-out query rather than the original prompt.

Keyword tools built on search-volume data are structurally blind to roughly a third of citation opportunities.