Pew's 1% counts 12,593 searches. The sentence quoting it says 68,000.

What is the 1% a share of?

Pew’s study tracked 900 US adults who agreed to share their browsing activity. Of those, 58% ran at least one Google search in March 2025 that produced an AI summary. Across the full dataset, “Some 18% of all the Google searches in our study generated an AI summary as part of the search results.”

The 1% sentence is explicit about its denominator: “Google users who encountered an AI summary also rarely clicked on a link in the summary itself. This occurred in just 1% of all visits to pages with such a summary.”

So the quantity is: among visits where a summary appeared, the share where someone clicked a link inside the summary. The figure beside it, which gets quoted less often, is the one measuring traffic loss. “Users who encountered an AI summary clicked on a traditional search result link in 8% of all visits. Those who did not encounter an AI summary clicked on a search result nearly twice as often (15% of visits).”

Two measurement limits in Pew’s own text matter for anyone reusing the 1%. A click counted only “if the next-visited URL matched any of the links available on the first page of search results or in the first three links cited in a page’s AI summary”, so links past the third cited source were outside the count. And the summaries were captured later than the browsing: “Search results were collected April 7-17, 2025.” A summary can change between the visit and the scrape.

One click rate, measured four ways

The Reuters Institute has published three other figures for roughly this question, in its own reports, on different instruments.

Measurement	Unit	Who was asked or tracked	Window	Figure
Pew	share of visits with a summary where a summary link was clicked	900 US adults, browser tracking	March 2025 visits, results scraped 7 to 17 April 2025	1%
Reuters Institute, Generative AI and News 2025	self-stated frequency, among people who saw AI answers	survey, six countries	fielded 5 June to 15 July 2025	33% always or often
Reuters Institute, Digital News Report 2026	self-stated frequency, full sample including non-users	survey, 27 markets	2026	4% always or often
Same chart, our arithmetic on its own rows	self-stated frequency, among AI chatbot news users only	survey, 27 markets	2026	about 44%

The last row is the one that makes the spread legible. The Digital News Report chart publishes four rows for each platform, and the note above it says the data are “for the full sample, meaning they include non-users (in light grey), to allow more meaningful comparisons across platforms with very different user bases.” For AI chatbots the rows read: always or often 4%, sometimes 3%, rarely or never 2%, and does not use for news 91%.

Subtract the non-users and 9% of people use an AI chatbot for news, which the chapter states directly: “The large gap between AI and the other platforms reflects the much smaller proportion of people who use AI chatbots for news in the first place (9% across 27 markets)”. Four of those nine points are the always-or-often group, so among users the rate is about 44%. Run the same arithmetic on search and 19 of 43 points is also about 44%. On social and video networks, 17 of 48 is about 35%.

The arithmetic there is ours rather than the report’s, and rounding makes it soft: 4 and 9 as printed could be anywhere from 3.5 to 4.5 over 8.5 to 9.5, which spans 37% to 53%. The useful part survives the rounding. On the full-sample basis AI looks four to five times worse than search at sending people onward. On the per-user basis the two are indistinguishable, and almost all of the headline gap is the size of the audience rather than the behaviour of it.

Why does “broadly in line” need a footnote?

The October 2025 Reuters Institute report publishes the 33% figure and then says this about it:

While we have to be careful about the limits of survey data in this context, as what people say they do might not always match what they actually do, our data are broadly in line with web-tracking data on the topic, which are better at capturing people’s behaviour (see Chapekis and Lieb 2025).

Chapekis and Lieb 2025 is the Pew page. Its reference list confirms it: “Chapekis, A., Lieb, A. 2025. Google Users Are Less Likely to Click on Links when an AI Summary Appears in the Results. 22 July. Pew Research Center.”

So 33% of people who saw AI answers say they always or often click a link in them, and the web-tracking study named as broadly in line records a summary-link click on 1% of visits where a summary appeared. The two are not the same quantity. One counts people and one counts visits, and a person who always clicks still only clicks on the visits they make.

Even allowing for that, the gap is wide. If every person in the 33% group clicked on only one tenth of their summary visits, the per-visit rate would land above 3%, which is already more than three times Pew’s figure. What the sentence probably means is that both sources point the same way: self-reported clicking is lower than people assume and tracked clicking is lower still. As written it reads as agreement on magnitude. We do not know which was intended, and the cause is not something an outside reader can establish.

What did Chartbeat actually print?

The 40% comes from Chartbeat, an analytics vendor whose publisher playbook ends in a demo request. The playbook is behind a lead form, though the file itself is served without a login.

Inside the PDF, the figures appear under a column headed “YoY Decline” with the values -21.9% and -40.2%. Across the whole seventeen-page text layer there is no definition of what the column counts and no count of sites. Two pages carry embedded images we could not read by eye, so a note inside a graphic is possible.

Elsewhere the same document gives a different shape of the same story: “Google Search made up about 9% of pageviews in July 2024 and fell to 5% in July 2026, a relative drop of almost 46% in two years. Google Discover fell 28.4% over the same stretch.” Chartbeat’s own web summary of the playbook repeats it as “a nearly 46% decline over two years”.

Those two presentations do not reconcile cleanly. Chained as yearly changes, 21.9% and 40.2% compound to 53.3%, not 46%. Read as a share of pageviews, 9% to 5% is a 44.4% relative fall, which is what “almost 46%” rounds from if the underlying values carry decimals. The first reading needs total pageviews to have fallen about 13.5% over the two years for both to hold. Chartbeat’s printed totals imply less than that: 31.1 billion average monthly pageviews from January to July 2026 against “about 32 billion” in the comparison period, which is under 3%. The printed numbers do not settle which unit the yearly column uses, and we are not going to guess.

The playbook also disagrees with itself twice in small ways. Direct traffic is “15.4%” in one paragraph and 15.9% in the table on the next page, and the web summary says 15.9%. AI referrals are “just 0.01% of traffic” in one line and “under 0.01% of pageviews” in another.

One more detail about the 13 October article: the words “has plummeted” link to a Search Engine Journal story rather than to Chartbeat. The 40% reaches the reader through trade press, which is ordinary practice and worth knowing when you decide what you are citing.

Which Chartbeat number is which?

Four overlapping windows of the same Chartbeat series now sit in public, three of them inside Reuters Institute publications.

Window	Figure	Sites, as described	Where it was published
July 2024 to July 2025	-21.9%	not stated	Chartbeat playbook, file dated 18 September 2026
November 2024 to November 2025	-33% globally, -38% in the US	”over 2,500 sites”, corrected to “2,576 sites worldwide, including 797 US sites, and 1,031 sites across Europe”	Reuters Institute, Trends and Predictions 2026, 12 January 2026
December 2024 to December 2025	-34%	not stated	Chartbeat Q1 2026 report
July 2025 to July 2026	-40.2%	“hundreds of publishers”	Reuters Institute, 13 October 2026, citing the playbook

None of these chain into each other, because the windows overlap and the samples are described three different ways. The January report uses “over 2,500 sites” and “2,576 sites” and, in a separate sentence, “hundreds of news sites”. The October article says “hundreds of publishers”. A sample that is 2,576 sites in January is unlikely to be hundreds in October, and nothing in either publication says the panel changed.

The forecast figure from the same January report is worth keeping separate from all of them. “Publishers expect traffic from search engines to almost halve (-43%) over the next three years”, on a base of 268 respondents. That is a three-year expectation across all search engines, surveyed from publishers. The -40.2% is a twelve-month measurement of Google Search only. They are close enough in size to be mistaken for each other and they measure nothing alike.

How to quote either number

Six words of context turn both of these into figures you can defend.

For the Pew 1%: of visits where an AI Overview appeared, 1% included a click on one of the first three links inside the summary, among 900 tracked US adults in March 2025. The companion figure, 8% against 15%, is the one that speaks to traffic.

For the Chartbeat 40%: Google Search referrals, July 2025 to July 2026, as reported by an analytics vendor in a gated playbook that does not state the unit or the number of sites. If you need a sourced panel size for the same series, the 2,576-site figure belongs to the November 2024 to November 2025 window, not to this one.

Both of those are longer than a headline wants, which is the trade you make. A figure with its base attached survives being checked, and a figure without one gets quoted until the base disappears entirely, which is how a 1% rate measured on 12,593 searches ends up attached to 68,000 of them in a sentence written by people who read the original.

Method and corrections

Every page named here was opened on 14 October 2026 and the quoted strings were matched against the text served. All nine returned HTTP 200.

Chartbeat’s playbook sits behind a lead form; the PDF address comes from the landing page’s own form redirect and the file is served without a login. Its text layer holds no unit note and no site count, and pages 3 to 6 carry embedded images that were not inspected visually, so a note inside a graphic cannot be ruled out. The Chartbeat Q1 2026 report PDF carries no internal date and its data run to 9 February 2026.

The arithmetic here is ours, with inputs shown: 68,879 divided by 12,593 is 5.47; 12,593 of 68,879 is 18.28%, which Pew prints as 18%; 1% of 12,593 is 125.93 and 1% of 68,879 is 688.79; 21.9% and 40.2% chained give 53.3%; 9% to 5% is a 44.4% relative fall; and on the Digital News Report rows, 4 of 9 is 44.4%, 19 of 43 is 44.2% and 17 of 48 is 35.4%.

We are not saying any of these sources got anything wrong. Pew states its base in its first methods line. The Reuters Institute states the limits of survey data in the same paragraph as its own survey figure. Chartbeat is a vendor and reads like one. What happens to a number between a methods line and a sentence that quotes it is a separate problem from whether the number is good.

Live pages can be edited after publication. If any figure quoted here changes, or if we have one wrong, we will correct it and say what changed.