ChatGPT accounted for 92.4% of the trackable standalone LLM referral traffic in Previsible’s dataset: 6.77 million LLM-driven sessions across 166 GA4 properties between November 2024 and May 2026. That is a meaningful result for GEO teams deciding where to look first. It is not, however, evidence that ChatGPT owns 92% of AI-influenced discovery, nor that every site should expect the same traffic mix. previsible.com
The practical mistake would be to turn a carefully bounded referral measurement into a universal market-share claim. The study measures a particular observable event: a user clicks a link in a standalone LLM product and arrives at a participating website. That makes the number useful, but it also means the sample, the attribution setup and the kinds of sites represented are part of the finding—not incidental footnotes. postforsuccess.com
What the 92.4% number actually measures
Previsible describes the figure as ChatGPT’s share of trackable LLM referral traffic. Its study includes standalone LLM platforms and excludes Google AI Overviews because those interactions operate under a different measurement model: their clicks are attributed to ordinary Google organic traffic rather than as identifiable standalone-LLM referrals. previsible.com
That boundary is decisive. A 92.4% share means that, within this dataset’s measured standalone-referral slice, ChatGPT generated nearly all observed referral sessions. It does not mean that nine out of ten people discovering brands through AI did so through ChatGPT. Previsible explicitly says AI discovery inside Google results likely represents a larger volume of AI-driven traffic than all standalone LLM platforms combined, while also saying that it cannot be measured in the same way. previsible.com
For GEO reporting, “AI traffic” is therefore too broad to be a decision-grade metric. The standalone chatbot referral channel is observable in this study; Google’s AI surfaces are not part of the same denominator. Combining the two conceptually and then presenting ChatGPT’s share as a share of all AI discovery would answer a question the dataset was not designed to answer. postforsuccess.com
A large dataset is not a census
The study is larger than the anecdotal screenshots that frequently drive AI-search discussion. It covers 166 GA4 properties, 6.77 million LLM-referred sessions and 19 months of observations. The participating sites span SaaS, ecommerce, finance, legal, health, insurance, education, publishing and ticketing. postforsuccess.com
That breadth gives the finding practical weight. The data is not limited to one publisher, one ecommerce catalogue or one vertical. It also captures a period in which total monthly LLM-referred sessions in the sample rose from 65,249 in November 2024 to 644,478 in May 2026, a 9.9x increase. postforsuccess.com
But the supplied methodology does not establish that these 166 properties are a random or representative sample of the web, of GA4 users, or of every audience using LLMs. It does not tell us how the properties were selected, how their referral traffic distribution compares with the wider web, or whether their measurement implementations were uniform. Those are not defects the report necessarily needs to solve; they are limits on what an external reader can infer from the reported aggregate.
That distinction matters because referral share is an outcome of both user behaviour and site opportunity. A dataset weighted toward sites with content, products, audiences or technical implementations that receive ChatGPT referrals may produce a different platform mix from a dataset weighted toward local businesses, consumer apps, closed portals or sites whose key journeys occur without a link click. The sources do not quantify those possible differences, so they should not be replaced with confident assumptions.
The tracking setup selects for clicks, not influence
The report counts visits where a person clicked from an AI product to a website. That is a valuable commercial event, but it is narrower than exposure, citation, recommendation, assisted conversion or a user who takes an answer from an LLM and later visits through another channel. The study is appropriately a referral-traffic study; it is not a complete measurement of every way an AI system can influence discovery. postforsuccess.com
This helps explain why the most important caveat is not a minor technicality. Google AI Overviews are absent from the chart precisely because their downstream visits are counted as Google organic traffic. Previsible’s conclusion is not that Google is irrelevant. It says the opposite: Google’s AI discovery may be larger than standalone LLM referrals, but the referral attribution available here cannot isolate it. previsible.com
GEO teams should resist treating what can be cleanly attributed as synonymous with what matters most. Attribution availability shapes dashboards. It does not settle the comparative influence of the underlying surfaces.
ChatGPT is still the correct first standalone referral priority
None of these caveats erase the operational conclusion inside the study’s actual frame. ChatGPT referrals increased from 47,606 sessions in November 2024 to 610,910 in May 2026, which Previsible describes as 12.8x growth. Its 92.4% aggregate share and continued growth make it the dominant standalone source of trackable LLM referrals in this sample. previsible.com
There is also no evidence here for an equal-effort, platform-neutral GEO programme if the immediate objective is measurable standalone referral acquisition. Gemini grew 3.2x over the tracked period, while Claude grew from 133 sessions in November 2024 to 8,528 in May 2026 and passed Perplexity in monthly referral sessions in March 2026. Perplexity declined 61% from its March 2025 peak, and Copilot declined 96% from its peak. previsible.com
The appropriate interpretation is priority, not exclusivity. Start analysis with ChatGPT because it is where the observed standalone referral volume is. Continue to inspect other platforms where their audience fit or trajectory warrants it. Do not claim that a low measured share proves a platform has little influence outside the referral events that the data can observe.
Landing pages expose a more useful GEO question
The report’s page-type findings may be more actionable than its market-share headline. Across the dataset, roughly a quarter of AI-referred traffic landed on internal search-results pages; for ChatGPT specifically, the reported figure was 28.8%. postforsuccess.com
That pattern changes the question from “How visible are we in AI?” to “Where does an AI-referred user land, and can that page complete the task?” Previsible’s SaaS analysis says search pages accounted for 34.6% of LLM referrals, while its ecommerce analysis identifies product pages as the primary landing surface. previsible.com
Those are vertical findings, not universal landing-page benchmarks. Yet they show why a site-wide AI-referral total can conceal the work that matters. A growing ChatGPT line is less valuable if the visitor is repeatedly delivered to a poor internal-search result. Conversely, a smaller number of referrals can be commercially significant if it consistently reaches an appropriate product, course, location or conversion page.
How to use the study without overstating it
I would treat the 92.4% figure as a well-supported allocation signal for one channel: standalone, trackable LLM referrals in a multi-industry GA4 sample. I would not use it as a forecast of your own platform mix or as a substitute for measuring the AI surfaces your analytics cannot distinguish.
- Keep standalone LLM referrals separate from broader organic search reporting, because Google AI Overview clicks are not isolated in this study’s model. previsible.com
- Break referral reporting down by source and landing-page type rather than relying on one blended “AI traffic” number.
- Prioritise ChatGPT when improving pages intended to acquire measurable standalone LLM referrals, because it is the dominant observed referrer in the dataset. previsible.com
- State the scope whenever presenting the 92.4% statistic: standalone LLM referral traffic, not all AI-influenced discovery.
The strongest reading of the research is neither “ChatGPT has won AI search” nor “the statistic is meaningless because it has boundaries.” ChatGPT has an extraordinary lead in the particular referral channel the study measures. For GEO practitioners, the discipline is to preserve that qualifier while using the result: optimise and measure where referral traffic is demonstrably arriving, but do not let a measurable slice become your entire model of AI discovery.