What AI Tools Actually Send Traffic To (And What They’re Erasing)

There is a chart in the Datos State of Search Q2 2026 report that tells you more about the future of the web than any headline about AI adoption. It ranks the domains people visit right after they use an AI tool like ChatGPT, Gemini, or Claude. It is a map of what AI hands you off to, and what it quietly swallows.

Read it closely and you see the reference web being erased in real time, while a different kind of website survives.

The data covers US desktop clickstream from tens of millions of users, comparing Q2 2025 with Q2 2026. “Destination domain” means the next domain visited after interacting with an AI website.

Key takeaways

• Three AI assistants now sit in the top 10 destinations from AI tools, up from one a year ago. AI increasingly sends people to more AI.
• Google is still the number one place people go after using an AI tool. AI became the research step, not the replacement for search.
• Wikipedia fell out of the top 15 entirely and NIH dropped to #15. The reference sites AI was trained on are the ones it stopped sending traffic to.
• The destinations that survived are places you do something, not places you only read. Being read by AI is not the same as being visited.

AI now feeds AI

A year ago, exactly one AI assistant appeared in the top 10 domains people visited after using an AI tool. ChatGPT sat at #10, and that was it.

By Q2 2026 there are three. ChatGPT climbed to #4, Gemini entered at #6, and Claude entered at #9. People finish with one AI tool and go straight to another.

This is what a maturing multi-model world looks like. The behavior underneath it is probably a mix of comparison, asking the same question of a second model, and using different assistants for different jobs. Whatever the exact motive, the pattern is a loop. Increasingly, the answer to “where do people go after AI” is more AI.

For anyone hoping AI traffic would spill out onto the open web, that loop is the opposite of good news.

Google survives as the action layer

Here is the part that complicates the “AI is killing search” story. After all that AI use, the single most common next destination is still Google, holding the #1 position year over year.

People ask an AI to explain something, then go to Google. To verify a claim. To find the actual link the AI mentioned but did not give them. To do the transactional or navigational follow-up a chatbot cannot finish.

That is a division of labor, not a replacement. AI became the research step and moved upstream of search. Google remained the action step. I made the platform-level version of this argument in why the data disagrees with the Google-is-dying narrative, and this chart is the behavioral proof of it.

The reference web is being erased

This is the movement that should stop you. Wikipedia was the #7 destination from AI tools in Q2 2025. In Q2 2026 it is gone from the top 15 completely. NIH fell from #8 to #15, barely hanging on.

AI tools used to send people to Wikipedia and NIH to read more. Now the AI is the read-more. It absorbed the answer and kept the click.

Sit with the irony. Wikipedia is one of the most heavily used sources in AI training and grounding on earth. Being that source is exactly what pushed it off the destination list. The model learned from Wikipedia so thoroughly that it no longer needs to send anyone there.

That is the fate worth understanding, because it is the one every information-only website is pointed at. Being read by AI is not the same as being visited by a person. You can be the source behind a million answers and receive zero of the clicks.

Reddit shows the same pattern at a slower pace, slipping from #6 to #10 as an AI destination even while its overall audience grows. AI leans on Reddit as a source and sends relatively fewer people to it. Reddit’s saving grace is that people also go there directly, which Wikipedia cannot say.

What survives tells you the rule

Look at who is still on the 2026 list and climbing. Google, Microsoft (up from #5 to #3), GitHub, Amazon, plus first-time entries Apple and WhatsApp.

None of these are places you go to read. They are places you go to do. Log in, buy, write code, message someone, complete a task. The AI can describe all of them, but it cannot perform them for you, so the handoff still happens.

Watch the split inside Microsoft. The company climbed to #3 while Bing dropped out of the top 15 entirely. The ecosystem you log into survived. The search box you could have just asked the AI instead got skipped.

So the survival rule is simple to state and hard to live by. Sites that only offer information the model can restate get absorbed. Sites that offer an action the model cannot complete get the visit.

What this means for your website

The goal was never to get cited by AI as an end in itself. Getting cited the way Wikipedia gets cited means getting consumed. The durable position is to be the place someone still has to visit to finish the thing the AI started.

• Audit what your site actually offers beyond information. If a model can fully answer what you do without anyone visiting, you are on the reference-site trajectory.
• Give people a reason to arrive that requires arriving: booking, buying, a login, a tool, a quote, a conversation with a real person.
• Still get cited, because citations are how you enter the consideration set. Just do not mistake being cited for being chosen.
• Measure visits and actions, not just mentions. A brand that shows up in a hundred AI answers and gets no bookings has a distribution win and a business problem.

This is the same shape as a pattern I documented earlier: AI users and Google users visit measurably different websites. You can see the original version in how AI searchers visit different websites than Google searchers. The Q2 data shows that gap widening, with AI pulling its users into a tighter loop.

It gets sharper still when the visitor is not a person. When an agent does the browsing, the reference-site problem becomes total, because software never clicks through to read more for its own sake. I covered that in information agents, when the searcher is software.

Two things this chart does not show

Read it with two limits in mind. It is desktop only. A large and growing share of AI use happens in mobile apps that never appear in this panel, which most likely means the closed-loop effect is understated here, not overstated.

And the AI-to-AI movement partly reflects power users running several tools at once, not people abandoning one assistant for another. The loop is real, but it is a story about how AI users behave, not proof that any single tool is losing.

FAQ

The web AI is building routes traffic to two kinds of places: more AI, and sites where you do something a model cannot do for you. Everything in between, the vast middle of pages that exist only to inform, is being read by the machine and visited by no one.

If you are not sure which side of that line your site sits on, the honest test is whether an AI can fully answer what you offer without anyone ever arriving. If it can, the guide to getting cited by AI is where to start, and then the real work is giving people a reason the citation alone cannot satisfy.

Sources

1. Datos (a Semrush company), State of Search Q2 2026: Behaviors, Trends, and Clicks Across the US & Europe. Top US AI destination domains, desktop clickstream, Q2 2025 versus Q2 2026.
2. Datos methodology note: destination domains are defined as the next domain visited following interaction with an AI website.