Why AI summary visibility and website visits should be measured separately

Pew Research Center analyzed the Google search behavior of 900 US adults in March 2025. Users clicked a standard search-result link on 8% of visits to pages with an AI summary, compared with 15% of visits to pages without one. They clicked a source link inside the AI summary on 1% of visits to pages where a summary appeared. This is an observed difference between search conditions, not a controlled experiment proving a causal effect from AI summaries alone.

After revising its bot-detection method, Wikimedia reported that human pageviews on Wikipedia had fallen by about 8% compared with the same months in 2024. The foundation offered AI and social media as possible explanations, but the figure does not isolate the effect of any specific search feature.

Google, by contrast, said in August 2025 that the total volume of organic search clicks was broadly stable year over year. Click rates per user, pageviews for a specific site, and total clicks across search are different metrics. The three findings cannot be combined into one decline rate, nor does any one result automatically disprove another.

A Search Engine Journal article about Wikipedia, collected in September 2026, prompted this review. This note compares earlier behavioral evidence and platform explanations linked from that article; it is not a report of newly measured behavior in September 2026.

Even when a brand or piece of content appears in an answer during exploration, whether the user goes on to visit the site must be checked separately. Treating visibility, citation, and clicks as one outcome makes it difficult to tell where the customer encountered information and whether they moved to the next action.

Content planning can distinguish questions that an AI summary can fully resolve from questions that still require opening the original source. The idea that detailed comparisons or verification materials may create a reason to visit is an execution hypothesis. The current evidence does not support prescribing a particular content format as a way to increase clicks or purchases.

This document is categorized under the exploration stage because it draws on observed post-search clicking. It does not measure purchase intent, conversion, or satisfaction, so it should not be extended to evaluation or purchase stages.

• For the same topic and similar search intent, do original-source visits and follow-up searches differ depending on whether an AI summary appears?
• How often does a brand appear in an answer, receive a source citation, and receive an actual visit?
• Is the same direction observed in Korean-language search and in specific product categories?

Measurement should hold constant, or explicitly separate, query, country, language, device, and search time, while retaining the denominator for every rate. These are proposals for follow-up measurement; no such measurement was run in this work.

• Google users are less likely to click on links when an AI summary appears in the results · Pew Research Center · 2025-07-22 · Original. Primary evidence on click behavior among US users.
• New User Trends on Wikipedia · Wikimedia Foundation / Diff · 2025-10-17, with 2026-02-13 also displayed on the page · Original. A separate observation of visit changes across all language editions.
• AI in Search is driving more queries and higher quality clicks · Google · 2025-08-06 · Original. The platform's account of total organic search clicks.
• What Wikipedia Reveals About AI Overviews And Web Traffic · Search Engine Journal · 2026-09-11 · Original. The S1 article that prompted this review; it is not counted as a fourth independent behavioral sample.

• Pew's behavior logs cover March 2025, while search results were recollected from April 7 to 17. The AI summaries shown at the time of use may differ from those seen during recollection.
• The classification centers on a US Google sample. The Wikipedia data is global and covers all language editions, so it is not directly compared or averaged with the US sample. Applicability to Korea is unconfirmed.
• Wikimedia's bot-detection rule change constrains the time-series comparison. Google's statement about total clicks does not provide raw data in the post that could be independently reproduced.
• The roughly 5% estimate from a paper mentioned by the collected article was not used as factual evidence because the original paper was not verified for this draft.
• The four sources are four reference documents, not four repeated observations under identical conditions. Verisca's synthesis remains E1 and does not establish a causal effect of AI summaries or a purchase outcome.