ChatGPT Lost 20 Points of Prompt Share and Still Sends 92% of AI Referrals

ChatGPT Lost 20 Points of Prompt Share and Still Sends 92% of AI Referrals

Comscore says ChatGPT lost 20 points of the prompt market in six months. The largest AI referral study ever run says ChatGPT sends 92% of AI clicks. Both numbers are right. They measure different stages of the same funnel, and the gap between them answers the budget question every GEO team is being asked this quarter.

Comscore’s Q2 2026 AI Intelligence Report, published today, carries the kind of headline that resets strategy decks: ChatGPT’s share of all AI prompts fell from 70% in January to 50% in June 2026, while Gemini nearly doubled from 17% to 30% and Claude climbed from 2% to 11%. Read it one way and it says the ChatGPT era of AI search is ending.

Eleven weeks earlier, Search Engine Land published the largest study of AI referral traffic to date: 6.77 million LLM-driven sessions across 166 websites, tracked over 19 months. Its finding points the other direction. ChatGPT sends 92.4% of all trackable AI referral traffic, up from roughly 84% six months before. The field is consolidating toward the leader.

Both measurements are credible, and neither one is wrong. They sit at different ends of the same funnel: Comscore counts prompts people type into assistants, while Previsible counts sessions where an assistant sent a visitor to a website. Between those two stages sit several decisions the assistants make on their own, and each assistant makes them very differently. That gap is the story. It tells you where AI-driven discovery happens, why referral dashboards systematically miss Gemini, and which scoreboard deserves your GEO budget.

Two studies, one assistant, 42 points apart

Comscore · Q2 2026 AI Intelligence Report
Counts prompts · opt-in panel of 500,000–1,000,000 US users · Jan–Jun 2026
70% → 50%
ChatGPT’s share of AI prompt volume lost 20 points in six months. Gemini 17% → 30%. Claude 2% → 11%.
Previsible · 3rd AI Traffic Study (via Search Engine Land)
Counts referral sessions · 6.77M sessions, 166 GA4 properties · Nov 2024–May 2026
92.4%
ChatGPT’s share of trackable LLM referral traffic, up from ~84% in December 2025. Growth of 12.8x in 19 months.
Same assistant, overlapping months, opposite directions. Comscore measures how people split their questions across assistants. Previsible measures how assistants split their outbound clicks. Sources: Comscore press release, Digiday, Search Engine Land.

The 42-point spread between a 50% prompt share and a 92.4% referral share is the distance between asking and clicking. And Comscore is not the only outfit on the referral side: an independent study by SE Ranking published in June found ChatGPT sending 74.8% of AI referral traffic in 2026. Two referral datasets, built on different samples and methods, disagree with each other by 18 points on the level and agree completely on the direction.

DatasetWhat it countsSampleWindowChatGPT headlineWatch-outs
Comscore AI Intelligence Report, Sep 22, 2026Share of AI prompts typed by consumersOpt-in panel, 500k–1M usersJan–Jun 2026Prompt share fell 70% → 50%Panel skews to measurable consumer behavior; excludes prompts inside Google Search surfaces
Previsible 3rd AI Traffic Study, Jul 6, 2026Referral sessions from standalone LLMs6.77M sessions, 166 GA4 propertiesNov 2024–May 202692.4% of trackable LLM referralsAuthor is Previsible’s CPO; excludes AI Overviews; agency-managed property mix
SE Ranking AI traffic study, Jun 18, 2026AI referral share of all website trafficGlobal referral panel2024–202674.78% of AI referrals; absolute volume +27% YoYAI is still 0.32% of all web traffic in this panel
Demandbase Labs, Aug 12, 2026Referral visits to B2B web properties11B visits, 1,584 tenantsJun 2025–Jun 2026B2B referrals +303% (645k → 2.6M monthly)One vendor’s estate; Gemini and Claude referrals flat here
Semrush clickstream study, Apr 7, 2026US clickstream, prompt-level visibility1B+ lines, 200M-user panelOct 2024–Feb 2026Outbound referrals +206% YoY; search enabled on 34.5% of queriesUS only; sees prompts, not citations
Five datasets, five scoreboards. Each counts a different stage of the AI discovery funnel, which is why the same company can look like it is collapsing and consolidating at the same time. Links in the sources list below.

ChatGPT’s share fell because the pie grew

A share falling from 70% to 50% sounds like decline, but Comscore’s own report says the opposite happened to ChatGPT’s absolute numbers. It reached 168 million desktop conversations in June 2026, and its multiplatform visitation grew from 59 million to 99 million year over year. Overall assistant reach climbed to 35% of desktop users and 29% of mobile users, up from 25% and 12% in January 2025.

The growth came from assistants reaching people where they already are. Gemini’s growth is largely a distribution story: it ships inside Google Search, Android, and Workspace, so it collects people who never went looking for a chatbot. Claude’s growth is a workplace story, concentrated among knowledge workers and technical users. Michael King, founder of iPullRank, told Digiday it resembles the early social era, “where audiences spread across platforms and each one developed its own behavior and its own user base.”

Jan 2026
70% 17% 11%
Jun 2026
50% 30% 11% 9%
ChatGPT Gemini Claude Other assistants (incl. Perplexity)
Share of AI prompt volume, January vs June 2026. ChatGPT’s 20-point loss maps almost exactly onto Gemini’s +13 and Claude’s +9. Total prompt volume grew over the same period, so the loss is dilution. Source: Comscore Q2 2026 AI Intelligence Report.

Semrush’s clickstream data adds the demand-side detail: ChatGPT’s total traffic plateaued near 1 billion monthly visits around November 2025, but users who stayed are working it harder, with average queries per session jumping 50% in the last four months of that study. Fewer casual tourists, more research sessions. That mix matters, because research sessions are the ones that end in clicks.

Prompts do not become clicks at the same rate

Neither study states this on its own: each assistant converts its prompt share into tracked referral traffic at a completely different rate.

ChatGPT
50%
74.8%
×1.5 over-yields prompts into clicks
Gemini
11.6%
30%
×0.4 under-yields by ~2.6x
Claude
2.6%
11%
×0.2 under-yields by ~4x
Prompt share (Comscore, Jun 2026) Referral share (colored by platform, SE Ranking 2026)
Referral yield: share of tracked AI clicks divided by share of prompts. Ratios are directional, because the two studies use different samples, geographies, and windows. Under Previsible’s stricter lens (92.4% of referrals), ChatGPT’s yield rises to roughly ×1.8. Perplexity is excluded because Comscore does not report it separately; at 7.2% of referrals from a single-digit prompt share, it also over-yields. Sources: Comscore, SE Ranking, Previsible.

ChatGPT holds half of prompts and sends three-quarters of tracked AI clicks. Gemini holds nearly a third of prompts and sends about one in nine tracked clicks. Claude holds 11% of prompts and sends 2.6% of tracked clicks. If a client asked you this morning whether to shift GEO budget toward Gemini because of the Comscore headline, this is the chart that answers them: prompt share measures where people ask, and these assistants are not equal machines for turning questions into visits.

Where Gemini’s prompts go

The under-yield follows from where each assistant lives and what its product does with a question.

A ChatGPT prompt

User asks a question. ChatGPT enables web search for 34.5% of queries (down from 46% in late 2024), answering the rest from training data.
Searched answers carry source links the user can click.
The click arrives with referrer chatgpt.com.
Visible in every analytics tool as an AI referral ✓

A Gemini prompt

User asks a question inside Google Search, Android, or Workspace. Gemini’s desktop conversations grew from 8M to 89M in a year, mostly via bundling.
The answer often resolves inside Google’s surfaces. AI Overviews now appear in 39.4% of Google desktop searches, up from 25.8% in July 2025.
when a click does happen ↓
It typically stays inside Google’s ecosystem and attributes as google / organic in analytics.
Invisible to referral dashboards as “AI traffic” ✗
Why one of these prompts counts as AI traffic and the other does not. The split between these paths is not quantified in any public dataset; referrer-based analytics structurally undercount Gemini and over-represent assistants that live outside Google. Sources: Semrush clickstream study, Comscore.

Comscore’s own report contains a version of this same conflation one stage earlier in the funnel. Tripadvisor accounts for 61% of travel sourcing by AI platforms but only 21% of citations. “Being used by AI is not equal to being visible in AI,” Comscore’s chief analytics and technology officer Smriti Sharma said. King describes the mechanics: an assistant first fans out into sub-queries and pulls a wide set of pages, then synthesizes and attaches citations only to the passages it leaned on. “For measurement, that means citations alone are a lagging, incomplete metric.”

Add the stage before sourcing and you get the full rule: there is no such thing as “AI market share.” There are at least four separate shares, one per stage of the funnel, and they are allowed to disagree: prompt share, retrieval share, citation share, and referral share. Quoting one of them as if it stood for the others overstates what it shows.

The referral wave grew 10x and stayed concentrated in ChatGPT

While prompt share fragmented, the referral side consolidated and grew. Previsible’s panel went from 65,249 monthly LLM-referred sessions in November 2024 to 644,478 in May 2026, a 9.9x increase. Semrush measured ChatGPT’s outbound referral traffic up 206% in 2025. Demandbase measured ChatGPT referrals to B2B properties quadrupling from roughly 645,000 to 2.6 million monthly visits between June 2025 and June 2026, while Perplexity referrals fell and Gemini and Claude stayed flat on its estate.

600k 400k 200k 0 the crash zone 65k 396k 443k 428k 644k Nov 24 Aug 25 Dec 25 Feb 26 May 26 Nov 2025: ChatGPT referrals 448,412 → 213,345 in one month
Monthly LLM-referred sessions across Previsible’s 166-property panel. Growth of 9.9x in 19 months, interrupted by a one-month crash in November 2025 when ChatGPT began favoring Wikipedia and Reddit; total sessions recovered to 442,609 by December. Dashed segment marks months with no published total. Source: Previsible via Search Engine Land.

The crash in that chart deserves as much attention as the growth line. In November 2025, ChatGPT referrals to the panel fell from 448,412 to 213,345 in a single month while every other platform held steady, apparently because the model changed which sources it preferred. Sessions recovered by December. David Bell, Previsible’s CPO and the study’s author, draws the operational conclusion: “One vendor’s product decisions can halve your AI traffic overnight.”

The same study also found where a surprising share of those clicks land. Roughly 25% to 28.8% of ChatGPT referral traffic arrives on sites’ own internal search results pages, and in SaaS that figure reaches 34.6%, making internal search the single largest landing surface for AI referrals: the model trusts the domain but cannot pick the exact page, and sends the user to the search box to finish the lookup. That is a fixable problem and one of the cheapest GEO wins available right now.

AI platforms account for about 0.32% of all website traffic in SE Ranking’s global panel, up from 0.02% in 2024. Demandbase’s ChatGPT referrals work out to roughly a third of one percent of the traffic it measures. The growth is real and compounding, and it is still small in absolute terms.

How to spend a GEO budget when the scoreboards disagree

Prompt-share news alone is a weak reason to rebalance away from ChatGPT. For any GEO goal that involves citations, recommendations, or referrals, ChatGPT remains the dominant channel by a factor of three to twelve over its nearest rival, depending on which referral dataset you trust. Bell’s phrasing is blunt: “Optimizing for ‘AI visibility’ without prioritizing ChatGPT means optimizing for an abstraction.”

Read prompt share to find where your audience is asking; let referral and visibility data set the budget. The Comscore shift tells you where different people are asking questions: Gemini collects mainstream, mobile-heavy, bundled users; Claude collects knowledge workers and technical buyers. King’s advice is to measure visibility per assistant and weight it by where your readers are, because “the fundamentals that get you cited are largely the same across all of them.”

Claude is a cheap early window for technical audiences, priced correctly. Claude’s referrals grew 64x in Previsible’s panel and overtook Perplexity in March 2026, and SE Ranking puts its growth at 320% year over year. But Demandbase found Claude’s B2B referrals flat, and its absolute volume is still small. Expect Claude to earn citations and sourcing influence among professional buyers before it earns sessions.

Fix the internal-search leak. If a quarter of your ChatGPT referrals land on a search results page, the model is asking your site to do the last mile of navigation. Give it deep-linkable, self-contained pages with clean URLs and explicit answers, and make sure your site architecture exposes one canonical page per question worth owning.

Price in the concentration risk. When one vendor sends ~92% of your AI referrals (or ~75%, per SE Ranking), your AI channel is a single point of failure that can halve in a month, as it did in November 2025. Multi-assistant presence is the hedge, with the honest expectation that the hedge buys visibility rather than clicks.

Track the surface no referral dashboard shows. AI Overviews now appear in 39.4% of Google desktop searches, up from 25.8% a year earlier, and Bing’s Copilot presence rose from 11.8% to 17.3%. The largest AI answer layer in the world is invisible to the metric most GEO dashboards report, so pair referral tracking with visibility measurement inside Google’s own AI surfaces.

Where the data is shaky

  • The studies use different measurement windows: Comscore’s prompt shares cover January to June 2026; Previsible’s 92.4% spans November 2024 through May 2026; SE Ranking reports full-year aggregates. The dumbbell above mixes them deliberately and labels the result directional.
  • The referral studies disagree with each other: 92.4% versus 74.8% is an 18-point spread driven by sample composition. Previsible uses 166 GA4 properties (heavily agency-managed sites), SE Ranking a global panel. SE Ranking also shows ChatGPT’s own referral share sliding from 79.7% to 74.8% in a year, so the leader’s grip on clicks is strong but not tightening forever.
  • Conflict-of-interest flags: Previsible’s study was written by its CPO and published through a contributor channel on Search Engine Land, which Semrush owns. Demandbase measured its own customers’ estates. Comscore’s panel is opt-in.
  • The referrer undercount runs in both directions. If Gemini’s true click influence is larger than its referrer share suggests, then the ×0.4 yield understates Gemini’s real value to brands, and the case for measuring visibility inside Google surfaces gets stronger.

The Comscore report and the Previsible study point two instruments at two different stages of the same funnel. ChatGPT is losing the prompt race at exactly the same time it is winning the click race, because the assistants gaining prompts were never click machines to begin with. Pick the scoreboard that measures the stage you are optimizing for, and stop letting a prompt-share headline answer a referral question.