ChatGPT sends 92% of AI search traffic - what 6.77 million sessions reveal

Everyone talks about the AI-search race as if it is close. New traffic data says that, among standalone assistants, it is not. Previsible analysed 6.77 million sessions and found ChatGPT sends the overwhelming majority of trackable AI referral traffic. But the same report carries a caveat that changes the whole picture.

When a market looks like a race, you hedge across the runners. When one runner is this far ahead, you prioritise. But the details of this data matter enormously, so let us be precise about what it does and does not say.

Our founder, Somya Goyal, shared her quick take on this data on LinkedIn. This post is the longer companion: what the numbers say, the caveat most coverage skips, and what to actually do about it.

The headline number

Previsible's 2026 State of AI Discovery report, published on 6 July 2026, analysed 6.77 million sessions across 166 GA4 properties from November 2024 to May 2026. Among standalone large language models, ChatGPT accounts for about 92.4% of trackable referral traffic. The rest of the field splits the remaining sliver.

The movement underneath is almost as striking as the top line. Claude grew roughly 64x over the period and overtook Perplexity around March 2026, while Perplexity fell sharply from its peak. So the non-ChatGPT field is real and moving - it is just small.

"Among standalone assistants the race is not close. But the biggest AI surface is not even in the 92%."

Curious how AI engines describe your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

The caveat that changes everything

Here is the part that a headline number hides, and Previsible states it plainly: the report excludes Google's AI Overviews, and in-Google AI discovery is likely larger than all standalone LLMs combined. So the 92% is 92% of a slice - the standalone-assistant slice - not 92% of all AI-driven discovery.

That reframes the strategic lesson. If you read "ChatGPT is 92%" and pour everything into ChatGPT, you may be optimising hard for the smaller of two arenas while ignoring the larger one that does not even appear in the chart. The honest picture is: ChatGPT dominates standalone referrals, and Google's AI surfaces quietly dominate the total.

What to actually do

1. Prioritise ChatGPT among standalone assistants. If you are going to optimise for one chat assistant's referral traffic, the data says start here.
2. Do not skip Google's AI surfaces. AI Overviews and AI Mode are where the largest share of AI discovery happens, even though they are absent from standalone-LLM referral stats. Being cited there is its own priority.
3. Assume the mix will shift. Claude's 64x rise and Perplexity's fall in a single period show how fast this moves. A strategy pinned to today's shares ages badly.
4. Measure your own mix, not the market's. The 92% is an average across 166 sites. Your buyers may skew differently. Track where your citations and referrals actually come from.

The takeaway

Two facts, held together: among standalone assistants, ChatGPT is not just winning but lapping the field - and yet the surface that matters most for total AI discovery, Google's AI answers, is not in that number at all. The takeaway is not "go all-in on ChatGPT." It is "prioritise ChatGPT, respect Google's AI surfaces, and measure your own reality rather than the headline."

Why one engine pulled so far ahead

A 92% share does not happen because one model is 92% better. It happens because of distribution. ChatGPT reached a mass audience first, became the default verb for "ask the AI," and then wired itself into the places people already work: a phone app, a browser extension, a search box that answers rather than lists. When the front door is that wide, referral traffic follows, almost regardless of how the underlying answers compare.

This matters for how you read the rest of the field. Claude growing 64x and overtaking Perplexity is not a story about answer quality either. It is a story about surfaces - more integrations, more default placements, more moments where a user is one tap from an answer. The lesson is that referral share tracks reach and habit far more than it tracks raw capability, and reach is exactly the thing that can shift when a new default appears.

So the strategic question is not "which model is best." It is "where do my buyers already ask, and am I being named when they do." Those are different questions, and only the second one is inside your control.

Getting named, not just linked

Referral traffic is the visible tip of AI discovery. Underneath it sits a larger, quieter behaviour: people read the answer, absorb the recommendation, and never click through at all. If the assistant names three vendors and describes one as the obvious choice for your use case, that brand has won the moment even if no session shows up in anyone's analytics. Optimising only for the click misses most of the value.

That changes what "optimisation" means here. You are not chasing a blue link on a results page. You are trying to be the name the model reaches for when it composes an answer, and the source it cites when it wants to look credible. In practice that rewards a few concrete things:

• Clear, quotable claims. Models lift sentences that state a fact plainly. Bury your differentiator in a paragraph of adjectives and it will be paraphrased away or skipped.
• Structured, checkable content. Comparison tables, defined terms, and specific numbers give a model something it can cite with confidence rather than hedge around.
• Corroboration off your own site. A claim repeated on third-party pages, reviews, and industry write-ups reads as consensus, and consensus is what an assistant is trying to summarise.
• Freshness and specificity. "Best for teams under 50 handling regulated data" gets named on the exact query it answers. "Best for everyone" gets named on nothing.

None of this is a trick. It is the same instinct that made a page rank in classic search, pointed at a reader who now summarises instead of linking.

What this data does not mean

It is easy to over-learn from a clean headline, so a few guardrails. A 92% share does not mean the other engines are safe to ignore. Share is a snapshot, and this report captures a period in which one runner fell 61% from its peak while another grew 64x. A snapshot is a starting point for attention, not a licence to bet the strategy on today's ranking.

It also does not mean AI referral traffic is the whole game. The report's own caveat about Google's AI surfaces is the reminder that the largest arena is not even counted here, and that the mention itself often matters more than the click. Treating the 92% as a map of all AI discovery is exactly the mistake the data warns against.

"Being named where your buyers already ask beats winning a race that is measured somewhere else."

And it does not mean the market average is your average. These are 166 sites blended together. A B2B security vendor and a consumer recipe site do not share an audience, and they will not share an engine mix. The only number that should drive your roadmap is your own, measured per engine, on the queries your buyers actually type.