Zero-click search

A zero-click search ends without a visit. The person typed a question, read what the results page or the AI answer told them, and left. Nothing went wrong for them. For the sites whose pages supplied the answer, the visit that used to be the payoff never happened.

How big is the effect?

The best public measurement is Pew Research Center's. It tracked the real browsing of 900 US adults through 68,879 Google searches in March 2025, 18% of which showed an AI summary.

• Clicked a traditional result

  With an AI summary

  8% of visits

  Without one

  15% of visits

• Clicked a link inside the summary

  With an AI summary

  1% of visits

  Without one

  none shown

• Ended the browsing session there

  With an AI summary

  26% of pages

  Without one

  16% of pages

Before AI summaries the pattern already existed. A SparkToro and Datos panel found 58.5% of US Google searches ended without a click between September 2022 and May 2024, through featured snippets, knowledge panels, instant answers and searches that were simply refined. AI summaries extend it to longer, more considered questions, the kind where a buyer used to open three tabs.

What does a zero-click answer look like?

Here is one from our own tracking. Attensira asks ChatGPT and Perplexity "how to block AI crawlers in robots.txt" every day. In the 30 days to 4 October 2026 the two engines gave 63 answers, 33 from ChatGPT and 30 from Perplexity, read over the Attensira MCP:

• No answer named a brand. Not ours, not anyone's. The engines answered with the rule itself: which user-agent tokens to write, and what each does.
• ChatGPT built its answers from Google and the AI operators. developers.google.com took 608 of the 1,507 citations in its 33 answers, anthropic.com 317 and help.openai.com 314.
• Perplexity built them from publishers. searchenginejournal.com (20), github.com (19) and cyberciti.biz (19) were among the most-cited of its 290 citations.

Somebody asking that question gets the complete answer, with the directive they need, in the chat window. Whether they ever click through to one of those pages, nobody in this picture can see. The pages that were cited still did the work: they decided what the answer said. How-to questions follow the same pattern: no brand was named in 63 answers to "how to block AI crawlers in robots.txt" or 66 to "what is GPTBot", while buying questions name products, Profound in 92 of 114 answers to "best GEO tools".

How is zero-click different from a low click-through rate?

Click-through rate is a property of one result: impressions in, clicks out. Zero-click is a property of the whole search: did the session reach any website? A page can have a healthy click-through rate on searches that still mostly end without a click, because most of them never showed it a click to win. Treating them as the same number leads to fixing titles when the real change is the answer above them.

What should you measure instead?

Measure the answer. Two questions replace "did they visit":

1. Was your brand in the answer? How often each engine names you when it answers your buyers' questions: share of voice.
2. Was your page the evidence? How often it links your domain: citation rate.

Search Console still reports Google impressions and clicks, including those from AI Overviews inside the Web search type. It cannot tell you what ChatGPT, Perplexity or Claude said about you. That has to be read from the answers, which is what Attensira does daily, per engine, with the sample size beside each rate.

What do people get wrong about it?

• Reading it as "SEO is dead". The pages cited in AI answers are largely pages that rank. Being retrievable still decides whether you are read.
• Expecting a fix that brings the click back. The searcher's need was met. The opportunity is to be what met it.
• Blaming a traffic dip on AI without checking. A drop can come from rankings, seasonality or tracking. Compare the questions that show AI answers with those that do not before you conclude anything.
• Linking presence to revenue as if it were measured. Nothing joins an answer a person read to a later purchase. Say it is an inference when you make it.

What does Attensira do about it?

The crew starts from the questions your buyers ask, reads what each engine answers today and finds the pages those answers lean on. Then it does the work to change the answer: your own page rewritten to answer the question in its first lines, a placement on the pages that decide it, a fixed crawler rule. In the default Draft mode, each piece waits for your approval, and the next readings show whether the answer moved by more than the noise.