AI Citation Overlap 2026: Why 67% of Brands Are Named but Only 10% of URLs Are Reused

AI search visibility is not one number — it is three numbers that move independently. Across 596,723 prompts answered by two or more engines, only 10.2% of cited URLs appeared on more than one engine. Measured at the domain level, overlap rose to 17.9%. Measured on brand names in the answer text, it reached 67.4% (Wellows citation data via TechTimes, 1 September 2026).

That 6.6× gap between name overlap and URL overlap is the most consequential measurement finding of 2026. It means AI engines broadly agree on which brands belong in the conversation and almost entirely disagree on which page should be used as proof. A brand can win the recommendation and still lose the citation — and most GEO dashboards cannot tell the difference, because they count one of those things and report it as both.

October 2026: retrieval explains the gap, not citation selection

This page has argued since September that three decisions — discovery, retrieval and citation selection — must be measured separately. October turned that from an inference into a measurement. Chen & Koudas replayed the same commercial questions through the engines' own APIs and compared two candidate explanations for why one page gets cited over another page on the same website:

Source: Chen & Koudas, arXiv:2610.15033, 10 October 2026 (158,099 answers / 468,037 citations; within-website design; replicates in a small-B2B market).

Two of the paper's findings are directly actionable for anyone running a cross-engine program:

1. 1.
   Being named and being cited are separate mechanisms, on separate timelines. The model "names brands without any supporting page," and controlled experiments showed the brand name alone reorders its product rankings even when the facts are identical. That is why brand-name overlap sits at 67.4% while URL overlap sits at 10.2%: the name lives in the model, the citation lives in the retrieval pool.
2. 2.
   Query rewriting, not query matching, is the real gate. The engines rewrite sub-queries on every answer, and those rewritten queries "rarely coincide with Google's or Bing's results." This is the mechanism behind the ~80% of AI-cited pages that never appear in Google's top 10 — the citation pool is a different list, generated fresh each run.

"Pages are thus cited when searched; brands are also named when remembered."

Why exact-URL overlap is partly unmeasurable

Jiang et al. add the measurement caveat that reframes the entire "overlap" question. Their observational study of four ChatGPT Web settings (5.2 through 5.5, Instant and Thinking modes) tested Google Search and the OpenAI Responses API as external proxies for ChatGPT's own retrieval. Both matched ChatGPT better at the hostname and entity levels than at the exact-URL level, and even combined they recovered only about a quarter of cited URLs.

Source: Jiang, Zhang, Xing, Chang & Xiao, "Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations," arXiv:2610.15382, 11 October 2026. Observational study of four ChatGPT Web settings; both proxies combined recover about one quarter of cited URLs; supervised combination reaches AUC above 0.74 at all three levels for GPT-5.3 and GPT-5.5.

The paper's framing conclusion is worth adopting verbatim as a measurement rule: GEO should distinguish retrieval visibility from citation visibility, account for rewritten queries and run-to-run variation, and consider hostname- and entity-level presence in addition to exact URLs. And because the settings differ from each other, it adds a maintenance warning — GEO strategies should be re-evaluated whenever the deployed model setting or serving environment changes, a point that lands hard given the Reddit citation swing documented on the citation volatility page.

The finding: 10.2% at the URL, 67.4% at the name

The September 2026 dataset covering ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode measures the same prompts at three levels of granularity. Read together, the three figures describe a funnel rather than a score.

Source: Wellows citation data across ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode, published via TechTimes, 1 September 2026. Sample: 596,723 prompts answered by two or more engines.

The practical reading is uncomfortable for anyone reporting a single "AI share of voice." If your dashboard shows strong brand-name presence, you have established entity association — the engines believe you are relevant. That is genuinely valuable and genuinely hard. But it does not mean any page you own will be retrieved when an engine needs to attach a source, and the 10.2% figure is the evidence that it usually will not be.

How much do two engines' source pools actually overlap?

The 10.2% figure is an average across five surfaces. Pairwise, the spread is wide — and which pair you care about changes the strategy. A July 2026 study of 161,286 prompts compared six platform pairs at the domain level (Writesonic). Two results bracket the whole field:

Source: Writesonic, 161,286 prompts, July 2026, domain-level overlap across six platform pairs. Because the measure is domain-level, strict URL-level comparison would produce lower figures still.

The practical consequence: the two engines you would most expect to converge — ChatGPT and Gemini — are the furthest apart. A program built to win ChatGPT citations is, on this evidence, doing almost nothing for Gemini. Perplexity and Google AI Overviews are the closest pair, which makes sense: both run retrieval patterns that sit nearer to classic search ranking than ChatGPT's source pool does.

Citation selection vs. citation absorption: ChatGPT cites least, absorbs most

Counting citations understates engines that use fewer sources but rely on them harder. A peer-reviewable 2026 study of 602 controlled prompts that logged 21,143 citations and 23,745 feature records split the problem in two: citation selection (is your page named as a source?) and citation absorption (how much does the page actually shape the answer?).

Two implications follow. First, citation count is not a measure of content usage — ChatGPT cites fewer than half as many sources as Perplexity and leans harder on each one. Second, the study found pages containing code blocks scored 76.88% higher on influence than pages without, and that high-influence pages differ sharply in length and heading structure. Formatting is a citation variable, not a style preference.

Google rank is a weak predictor — except in AI Overviews

Where an engine draws its pool from determines how much classic ranking still matters. An Ahrefs test of 15,000 long-tail questions run through Google, Bing, ChatGPT, Gemini, Copilot and Perplexity produced a split that should end the "just rank higher" advice:

A separate line of work reaches the same conclusion from the source-type direction: Chen, Wang, Chen and Koudas (arXiv:2601.16858, University of Toronto, 2026) found AI engines show a systematic bias in source type — Wikipedia, news media and government sites are cited at rates far above their share of classic search results — and that brand visibility consistency across models is far lower than expected. Being cited in ChatGPT tells you almost nothing about Gemini.

Why naming and citing are different operations

The 6.6× gap is not a measurement artifact. It follows directly from how generative search is built, and both major platform disclosures describe the same three-stage pipeline:

1. 1.
   Discovery. A page must be crawled and eligible at all. OpenAI describes ChatGPT search as a system that retrieves web sources and attaches citations; Google's documentation states pages need ordinary Search eligibility and snippet eligibility before they can be used in AI features. Without this stage, the other two cannot happen.
2. 2.
   Retrieval. The engine selects candidate passages. This is where engines diverge most, because each runs its own retrieval system over its own source pool — Bing's index for ChatGPT Search, Googlebot for AIO and AI Mode, PerplexityBot's index for Perplexity.
3. 3.
   Citation selection. A specific URL is attached as proof. The engine optimizes for the passage that best supports the sentence it just wrote, not for the brand it just named.

"A brand can become part of the category answer while its own evidence layer remains fragmented. The model may know the brand belongs in the conversation, but each engine may still reach for a different proof source when it has to cite the answer."

Brand-name overlap stays high because entity association is built from repeated descriptions across the entire web — press coverage, review sites, listicles, forums, directories — not from one optimized page. That is a slow, distributed, cumulative process, and it is largely why the same handful of companies get named everywhere. Citation overlap stays low because step three is a local, per-query decision made against each engine's own candidate set.

The 2.8% problem: recommendation without citation

A second September 2026 dataset makes the commercial stakes concrete. Shero Commerce analyzed buying prompts across Google AI Mode, ChatGPT and Perplexity and examined 1,851 cited sources. The result:

This is the gap most GEO programs are blind to. Standard reporting counts name mentions, so a brand can show rising "AI visibility" while the evidence layer underneath every answer about it is owned by competitors, affiliates, listicle publishers and review sites. Those third-party pages decide the framing, the price shown, the feature list, and the caveats attached to your name.

The risk compounds with the finding from Counter-GEO-Bench (arXiv preprint, 2 September 2026, accepted to EMNLP 2026): a brand can be cited next to distorted claims it never published, and the tested safety filters reduced attack success by no more than 5.7%. Recommendation without a citation you control is a reputation exposure, not just a missed link. See the GEO evidence audit for the full treatment.

How each engine's citation pool behaves

Coverage has to be planned per engine, because the pools barely intersect. The table below summarizes what 2026 research has established about each surface.

Sources: Sensor Tower "State of AI Report 2026" and Similarweb Global AI Tracker for assistant share; OpenAI and Google Search Central documentation for index architecture; Search Engine Journal / Ahrefs for AIO and AI Mode source preferences; Nico Digital for Perplexity citation timelines.

The 4-step playbook for a citation pool that survives contact with any engine

Because you cannot optimize a 10.2% overlap away, the goal is not to force engines to agree. It is to be a plausible proof source in every pool at once. Four steps:

1. 1.
   Get into every pool first. Nothing downstream matters if a crawler has never fetched you. Allow the retrieval crawlers — OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot, Google-Extended — and note that a September 2026 audit found 17% of major sites still block PerplexityBot. Full workflow in the indexing guide.
2. 2.
   Build claim-level pages, not brand pages. Engines cite passages that prove one specific assertion. A page titled around a single claim — with the answer in the first two sentences, a named statistic carrying a source and date, and a current dateModified — is retrievable by every engine, not just the one you were thinking about.
3. 3.
   Make the same facts verifiable off your own domain. With only 2.8% of citations going to brand-owned pages, third-party corroboration is not optional. Ensure pricing, certifications, limits and specifications are consistent across your site, review platforms, partner pages and press coverage — inconsistency suppresses credibility scoring.
4. 4.
   Apply the Princeton lifts to the page you want cited. Expert quotations +41%, statistics with named sources +33%, fluency +29%, external citations +28% — and keyword stuffing −8%. These operate at the passage level, which is exactly where citation selection happens. See the Princeton benchmark.

How to measure the three columns

Replace the single visibility percentage with three tracked metrics, sampled across a fixed prompt set, per engine, at a fixed cadence:

Two cautions follow from the wider evidence base. First, sample repeatedly, never once: the July 2026 critical survey of 45 GEO studies found day-to-day source overlap as low as a Jaccard similarity of 0.34–0.42 across repeated runs, and an August 2026 volatility event saw Reddit's ChatGPT citation share fall 86% in four days. A single screenshot proves nothing — see the citation volatility analysis. Second, track accuracy alongside volume, because a rising mention count with deteriorating framing is an incident rather than a win.

Frequently asked questions