AI Search

AI summaries, citations and clicks: how to read the evidence

What Pew's 68,879-search study can tell marketing teams about citations and clicks, with complete tables, interactive charts and the limits of the comparison.

Abstract source cards feeding a translucent answer panel beside a separate path to a website doorway

Key takeaways

  • An AI citation and a website visit answer different measurement questions.
  • Pew's historical browsing study describes observed differences; it does not isolate the causal effect of showing an AI summary.
  • Keep study dates, query groups and denominators visible whenever you reuse a percentage.

An AI answer can reference a page without sending its reader to that page. That makes citation visibility useful, but it gives a marketing team a different signal from a recorded website visit. Before turning a research headline into a content target, identify which signal the underlying number actually measures.

Pew Research Center's July 2025 analysis provides a concrete case. Across the visits it examined, traditional-result clicks were less common on result pages associated with an AI summary. The practical lesson is to read the comparison carefully before using it to explain your own results. A lower observed rate is a finding about those groups, during that study period. It is not a ready-made forecast for your brand. [1]

Start with what was observed

The research combined browsing records from 900 U.S. adults during March 2025 with search-result snapshots collected in April. Its dataset contained 68,879 unique Google searches; 12,593 produced a summary in the collected results. That timing matters: the reconstructed results were not necessarily identical to the page each person saw during the original browsing session. [1]

For a reader evaluating the finding, this is the first boundary to keep intact. The study offers historical observational evidence. It does not randomly assign the same searcher to otherwise identical result pages with and without summaries. Differences in the people, questions and situations represented by each group can therefore remain part of the comparison.

Different observed outcomes on Google result pages

Published percentages for visits associated with an AI summary and visits with only traditional results.

2619.5136.50Clicked a traditional reEnded browsing sessionObserved action after the result page
  • With an AI summary
  • Without an AI summary
Full data table
Different observed outcomes on Google result pages · % of visits
Observed action after the result pageWith an AI summaryWithout an AI summary
Clicked a traditional result815
Ended browsing session2616

Pew Research Center: March 2025 browsing and April 2025 result snapshots. These are observed groups, not randomized treatment groups. The chart does not estimate a causal effect. [1]

A percentage-point difference is not a causal estimate

The traditional-result percentages differ by seven percentage points: 8 minus 15. Expressing the same arithmetic relative to the 15% comparison value produces a 46.7% lower observed rate. Both calculations describe the published percentages. Neither calculation tells you how many visits a particular site lost because of an AI summary.

CalculationResultInterpretation
Traditional-result rate: 8% minus 15%−7 percentage pointsAbsolute difference between the reported groups
Relative comparison: (8 − 15) ÷ 15 × 100−46.7%Descriptive difference using 15% as the reference
Session-ending rate: 26% minus 16%+10 percentage pointsAbsolute difference between the reported groups

These are Ansvra calculations from Pew's rounded published values. They do not add precision to the underlying observations. A content plan should not convert them into an expected traffic loss, an assumed recovery target or a promised return from improving citations.

Use the two forms of arithmetic for different communication needs. Percentage points keep the comparison close to the original rate. A relative percentage emphasizes the size of the gap compared with a reference value. When sharing either, retain the original rates beside it so a reader can reconstruct the calculation.

Two glass frames holding different paper shapes, with a magnifying lens between them
Different comparison groups require a careful reading of the evidence.

The question being asked changes the context

The same research reported different summary frequencies across query groups. Short searches, long searches, question-word openings and full sentences did not produce the same observed pattern. The following chart keeps those groups separate; they overlap, so their percentages should not be added together. [1]

Summary frequency varied with the kind of query

Published percentages of searches in each query group that produced an AI summary in Pew's collected results.

604530150One or two wordsTen words or moreQuestion-word openingFull sentenceQuery group
  • Summary present
Full data table
Summary frequency varied with the kind of query · % of searches in group
Query groupSummary present
One or two words8
Ten words or more53
Question-word opening60
Full sentence36

Historical March/April 2025 study. Groups overlap and have different denominators; these values must not be added together or treated as a market-share composition. [1]

For a brand, the useful editorial question is which kinds of customer questions its content can answer clearly. A broad category query, a comparison and a detailed planning question can require different evidence on the destination page. Start by identifying the decision the reader is trying to make, then check whether the page provides the information needed for that decision.

Later platform observations add context without updating the older study's rates. In May 2026, Google described average AI Mode queries as three times the length of traditional Search queries. That is a first-party statement about a different experience and observation context. It should not be plotted as a later point in Pew's browsing series or treated as proof of a change in the study's click behavior. [2]

A citation is evidence of a reference, not a visit

A citation count tells you that content was referenced within the measurement's coverage. A visit requires a separate observation of someone reaching the website. Keeping those definitions explicit prevents a visibility report from quietly becoming a traffic report.

Microsoft makes this distinction concrete in its AI Performance definitions. Its citation reporting counts references displayed across supported experiences, while the definition does not establish placement within an individual answer. Its grounding-query data is also sampled. Those boundaries are a reason to inspect the supporting evidence before drawing conclusions from a total. [3]

For Ansvra readers, a useful sequence is to open the saved answer, inspect whether the brand was mentioned, identify any cited page, and then ask which separate evidence would establish a visit or a business outcome. This keeps a discussion about presence grounded in what was actually recorded.

You can use AI citation analysis to investigate references and source analysis to understand the pages behind them. Those views help frame the next question; they do not turn an external study's click rate into a measurement of your own audience.

A practical route from a statistic to a decision

Ansvra's editorial framework for keeping evidence and interpretation connected.

  1. Locate the number

    Open the original source and identify the exact result and collection period. [1]

    Next: Name the denominator

  2. Name the denominator

    Write down whether the percentage describes searches, visits, sources or another population. [1]

    Next: Check the comparison

  3. Check the comparison

    Identify differences between the groups and distinguish observational evidence from a causal estimate. [1]

    Next: Separate the signals

  4. Separate the signals

    Keep a recorded mention, displayed citation and website visit as distinct observations. [3]

    Next: Choose a focused review

  5. Choose a focused review

    Inspect the saved answer and cited page, then define the next evidence needed before changing content. [1] [3]

This framework is an editorial interpretation of the source methodology and measurement definitions, not a measured conversion funnel. [1] [3]

Turn the research into a decision note

A short decision note is often more useful than repeating a striking percentage in a presentation. Write down the question you are investigating, the evidence you already have, what it cannot establish and the next observation that would change your decision.

If a useful page is cited but referral activity is unclear, start with the measurement gap. Check the relevant traffic records and their definitions before concluding that the citation succeeded or failed commercially. If the saved answer references a different page, inspect what that page contributes to the answer before proposing a rewrite.

When a page is absent, keep the diagnosis open. A missing citation in a recorded answer does not, by itself, establish whether the reason was content coverage, retrieval, access, a different source choice or variation in the answer. Record what the answer supports, then choose a narrowly defined investigation.

Make the proposed content change specific enough to review. Identify the reader question, the information the page should provide and the evidence needed to support it. Avoid a general instruction to “optimize for AI” when a clearer action would be to explain a comparison, document a limitation or make a definition easier to find.

Finally, define the later comparison before making the change. Keep the tracked questions and the reporting scope visible. A follow-up review can then describe what changed in the recorded answers while leaving unrelated movements in traffic or demand open to separate investigation.

What this analysis leaves unanswered

The published percentages do not supply your site's conversion rate, the value of an individual citation or a current benchmark for every market. This article also does not reconstruct respondent-level records or estimate uncertainty from rounded aggregates. The complete tables above contain the selected published values used in this analysis, rather than a claim to reproduce the full underlying browsing dataset.

That limit is useful. It tells a team where to stop repeating the external number and start asking for evidence specific to its own situation. Research can sharpen a question without resolving every operational decision that follows from it.

Your evidence review checklist

  • Write the exact outcome being measured: a mention, citation, visit or business result.
  • Keep the original study period and population next to any reused statistic.
  • Show the denominator and original rates beside a derived percentage.
  • Read saved answers before deciding which content deserves attention.
  • Define the next comparison before changing the page.

Questions about the evidence

Does this study prove that AI summaries caused a particular site's traffic to fall?

No. The published comparison describes observed groups and the study's collection method. It does not isolate a causal effect for an individual website. [1]

Can a citation count be used as a website-visit count?

No. A displayed source reference and a visit are separate observations. Microsoft's citation definitions are a useful example of why the measurement's coverage and meaning must remain explicit. [3]

Do the query-group percentages add up to the whole search audience?

No. A long question can also be a full sentence and begin with a question word. The groups overlap, and each percentage uses the searches in that group as its denominator. [1]

Sources

  1. Pew Research Center: Google users are less likely to click on links when an AI summary appearsAccessed
  2. Google: How AI Mode is changing the way people search in the U.S.Accessed
  3. Microsoft Bing: Introducing AI Performance in Bing Webmaster ToolsAccessed

Discussion

Join the discussion

Start the conversation

Share a question or tell us what worked for you.

Comments appear after review. How we handle comments

The Ansvra journal

Keep reading

Browse all articles

Ansvra

A clearer next step.

MEET YOUR AI GUIDE

A little clarity.
A clearer next move.

Ask about Ansvra, AI visibility, or your next step with content.

Opening chat…