What 17 months of clickstream data say about ChatGPT traffic

The aggregate growth of AI traffic hides a fight that got harder for anyone outside the top. Concentration is what 17 months of numbers show.

Disclosure: I am a Semrush ambassador, and the study cited in this article was published by Semrush. The data is labeled with source and access date, and my reading of it is my own.

Studies on AI traffic usually travel as headlines, with one big number and no reading of the trend. Semrush’s clickstream study lets you do the opposite, because it covers a period long enough to show inflection points instead of snapshots.

Why concentration is a different reading from growth

The difference between the two readings has a direct consequence for content planning, and it is worth being explicit about the mechanism.

On a growth curve, entering later costs less, because the surface is still expanding and new space is created every month. That is how traditional search behaved for a decade, and it is the intuition most teams still carry around.

On a concentration curve, entering later costs more, because the space stopped growing and a new domain coming in implies another one going out. The fight becomes one against a specific incumbent that is already in the set the system consults. When that incumbent loses ground, the competitor traffic drop diagnosis helps you read where the demand went.

The study’s numbers describe the second case. The volume plateau since November 2025 rules out the expansion hypothesis, and the drop in domains shows that selection got narrower over the same period.

The behavior of the people asking also changed

The study brings a second set of numbers that rarely makes the headlines and that changes what gets written, on top of how much.

According to Semrush, search was triggered in 34.5% of queries in February 2026, against 46% at the end of 2024. In the same study, search-enabled prompts went from 4.7 words in early 2025 to 8.7 words in early 2026, while prompts without search dropped from 24.9 to 13.5 words.

The study also records 1.75 queries per session in February 2026, up 50% over the last four months of the period. That points to chained conversation, with refinement, instead of a single question. For whoever is measuring, it means the session now has more than one possible entry point, and giving the conversion to the last touch loses part of the path.

Where the traffic goes, which contradicts the replacement thesis

The most uncomfortable data point for both current narratives is the destination. Again according to Semrush, Google receives 21.6% of all referrals from ChatGPT, and the ten largest domains hold a little over 30% of referral traffic.

What I would do with these numbers in planning

• Stop projecting linear growth in AI traffic. The plateau observed since November 2025 argues against that projection, and a plan that depends on it will miss its target by design.
• Prioritize getting into the band of domains that still receives referrals. In a concentration scenario, effort pays off more by displacing an incumbent in narrow territory than by covering broad territory with average pieces.
• Rewrite the base with longer, more specific prompts in mind, since that is the format that grew, and let the short keyword in the spreadsheet serve only as a starting point.
• Measure the chained journey, because a session with multiple queries changes what counts as an assisted conversion.

The limits of this study, stated

Three caveats that need to come with any use of these numbers.

The coverage is the United States, with a panel of 200 million users. There is no basis for extending the reading to other countries or languages without local verification, and regional difference is a legitimate open question.

The measurement is clickstream, which captures the click leaving the system. It cannot see the answer that settled the question without any click at all, which is exactly the behavior that grew most in the period. That behavior only shows up in citation measurement, like the one described in how to track AI visibility.

The study comes from a vendor that sells an AI visibility tool. That does not invalidate the data, which comes with declared methodology and sample, and it recommends treating the market reading that comes with it by the same standard you apply to any manufacturer’s material.