The short version
- 34.1% of AI assistant sessions had no observed web step in a February 2026 panel of US and British users (Iannelli and Ai, 2026).
- For the same people’s search sessions the figure was 19.5%, a gap of 13.0 percentage points that held at 13.6 in March 2026.
- The share depends heavily on where one session ends and the next begins: it ran from 46.9% to 29.2% across the researchers’ three choices.
- It also varied by market and by assistant: 50.9% in the United States against 20.7% in Great Britain, and 48.8% for Gemini against 27.3% for ChatGPT.
- No visit does not mean no web: in our study, every assistant answer that ran a search cited at least one page.
How many AI assistant sessions involve no website visit?
About a third, in the one large study that has measured it.
Iannelli and Ai (opens in a new tab) linked the prompts and responses of an opt-in panel to the same people’s searches and page visits in February 2026. The assistants were standalone ones, mainly ChatGPT and Gemini with a small share of Perplexity. A session ended after 30 minutes with no activity.
Counting each person equally, 34.1% of sessions with an assistant had no observed outside web step: no search and no page visit. Counting every session equally, the share was 37.8%.
The authors work at Scrunch AI, a commercial company, so this is vendor research. The paper was under review and not yet peer reviewed.
| Where outside web activity fell | Share of assistant sessions |
|---|---|
| Nowhere: none observed | 34.1% |
| Only before the assistant | 18.3% |
| Only after the assistant | 10.5% |
| On both sides, or between prompts | 37.1% |
Only about one session in ten matched the familiar picture of asking the AI first and then clicking out.
Is that more than with ordinary Google search?
Yes: only 19.5% of the same people’s search sessions had no visit to a page beyond the search engine. The researchers built search sessions the same way and checked whether any non-search page was visited. Comparing each person’s two kinds of sessions, assistant sessions stayed self-contained 13.0 percentage points more often. We compare clicks per question in our article on answer engines and website clicks.
The finding repeated a month later. In March 2026, 35.3% of assistant sessions and 20.4% of search sessions had no visit, a gap of 13.6 points.
Some of those searches may have been answered by Google’s own AI summary on the results page. The authors expect that to shrink the gap, not widen it.
Google is not immune to sessions that stop on its own page. In a Pew Research Center panel (opens in a new tab) of 900 US adults in March 2025, people ended their browsing on 26% of Google pages showing an AI Overview. That is the AI summary at the top of results; the figure was 16% for pages without one.
Why does the number move so much?
Because it depends on how a session is defined and how much browsing the panel captured.
The researchers tried three gaps of inactivity to separate one session from the next. With a 15-minute gap, 46.9% of assistant sessions had no web step; with a 60-minute gap, 29.2% did. The shares with web activity only before, or only after, the assistant barely moved.
Coverage matters too. Among people with at least one day of recorded browsing, 37.7% of sessions had no web step. Among people with browsing on all 28 days of the month, it was 26.7%.
Assistant use in mobile apps and through APIs was not observed at all. So the honest range is roughly a quarter to nearly half of assistant sessions. The point the authors stand behind is the ordering: assistant sessions stay self-contained more often than search sessions, under every definition they tried.
Does it differ by country or by assistant?
Sharply: US sessions stayed inside the assistant far more often than British ones. Looking at sessions that used a single assistant, 50.9% were self-contained in the United States against 20.7% in Great Britain. By assistant, the share was 48.8% for Gemini and 27.3% for ChatGPT.
In every split, assistant sessions were still more self-contained than search sessions, which stood at 21.6% and 17.9% in the two markets. The authors conclude that the levels “are clearly population- and provider-dependent,” while the ordering holds.
Does a session with no visit mean the question was answered?
No: the data shows that no visit happened, not that the need was met. The authors put it plainly: “Assistant-contained also does not mean resolved.” Their data held timestamps only, never the text of prompts or answers.
The self-contained sessions break down into three patterns, as shares of all assistant sessions:
| Pattern | Share |
|---|---|
| One prompt and one response | 19.0% |
| Several exchanges in one conversation | 7.4% |
| More than one conversation | 7.7% |
Many people went back to the web later. Of the self-contained sessions, 12.0% were followed by a web visit within one hour and 49.5% within six hours. The authors warn that those later visits may have nothing to do with the earlier question.
What does this mean for measuring AI’s influence on your brand?
It means referral clicks from AI assistants will undercount their influence. A user who never leaves the assistant can still be reading about you, because the assistant searches the web on their behalf.
In our hidden-searches study, ChatGPT ran a mean of 3.7 searches per buyer question before answering. Every answer that ran a search cited at least one page, and none of the 42 answers without a search cited anything. Your page can shape an answer that sends you no visitor.
Google’s conversational AI Mode shows a similar trade. In a 2026 field experiment (opens in a new tab) with 1,100 US users, sending every search to AI Mode cut clicks to outside websites by 18.8 percentage points. Yet it raised the time people spent per search session by 0.43 minutes.
We cover that experiment in our article on AI Mode and traffic.
What should you do about it?
Measure AI visibility directly, because traffic reports will miss much of it. Practical steps:
- Ask the assistants. Run the questions your buyers ask through ChatGPT, Gemini and others, and record whether you are named and what is said.
- Do not judge AI by referral traffic alone. A low click count from an assistant can sit alongside frequent mentions.
- Watch indirect signals. Branded searches, direct visits and “how did you hear about us?” answers can pick up influence that never arrives as a click.
- Make your key facts easy to quote. If the answer is all the buyer reads, prices, features and proof points need to be stated plainly on pages assistants can find.
If you want help tracking how assistants describe your brand, see our generative engine optimization service.
What does the research not tell us yet?
Much remains open, because one panel study carries most of the evidence. The main gaps:
- Breadth. The figures come from one AI-forward opt-in panel in two countries over two months, February and March 2026.
- Mobile apps. Assistant use in native mobile apps and through APIs was not observed, so quick phone lookups are missing.
- Outcomes. With no prompt or answer text, nobody can say whether self-contained sessions ended satisfied, confused or interrupted.
- Business buyers. No study has measured self-contained sessions for B2B purchase research.
- Cause. The design cannot tell whether assistants reduce website visits overall or simply attract tasks that need fewer visits. For the wider question, see whether ChatGPT is taking Google’s place.
Frequently asked questions
What share of ChatGPT sessions end without a website visit?
In one 2026 panel, 27.3% of single-assistant ChatGPT sessions had no observed web step, counting each person equally. Gemini’s figure was 48.8%, so the level depends on the assistant.
Do AI assistants send traffic to websites?
Sometimes, but less than search does. In the same panel, web activity came only after the assistant in 10.5% of sessions, and on both sides or between prompts in 37.1%.
Are zero-click sessions new with AI?
No. In the same panel, 19.5% of ordinary search sessions also ended without a page visit, though assistant sessions did so more often.
How do I measure AI influence if users never click?
Check the answers themselves. Assistants search the web before answering, ChatGPT 3.7 times per buyer question in our study, so they can rely on your pages without sending a visit.
Sources
- Iannelli, M. and Ai, A. (2026), The New Shape of Search: How Conversational AI Recomposes Information Seeking (opens in a new tab), arXiv:2607.04282.
- Chapekis, A., Lieb, A., Shah, S. and Smith, A. (2026), Investigating Click Behaviors On Google Search Result Pages That Produce an AI Overview (opens in a new tab), arXiv:2608.04831.
- Wang, S. T., Gleason, J., Bart, Y., Wilson, C. and Metaxa, D. (2026), AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence (opens in a new tab), arXiv:2608.18352.
- Underneath (2026), The hidden searches AI assistants run before they answer
- Underneath (2026), How much traffic would AI Mode as Google’s default cost you?