What ChatGPT Actually Cites in 2026 (Data Analysis Across 6 Public Studies)

Name the three domains ChatGPT cites most, right now, without opening a tracker. If you can, skip this. If you can't, keep reading.

Six public studies have measured ChatGPT citation behaviour at scale. Samples range from 100 queries to 680 million citations. Methods vary. They converge anyway. ChatGPT cites less than the industry pretends. When it does cite, the pool is small enough to memorise. What follows is a synthesis of the six, cross-checked against 17,112 first-party Copilot citations from GoGoChimp's own tracker (Bing Webmaster Tools, verified 2026-07-29).

I'm tired of watching agencies quote one chart they half-read on LinkedIn and pitch a full retainer off it. One study is a data point. Six studies stacked and cross-checked against your own tracker is what you can actually price a strategy off.

How six public studies of ChatGPT citation behaviour actually converge

Six independent measurements, from a 100-query study to a 680-million-citation sweep, land on the same two findings. ChatGPT often cites nothing at all. In the US, its zero-citation rate doubled from 28% to 48% in March 2026 (seoClarity, 2026). Citations rebounded in May. When it does, Wikipedia carries 47.9% of the top-10 source share and 7.8% of all ChatGPT citations (Profound, 2025).

Everything else is texture on those two.

Most agencies pick their favourite of the six and quote it forever. That's how you end up quoting Perplexity data at a ChatGPT client. The six studies overlap to some degree but the finer details are where they vary. Anyone who doesn't understand this is quoting you the wrong retainer.

Study	Sample	Headline finding	What it settles
Lily Ray, self-promotional listicle study	100 B2B "best software" queries, 323 self-promotional listicle citations, April to June 2026	When a brand's own listicle was cited, AI Overviews left that brand out of the recommendation 69% of the time (Ray, 2026)	Citation isn't recommendation
Nelson Brassell (Ten Speed, data via Peec AI)	7,387 citations, 170 B2B evaluation prompts	88.3% of B2B evaluation-stage citations are brand-controllable, product pages first at 24.1% (Ten Speed, 2026)	Funnel-stage matters
Superlines platform tracking	34,234 AI responses on one brand, 10 platforms, 30 days	ChatGPT: 0.59% citation rate, 0.14% brand visibility (Superlines, 2026)	Engines differ wildly for the same brand
Nectiv fan-out study	About 4,000 prompts, GPT-5.6 Sol (Nectiv, 2026)	Fan-out searches per prompt more than tripled, from 2.17 to 7.61 on average	Each prompt now triggers far more searches
Profound Citation Patterns	680 million citations analysed	Wikipedia 47.9% of ChatGPT top-10; Reddit 46.7% of Perplexity top-10	Each engine has its own corpus
seoClarity citation tracker	Millions of ChatGPT interactions, 5 markets, February to May 2026	US zero-citation rate doubled from 28% to 48% in March 2026, then citations rebounded in May (seoClarity, 2026)	Silence is common, and it swings

Where do they disagree? Nectiv shows ChatGPT running more than three times as many fan-out searches per prompt as a year ago. seoClarity shows its citation rate swinging within weeks. So any single citation-rate figure is a snapshot. Ray measured AI Overviews, not ChatGPT proper, so her 69% figure is directional for ChatGPT, not identical. Superlines tracked one brand, so its rates describe that brand, not the market. Read every study with the engine, model version, and funnel stage attached, or you're comparing apples to a chart.

ChatGPT stays quiet on most answers, and when it does cite, it names a small pool weighted toward Wikipedia. Your job is to be in that pool.

Six independent studies covering more than 680 million citations converge on one uncomfortable finding. ChatGPT often answers without citing anyone, and when it does cite, the pool is small. In the US, its zero-citation rate hit 48% in March 2026 (seoClarity, 2026). It rebounded in May.

Why ChatGPT often answers without a source

ChatGPT often skips citing anything at all. The mechanic is retrieval cost: when the model already knows the answer with high confidence from training, it answers direct and skips the source stack.

ChatGPT decides what to cite in two steps. First it retrieves sources when browsing or search is on. In the US, the share of ChatGPT answers with no citation at all doubled from 28% to 48% in March 2026 (seoClarity, 2026). It rebounded in May. Second, it recommends entities it already knows from its training corpus. Recommendation and citation are separate signals. Your page can be the cited source while a competitor is the named recommendation.

Agencies still pitch "if your content is helpful enough, ChatGPT will cite it." Helpfulness never enters the retrieval decision. Retrieval cost, index freshness, and prompt phrasing do. Silence is the default. Your prompt phrasing is the biggest lever you actually control. "Best X UK 2026" triggers retrieval far more often than an open-ended definitional question.

Why does ChatGPT skip sourcing so often?

Because the model treats most short factual answers as things it already knows. Retrieval costs money and adds latency. When training already carries the answer with high confidence, ChatGPT answers direct and skips the citation stack. That's why entity strength beats blog volume. If ChatGPT knows your brand from Wikipedia, it names you without fetching a page.

Does prompt length change the citation rate?

No large study isolates prompt length. Position in the conversation does matter. In Profound's analysis of about 700,000 ChatGPT conversations, turn 1 was 2.5x more likely to trigger citations than turn 10, and nearly 4x more likely than turn 20 (Profound, 2026). Ask your most important questions at the start of a fresh session.

Which domains does ChatGPT actually name?

When ChatGPT does cite, it names a smaller pool than most agencies claim. .com domains carry 80.41% of ChatGPT citations. .org carries 11.29%. .uk carries just 2.16% (Profound, 2025).

British content is structurally underweighted in ChatGPT's index. That's the single most important thing UK founders never hear. If you publish on a .co.uk and ChatGPT ignores you, part of the answer is the suffix. Anyone pitching your .co.uk domain as a "UK strength for AI" has the data upside down.

Top domains for ChatGPT are Wikipedia, Reddit at a smaller weight than in Perplexity, YouTube, major newspaper mastheads, and a long tail of niche trade publishers. Wikipedia's dominance is category-defining, not marginal. Compare Perplexity, where Reddit is 46.7% of top-10 source share on the same query class (Profound, 2025).

Which trade publishers punch above their weight?

Niche trades with strong Wikipedia entity links and dense internal citations. Search Engine Journal, Search Engine Land, Ahrefs' blog, and Baymard Institute all appear in the top-100 lists of every study we reviewed. Three properties in common: named authors, dated pages, dense inline citations. The retriever reads those as trust.

How does the .uk penalty affect UK brands?

You have to earn placement on the .com and .org surfaces to win ChatGPT citation share. Guest bylines on .com publishers. .org community-standard listings. A Wikipedia entity. Those are the three fastest counters to the domain bias. See our guide to entity SEO and brand mentions for the entity side.

The Wikipedia weighting nobody wants to sell you

Wikipedia is 47.9% of ChatGPT's top-10 source share and 7.8% of all ChatGPT citations (Profound, 2025). No other single domain gets close.

This matters more for ChatGPT than any other engine. Google AI Overviews leans on the classical Google index. Perplexity leans on Reddit and topical publishers. ChatGPT leans on Wikipedia. If your brand has no Wikipedia entity, ChatGPT has no persistent memory of you at the corpus level. It can still cite one of your pages on live retrieval. It won't name you when a user asks "who are the best X" without opening a browser.

The upstream implication runs against the industry's default advice. The "publish 20 more blog posts" pitch does nothing for this. A neutral, sourced, editor-approved Wikipedia article does. Wikidata does the same at machine-readable scale. Pick one entity move for ChatGPT visibility in 2026 and pick the Wikipedia entity. See how to get cited by ChatGPT for the full entity build.

Wikipedia is 47.9% of ChatGPT's top-10 source share. If your brand has no Wikipedia entity, ChatGPT has no persistent memory of you at the corpus level. Everything else in an AI SEO plan is downstream of that one build (Profound, 2025).

Why AI Overviews cite a vendor's listicle, then recommend its rivals

Lily Ray tracked 100 B2B "best software" queries in Google AI Overviews between April and June 2026. When a brand's own "best of" listicle was cited, AI Overviews left that brand out of the recommendation 69% of the time, in 224 of 323 cases (Ray, 2026). Ray measured AI Overviews, not ChatGPT. Treat it as directional for ChatGPT.

The mechanism repeats across engines. When a comparison query needs a shortlist, the retriever grabs the first coherent shortlist it can find. That shortlist is often the one a vendor wrote about itself. The retriever does not tell a neutral publisher listicle apart from a vendor's "best of" page written by the vendor. So a vendor with a well-structured, dated, well-schema'd "Best X 2026" page can get cited by ChatGPT while an actually neutral publisher gets skipped.

The industry sold "only neutral third-party listicles win citations." Ray's data shows vendor listicles do get cited. It also shows the vendor usually loses the recommendation to the category leaders it names. Two implications follow. First, your own comparison content can earn the citation, but only an established brand also wins the recommendation. Our best AI SEO tools 2026 list works on this principle. Second, source and subject often split. Your page can be cited while a competitor is the named recommendation. Measure both signals separately.

Why citation studies contradict each other at different funnel stages

Across AI search as a whole, 96% of citations point to third-party sources rather than the brand's own site, up from about 80% a few months earlier (Writesonic via Search Engine Journal, 2026). Reddit, YouTube, forums and industry publications carry the surface. This is the "earned media does the work" number.

At B2B evaluation stage, the split flips. 88.3% of citations are brand-controllable. Product pages are the #1 cited surface at 24.1%. Articles, blog and PR at 17.4%. Comparison content 13.3%. Listicles 13.2%. How-to 8.9%. Homepages 7.8%. G2 and Capterra directory profiles 7.2%. Reddit, YouTube and forums combined 4.2% (Nelson Brassell, Ten Speed, 2026). Sample: 7,387 citations across 170 B2B evaluation-stage prompts.

The single biggest measurement error in AI SEO right now is quoting one citation-share number without saying which funnel stage it came from. Every agency that hands you "earned media is 96% of citations" without naming the funnel stage is selling half the truth. Every one that quotes "product pages carry 24.1%" without naming it is selling the other half.

So the advice you give a client depends on the buyer's stage. At awareness, the ChatGPT playbook is Wikipedia, earned media, third-party listings. At evaluation, the ChatGPT playbook is a product page written for the retriever, a comparison page, and a G2 or Capterra profile. Both playbooks are true, but neither transfers between stages.

How do I tell which stage my target query sits at?

Read the query. "What is X" sits at awareness. "Best X 2026" or "X vs Y" sits at evaluation. "How to buy X in the UK" is late evaluation, going into transaction. Awareness routes through third-party citations. Evaluation routes through brand-controllable surfaces. Match the surface to the stage.

17,112 Copilot citations, and what the concentration curve looks like

GoGoChimp's Bing WMT report logged 17,112 Copilot citations in 90 days, running at roughly 1,050 a day. 87.25% of them concentrated in three listicle pillars. One page carried 49% of the total load. The top two pages carried 72% (Bing WMT, 2026-07-29).

Public studies measure ChatGPT. We can't measure our own ChatGPT citations directly. Nobody outside OpenAI can. What you can measure is Microsoft Copilot, because Bing Webmaster Tools ships a free AI Performance report for it, and Copilot shares the Bing retrieval index that ChatGPT Search also uses.

Here is what our tracker reads.

17,112 Copilot citations in the first 90 days (Bing WMT, 2026-07-29). Trailing 30 days: 14,347. Last 7 days: 7,353, at roughly 1,050 a day and climbing. Peak day: 1,420. Distinct cited pages per day up to 19.

The concentration matches the public studies almost line for line. 87.25% of our Copilot citations sit in three listicle pillars. One page carries 49% of the load. Copilot isn't ChatGPT. The concentration signature is the same one Profound describes: a small number of extractable, dense, dated pages doing almost all of the work.

The pitch that agencies sell hardest, spread thin content across dozens of topics, is the exact opposite of what wins here. Build ten listicles at proper depth and expect two to do 70%+ of the citation work. The rest earn tail traffic.

Across 17,112 first-party Copilot citations in 90 days, 87.25% concentrated in three listicle pillars, and one single page carried 49% of the total load (Bing WMT, 2026-07-29). The AI citation footprint is spike-shaped, not spread-shaped. Build for the spike.

The citation playbook that fits ChatGPT's actual bias

Five moves, ranked by lift per hour rather than by novelty. This is what we run for clients on the OperatorAI methodology (GoGoChimp's CRO methodology, distinct from OpenAI's Operator agent product released January 2025). Ordered from highest leverage.

Schema-first pitches sell easy, but the entity has to come first. If you buy schema before you build the entity, you paid to make an invisible brand marginally more machine-readable. That does nothing.

Fix 1: Own a Wikipedia entity

If Wikipedia is 47.9% of ChatGPT's top source share, the entity is the lever. A neutral, well-sourced Wikipedia article, plus a Wikidata Q-item with authority links, gives ChatGPT persistent memory of your brand. Nothing else touches this on cost-per-citation. Warning: Wikipedia's notability bar is real. Assemble the references first. Submit the article second. Expect at least one revision cycle.

Fix 2: Publish listicles with named entities in headings

Comparison content is the most-cited format in every study we reviewed. Build listicles with proper depth, one H2 per named entity, and an at-a-glance semantic HTML table near the top. The retriever lifts these into answers almost verbatim. Don't use markdown rendered as prose for comparison content. Use a real HTML <table>. That is the pattern behind our best AI SEO tools 2026 pillar.

Fix 3: Earn one third-party trade-press citation per quarter

Earned media at DA-70+ trade publishers earns disproportionate weight in ChatGPT's corpus. One quality byline on a recognised publisher does more than 40 blog posts on your own site. Target the trade publishers your buyers already read. Search Engine Land, Ahrefs, Search Engine Journal, Baymard Institute, and vertical-specific trades all qualify. Quarterly cadence is enough.

Fix 4: Answer-first structure, dated data, named authors

Every page opens with a 40-60 word answer capsule containing the entity, the claim, and a number. Every stat cites its source inline. Every page carries a visible "last updated" date. Every byline names a real author with a linked Person schema entry. 44.2% of ChatGPT citations come from the first 30% of a page, across 18,012 verified citations (Kevin Indig, 2026). Front-load the answer. Don't bury the lede.

Fix 5: Cite your own sources with ScholarlyArticle schema

Structured data doesn't cause AI citations directly. It does make your page cleaner for the retriever to grade. Ship Article schema with a named author, FAQPage schema on any question block, Organization schema with a complete sameAs list, and ScholarlyArticle schema on data-heavy pieces. See schema markup for AI SEO 2026 for the full stack. Combined with entity strength and citation density, schema turns a good page into an extractable one.

What the data does NOT show yet

Read this article as a snapshot, not a law. Anyone selling settled AI citation science in 2026 is selling something else.

ChatGPT isn't perfectly measurable from outside. OpenAI doesn't ship a first-party citation dashboard the way Bing does. Every public study samples prompts differently. Some route through browsing mode. Some through native ChatGPT. Some through the API. The results are directional, not definitive.

Model versions shift. Between GPT-4o and GPT-5.6, citation behaviour on the same prompt has moved. Semrush data showed the share of ChatGPT responses citing Reddit falling from close to 60% in early August 2025 to around 10% by mid-September, across more than 100 million citations (Semrush, 2025). What is true in September 2026 may not hold in December 2026. Track the direction, not the number.

Prompt-level variance is enormous. Identical prompts return the same list of brands less than 1% of the time on ChatGPT and Google's AI, across 2,961 runs (SparkToro, 2026). A single-query check is a data point. It isn't a trend. Run at least 20 queries across a month before drawing conclusions.

Nectiv's fan-out study found ChatGPT now runs an average of 7.61 fan-out searches per prompt on GPT-5.6 Sol, up from 2.17 a year earlier. That is more than triple (Nectiv, 2026). If that holds, the retrieval side becomes a fan-out problem, and the entity signals discussed here become even more load-bearing.

Sample bias, model version drift, snapshot-not-law. Treat this as a snapshot. Re-check quarterly.

FAQ

Does ChatGPT cite my domain? Almost certainly not, unless you're already a recognised entity in ChatGPT's training corpus. Superlines tracked its own brand across 34,234 AI responses over 30 days. ChatGPT gave it a 0.59% citation rate and named it in just 0.14% of answers (Superlines, 2026). Fastest check: run 20 buying-intent queries in your category through ChatGPT with browsing on, log which domains it names, and compare against a Bing WMT AI Performance export as a proxy signal.

Why does ChatGPT prefer Wikipedia? Because Wikipedia's editorial standard (verifiable citation, neutral tone, community moderation) matches ChatGPT's retrieval trust threshold better than any other web source. Wikipedia is 47.9% of ChatGPT's top-10 source share (Profound, 2025). No other single domain gets close. The corpus was trained heavily on it.

Can I influence ChatGPT's source pool? Indirectly. You can't push a page into ChatGPT's training corpus on demand. You can build a Wikipedia entity, earn third-party media citations, publish extractable content, and ship comprehensive schema. Over months, those signals compound into recognisable entity strength. Anyone selling faster than that is selling hope.

How is this different from Perplexity? Perplexity leans on Reddit as 46.7% of its top-10 source share (Profound, 2025). ChatGPT often cites nothing and leans on Wikipedia. Only 11% of domains are cited by both (Averi, 2026). Different engines, different playbooks. See ChatGPT SEO and Microsoft Copilot SEO for the engine-specific breakdowns.

Should I run my own study? Yes, at your scale. Pick 20-50 buyer queries in your category. Run each through ChatGPT once a month. Log which domains it names, which it links, and how the answer shape shifts. A three-month log will tell you more about your citation position than any generic industry report. Our AI visibility tracking guide covers the setup.

Do these findings apply to GPT-5.6 fan-out queries? Partly. Fan-out (internal query decomposition) multiplies the retrieval surface per user prompt. Nectiv's study found average fan-out searches per prompt rose from 2.17 to 7.61, more than a 3x increase (Nectiv, 2026). The entity and source-pool findings still hold. Volume scales with the fan-out. In practice the same small pool of sources gets consulted more often per prompt, which reinforces the Wikipedia and earned-media weighting rather than diluting it.

Where to go next

Six public studies plus one first-party cross-check point at the same answer. ChatGPT cites a small pool, weighted toward Wikipedia, earned media, and dense listicles, and it often cites nothing at all.

Build entity first. Then earn media. Then ship extractable content.

Read How to get cited by ChatGPT for the seven-move sequence. Read Generative Engine Optimisation for the multi-engine strategy layer. Read GEO vs SEO vs AEO vs AIO for the definitions.

If you run a UK brand and your Bing WMT AI Performance report shows fewer than 100 Copilot citations a month, book a free AI visibility audit. We'll show you where you sit in your category's ChatGPT source pool and the two moves that would change that.

References

• Gandhi, M. (2026, updated 19 June). Tracking the Volatility of ChatGPT's Citations: A Global Trend Analysis. seoClarity. https://www.seoclarity.net/chatgpt-citation-decline-analysis
• Superlines. (2026). AI Search Statistics 2026: 60+ Data Points on Visibility, Citations, and Traffic. 34,234 AI responses, 14 January to 13 February 2026. https://www.superlines.io/articles/ai-search-statistics/
• Profound. (2025). AI Platform Citation Patterns 2025-2026. https://www.tryprofound.com/blog/ai-platform-citation-patterns
• Profound. (2026). How ChatGPT sources the web. Brandon Punturo, 3 February 2026. https://www.tryprofound.com/blog/chatgpt-citation-sources
• Averi. (2026). AI Citation Tracking: ChatGPT, Perplexity, Claude. https://www.averi.ai/blog/ai-citation-tracking-chatgpt-perplexity-claude
• Brassell, N. (2026, 3 September). What AI Actually Cites for B2B Evaluation Stage Prompts. Ten Speed. Citation data via Peec AI, 7,387 citations across 170 prompts. https://www.tenspeed.io/blog/what-ai-cites-b2b-evaluation-stage
• Ray, L. (2026, 17 June). Why Calling Yourself the 'Best' Could Be Helping Your Competitors Win in AI Search. https://lilyraynyc.substack.com/p/why-calling-yourself-the-best-could
• Long, C. (2026, 13 August). What We Learned From Analyzing 28K+ ChatGPT And Gemini Fan-Out Queries. Nectiv. https://nectivdigital.com/blog/chatgpt-tripled-fan-out-queries-data-study
• Harsel, L. (2025, 10 November). The Most-Cited Domains in AI: A 3-Month Study. Semrush. 230K prompts, 100M+ citations. https://www.semrush.com/blog/most-cited-domains-ai/
• Indig, K. (2026, 16 February). The science of how AI pays attention. Growth Memo. 18,012 verified ChatGPT citations. https://www.growth-memo.com/p/the-science-of-how-ai-pays-attention
• Goodwin, D. (2026, 18 February). 44% of ChatGPT citations come from the first third of content: Study. Search Engine Land. https://searchengineland.com/chatgpt-citations-content-study-469483
• Baker, L. (2026, 7 July). SEO Study: 5 Lessons From Running AI Agents Across Every Search (Writesonic research). Search Engine Journal. https://www.searchenginejournal.com/seo-study-5-lessons-from-running-ai-agents-across-every-search/581663/
• Fishkin, R. (2026, 28 January). NEW Research: AIs are highly inconsistent when recommending brands or products. SparkToro. https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/
• Microsoft Bing Webmaster Tools. (2026, 29 July). AI Performance Overview Stats, gogochimp.com. First-party export.