ChatGPT Search & SEO Statistics 2026: How It Picks and Cites Pages

ChatGPT is no longer just a chatbot that marketers worry about in the abstract. It is a measurable search and discovery channel, with regulator-grade usage disclosures in the EU, clickstream studies tracking billions of prompts, and reverse-engineering work that shows exactly how it picks the pages it cites.

That matters because the mechanics are genuinely different from Google. ChatGPT decides on a prompt-by-prompt basis whether to search the live web at all, fans a single question out into many synthetic queries, and then cites only a sliver of what it retrieves.

This article pulls together the verified numbers on ChatGPT as a search and marketing channel in 2026: who uses its search function, how it selects sources, which pages win, and what marketers and businesses are actually doing with it. For the broader user-base figures and cross-platform citation league tables, see our separate ChatGPT Statistics and AI Citation Statistics articles.

Key ChatGPT search and SEO statistics for 2026

• ChatGPT search averaged 159.1 million monthly active recipients in the EU over the six months to 31 March 2026, according to OpenAI's DSA disclosure, nearly four times the 41.3 million a year earlier.
• The European Commission has confirmed OpenAI's numbers sit above the 45 million threshold for designation as a very large platform or search engine under the DSA, per PYMNTS reporting.
• ChatGPT triggered a live web search on 34.5% of prompts in February 2026, down from 46% in late 2024, in Semrush clickstream data.
• Between 65% and 85% of ChatGPT prompts could not be matched to any traditional search keyword across Semrush's 17-month study window.
• ChatGPT left 85% of the pages it retrieved uncited in an AirOps study of 15,000 queries and 548,534 retrieved pages.
• ChatGPT generated two or more fan-out queries on 89.6% of searches, and 95% of those fan-out queries had zero monthly search volume, per AirOps.
• Pages ranking number one in Google were cited 3.5 times more often than pages outside the top 20, and 55.8% of all cited pages ranked in Google's top 20 for at least one query, per AirOps.
• The average number of unique domains cited per ChatGPT response fell from 19 to 15, a drop of more than 20%, after OpenAI's 4 March 2026 default model switch, according to research published on Search Engine Land.
• Over 20% of ChatGPT's outbound referral clicks go to Google, and the top 10 destination domains take just over 30% of all referrals, per Semrush.
• Practical guidance (29%), seeking information (24%) and writing (24%) together account for nearly 80% of ChatGPT messages, in OpenAI's own usage research.
• More than 1 million businesses were paying OpenAI customers by November 2025, with over 7 million ChatGPT for Work seats, per OpenAI.
• 80% of marketers use AI for content creation, and 61% say marketing is experiencing its biggest disruption in 20 years, in HubSpot's 2026 State of Marketing report.

How many people use ChatGPT search?

The most reliable numbers come from a regulator, not a press release. Under Article 24(2) of the EU Digital Services Act, OpenAI must publish average monthly active recipients for ChatGPT search in the EU, and the trajectory is steep [1].

The latest disclosure reports approximately 159.1 million average monthly active recipients in the EU for the six-month period ending 31 March 2026 [1]. A year earlier, Search Engine Land reported the equivalent figure at 41.3 million, itself nearly four times the 11.2 million of the preceding half-year [2].

That is roughly a fourteenfold increase in about 18 months, in one region, for the search function alone. PYMNTS reports the European Commission has confirmed these figures sit well above the 45 million threshold that can trigger designation as a very large online platform or search engine, though it stresses any classification is assessed case by case [3].

Two caveats. These are EU-only figures calculated specifically for DSA compliance, so they are not directly comparable with weekly-active-user numbers for ChatGPT overall. And for scale, the 2025 disclosure round put Google's EU search recipients at roughly 20 times ChatGPT search's then-total [2]. ChatGPT search is growing extremely fast from a base that is still small next to Google.

What share of ChatGPT prompts are search-like?

Most ChatGPT usage still does not look like search. Across Semrush's analysis of more than 1 billion lines of US clickstream data from October 2024 to February 2026, between 65% and 85% of prompts could not be matched to any traditional search keyword in its database [5].

OpenAI's own research points the same way. The OpenAI and NBER working paper "How People Use ChatGPT", based on around one million de-identified consumer messages, found practical guidance accounts for about 29% of usage, seeking information about 24%, and writing about 24%, together nearly 80% of all messages [4].

The same paper classifies about 49% of messages as "Asking", 40% as "Doing" and 11% as "Expressing" [4]. "Asking" is the closest analogue to search intent, and combined with the 24% seeking-information share it suggests a meaningful minority, not a majority, of ChatGPT activity competes directly with a search box.

The direction of travel matters more than the snapshot. Semrush found the share of prompts using traditional search language nearly doubled between October 2025 and February 2026 [5]. People are learning to treat ChatGPT as a search engine, even though most of what happens inside it is still drafting, advice and problem-solving.

How often does ChatGPT actually search the live web?

Less often than most marketers assume, and less often than it used to. Semrush's clickstream data shows ChatGPT enabled web search on 34.5% of queries as of February 2026, down from 46% in late 2024 [5].

That single number reframes ChatGPT optimisation. On roughly two-thirds of prompts, the model answers from pre-trained knowledge and your live page never enters the equation. Your content only competes when retrieval actually fires, which happens mainly on fresh, factual, product or news-flavoured questions.

Search Engine Land's coverage of the study notes the decline in search-triggered queries directly limits how often sites can earn citations and traffic [6]. Research published on Search Engine Land by Olivier de Segonzac adds a structural reason: more than 90% of ChatGPT's weekly users are on the free plan, and the default free experience triggers fewer web searches, runs fewer queries and produces fewer citations [8].

There is also a behavioural difference when search is on. In Semrush's earlier 80-million-record study of July to December 2024 data, prompts without search averaged 23 words, whilst search-enabled queries averaged just 4.2 words [13]. People instinctively write short, keyword-ish queries when they want ChatGPT to look something up.

How does ChatGPT choose which sources to cite?

Through query fan-out. When ChatGPT searches, it rarely runs your user's question verbatim. It rewrites the prompt into multiple synthetic queries, retrieves results for each, then synthesises an answer from a shortlist.

The scale of this is now measured. AirOps ran 15,000 original queries that expanded to 43,233 total queries once fan-outs were counted, and found ChatGPT generated two or more fan-out queries on 89.6% of searches [7]. Strikingly, 95% of those fan-out queries had zero monthly search volume in traditional keyword tools [7].

The de Segonzac study reverse-engineered the internal tool behind this, called web.run, and found the behaviour varies sharply by model. GPT-5.4 Thinking distributes often more than 10 fan-out queries per response, uses site: operators to restrict searches to trusted domains, and can chain 5 to more than 10 rounds of search, whilst the default GPT-5.3 Instant typically runs 2 or 3 [8].

Model choice also changed the citation surface. After OpenAI switched its default model on 4 March 2026, the average number of unique domains cited per response fell from 19 to 15, and unique URLs from 24 to 19, across 400 daily tracked prompts over 14 weeks [8]. Same pie, fewer slices: fewer sites now share each answer.

One more finding worth knowing. OpenAI describes OAI-SearchBot as the crawler that builds ChatGPT's search index, but the study found the model leans on third-party search APIs, with Google's fingerprints visible in the backend, then sends its ChatGPT-User agent to fetch the selected pages [8]. Ranking in Google still feeds ChatGPT, indirectly.

What kinds of pages win visibility in ChatGPT?

Pages that already rank, from mid-authority domains, matching very specific sub-questions. AirOps found pages ranking number one in Google were cited 3.5 times more often than pages outside the top 20, with a 43.2% citation rate for top-ranked pages, and 55.8% of all cited pages ranked in Google's top 20 for at least one query [7].

Retrieval is not citation, though. Across 548,534 retrieved pages, ChatGPT left 85% uncited [7]. Being fetched gets you into the audition; only 15% of retrieved pages make the final answer.

Two numbers in that table should change how content teams plan. First, 32.9% of cited pages appeared only in results for a fan-out query, never for the original prompt [7]. A third of ChatGPT visibility comes from covering the narrow sub-questions no keyword tool will show you.

Second, domains with authority scores between 20 and 80 took 63.6% of citations, whilst the biggest sites (80 to 100) were retrieved often but cited at a lower 15% rate [7]. ChatGPT's synthesis step appears to reward specific, well-matched pages over raw domain size, which is genuinely better news for mid-sized publishers than the equivalent Google data.

Where do ChatGPT's outbound clicks go?

Mostly to a small club of big destinations, and remarkably often back to Google. Semrush found more than 20% of ChatGPT's referral traffic goes to Google itself, with the top 10 domains taking just over 30% of all outbound clicks [5][6].

Users treating ChatGPT and Google as complements, not substitutes, is the honest reading. People take an AI answer and go verify it, or complete the task, on the open web via Google.

Distribution beyond the top is broadening but volatile. The number of domains receiving any ChatGPT referrals peaked around 260,000 during 2025 before settling near 170,000 [6]. For most sites, ChatGPT visibility still translates into marginal absolute traffic, which is why conversion quality, covered in our AI referral traffic article, carries the business case.

How many marketers use ChatGPT and AI for SEO work?

Adoption is close to universal at the "any AI" level, and majority-level for content and optimisation tasks. HubSpot's 2026 State of Marketing report finds 80% of marketers use AI for content creation and 75% for media production, whilst 61% believe marketing is experiencing its biggest disruption in 20 years [9].

HubSpot's earlier AI Trends for Marketers survey of more than 1,000 professionals, updated June 2025, put overall AI use at 66% of marketers, with 48% using generative AI for research tasks and 47% for blog posts [10]. SurveyMonkey research puts day-to-day AI use higher at 88% of marketers, with 51% using AI tools to optimise content, including SEO keyword work [11].

The spread between 66% and 88% across credible surveys is itself informative. Definitions differ, samples differ, and "using AI" now ranges from a monthly ChatGPT dabble to fully AI-assisted production pipelines. Treat any single adoption figure as a band, not a point.

What has clearly changed by 2026 is the direction of the work. The question inside SEO teams has shifted from "should we use ChatGPT to produce content?" to "how do we get ChatGPT to retrieve and cite our content?", which is why fan-out data now features in content briefs.

How are businesses adopting ChatGPT?

Fast enough that OpenAI now reports enterprise numbers alongside consumer ones. OpenAI announced in November 2025 that more than 1 million businesses worldwide were direct customers, with over 7 million ChatGPT for Work seats, up 40% in two months, and Enterprise seats growing ninefold year on year [12].

OpenAI credits the consumer base, then more than 800 million weekly users, as the wedge driving workplace adoption [12]. Employees bring the tool to work before procurement catches up, the classic consumerisation pattern.

For marketing specifically, the business story and the search story are converging. Semrush's referral data shows online services, software and development, mass media and education receiving the most ChatGPT traffic [5], and Search Engine Land reported in July 2026 that OpenAI appears to be building chatbot-native ad formats [8]. A channel with regulator-verified nine-figure search usage and an incoming ad product is a media channel, whatever else it is.

How do you optimise for ChatGPT's query fan-out?

Cover the sub-questions, not just the head term. With 95% of fan-out queries showing zero search volume [7], keyword tools cannot enumerate them, but they are predictable: definitions, comparisons, pricing, specs, suitability, alternatives and evidence for the parent topic.

The data supports a few concrete moves. Keep earning top-20 Google rankings, since 55.8% of cited pages rank there and the backend still draws on Google-derived results [7][8]. Structure pages so a single passage cleanly answers a single narrow question, because citation happens at the level of the fan-out match, and a third of citations come only from fan-out SERPs [7].

For product and ecommerce queries, the mechanics are unusually literal. De Segonzac's work found ChatGPT now runs a separate shopping fan-out for each candidate product, fetching specs, reviews and pricing one by one [8]. If your product pages do not expose those elements cleanly to ChatGPT-User, you are invisible at the exact step where recommendations are assembled.

And accept the funnel maths. Retrieval fires on about a third of prompts [5], 85% of retrieved pages are discarded [7], and the citation surface shrank by a fifth in March 2026 [8]. ChatGPT visibility is winnable, but it is a narrow gate, which is precisely why the pages that do get through are seeing outsized value.

How to read these numbers

The EU DSA figures are the closest thing this market has to audited data, but they are averages calculated for a specific legal purpose, EU-only, and cover the search function rather than all ChatGPT use. Do not divide them into global totals.

Clickstream studies such as Semrush's are panel-based. A 200-million-user panel is large, but it skews towards consumers, towards the US in this case, and infers intent from text matching, so "search-like" is a modelled label rather than an observed one.

The retrieval studies (AirOps, de Segonzac) are experiments on researcher-chosen prompt sets. They reveal mechanics with unusual clarity, but citation rates would shift with a different prompt mix, and OpenAI can change the behaviour with a model swap, as the March 2026 drop from 19 to 15 cited domains demonstrated.

Survey figures on marketer adoption are self-reported and definition-sensitive, which is why HubSpot and SurveyMonkey land 22 points apart on essentially the same question. When credible sources disagree, the honest answer is a range.

What this means going into 2027

First, ChatGPT search is heading into formal search-engine regulation in Europe. A VLOSE-style designation would bring audits and transparency reports, and probably better public data on how sources are selected, which marketers should actively watch for.

Second, optimisation effort should follow the retrieval trigger. Content targeting fresh, factual, comparative and product questions gets a shot at citation; evergreen conversational topics mostly play out inside the model's training data, where brand mentions across the wider web matter more than any single page.

Third, plan for a shrinking citation surface. Fewer domains per answer and deeper trusted-domain bias mean AI visibility will concentrate the way early Google did. Mid-authority sites still win 63.6% of citations today; that window is worth exploiting before it narrows.

Fourth, build pages for fan-out, not just keywords. Passage-level answers to zero-volume sub-questions are now a measurable third of ChatGPT citations, and no rank tracker will show you the queries. Topic mapping beats keyword lists here.

Finally, treat ChatGPT as a channel with its own P&L logic. Regulator-verified usage in the hundreds of millions, over a million business customers, and ad formats in development: the sensible 2027 posture is to measure it properly now, whilst the traffic is still small enough that competitors ignore it.

If you want a plan for earning citations inside ChatGPT for your own pages, you can book a call with me.