AI summaries and fewer clicks: what the Pew data can tell your SEO team

A search result can be seen without bringing a visit

A page can appear in an answer, receive no visit, and still influence a buying decision. It can also receive more visits that produce no useful enquiries. These are different outcomes. A good SEO or answer-engine optimization review names the outcome before it recommends a change.

The practical question is not whether AI will replace every click. It is which part of your own discovery-to-customer journey changed, and what evidence would justify spending time on it. The research below gives a reason to investigate that journey. It does not supply a replacement traffic target.

Read the study with its dates attached

Pew Research Center published this analysis on July 22, 2025. It used March browsing records from 900 U.S. adults and identified 68,879 unique Google searches; 12,593 produced an AI summary in the collected results. Crucially, researchers reran searches to collect result pages on April 7–17. Those pages may differ from what people originally saw.

The reported traditional-result click rates were 8% for visits associated with an AI summary and 15% without one. Summary-source clicks occurred on 1% of visits with a summary. Session-ending rates were 26% versus 16%. These are observations about that study, not a randomized estimate of what adding an overview causes.

Use the chart without changing its denominator

Each bar is a reported percentage of visits to the relevant kind of results page. It is not a percentage of all participants, an estimate of AI market share, or your site’s click-through rate. The published unique-search counts should not be substituted for the raw visit denominators to manufacture exact click totals.

The two traditional-link bars belong together. The session-ending bars describe another outcome. Summary-source clicks apply to pages with summaries; an absent-summary category has no corresponding summary link. Do not invent a zero bar and treat that as a controlled comparison.

Make three small records instead of one dramatic traffic story

Keep a search record: date range, search type, country, device, query group, landing page, impressions and clicks. Keep an answer record: exact question, product or API used, search mode, model label when available, language, collection time, answer and source URLs. Keep an outcome record: relevant enquiries, sign-ups or completed tasks on the destination page.

Link these records by a documented page or question theme. Do not claim they identify the same person or the same impression unless your collection actually supports that join. In a small team, a weekly spreadsheet with clear definitions is often more useful than a daily blended visibility score.

Search Console is useful, but it is not an AI attribution switch

Google’s documentation says AI Overviews and AI Mode traffic is included in Search Console’s Web search reporting. It also says ordinary SEO foundations apply and that eligibility does not guarantee inclusion. Treat the report as search evidence, not an automatic explanation of why a page changed.

Start with a comparison you can inspect: the same landing pages, device category, geography and sensible date windows. Separate branded queries from unaided category questions. Check whether a new campaign, a changed title, tracking consent, seasonality or a different query mix could explain the result before attaching an AI label.

Turn the evidence into one page decision

Imagine a fictional product page for a booking tool. Its impressions rise, clicks fall, and trial starts stay flat. A sensible first task is to inspect which queries now reach that page and whether the result promises the right thing. Rewriting the whole website to chase an unspecified AI score would be premature.

If answers repeatedly omit a constraint buyers care about, improve the relevant public explanation and its supporting evidence. If visits arrive but buyers cannot find the price or next step, fix that page journey. If impressions and relevant visits are stable, an alarming industry headline alone is not a reason to replace useful content.

Write a decision note before you recheck

Use this format: “For this page and buyer question, we observed X in this dated sample. We changed Y because it addressed Z. We will review the same evidence fields after the page has been processed again. We will keep or reverse the change based on the buyer outcome and the evidence quality.”

Choose a recheck date and preserve the original answer rather than repeatedly refreshing until a favorable one appears. A changed answer can justify another investigation. It cannot, by itself, establish that the page edit caused the change.

A useful next step

Pick one commercially important question and one destination page. Save a baseline, check that the answer describes your product accurately, and make one evidence-backed improvement. Read /methodology for how DoneHat separates observations from claims, or /demo to inspect a sample workflow. If you later need repeat checks across a larger question set, evaluate /pricing against that concrete job rather than a promised traffic uplift.

Sources

• Google users are less likely to click on links when an AI summary appears in the results — July 22, 2025 — Pew Research Center (accessed 2026-10-17)
• AI features and your website — documentation reviewed October 17, 2026 — Google Search Central (accessed 2026-10-17)