The figures do not convert

A field experiment on Google Search was posted to arXiv on 18 August 2026. Along the way it reports, as background, that AI Overviews reduced Wikipedia's traffic, and cites a working paper by Khosravi and Yoganarasimhan for the size of the drop. That is an ordinary sentence in an ordinary paper, and the only reason to start here is that it can be checked in two clicks. A figure inside a citation is a summary of a measurement, and a score is not a record: the summary is the part that travels, and the measurement stays where it was made.

Open that paper's arXiv listing today and it reports something else. Version 4, posted on 12 May, reports 15% in its abstract; version 7, posted on 25 October, reports 5.45% and 4.82%. Retrieved 31 October 2026. The later versions are not a correction of the earlier arithmetic. They report a different quantity: the early abstracts describe a daily-traffic effect on English Wikipedia set against the Hindi, Indonesian, Japanese and Portuguese editions, across 161,382 article-language pairs; the current one describes monthly external-search referrals set against the German and French editions. Same title, same authors, different measurement.

Nothing went wrong here. When the experiment was posted on 18 August, version 4 was the newest on arXiv; version 5 appeared eight days later. That is the shape of the thing this piece is about. Not error, and not carelessness, but a figure still in circulation under its label after the measurement underneath it has moved. A number does not stop at the publish button either.

The ledger

Set that beside the figures about Google and publisher traffic in circulation this autumn. Nine of them appear in this piece, and each one gets its basis printed alongside it, because the basis is the subject.

• 40.2%, with 21.9% for the year before: a year-over-year decline for Google Search across Chartbeat's publisher network, July 2025 to July 2026, as Search Engine Journal reports it from Chartbeat's 2026 Publisher Playbook. Unit unstated. No site count for either year.
• Nearly 46%: a fall in Google Search's share of pageviews, roughly 9% to 5%, July 2024 to July 2026, on Chartbeat's own resource page.
• 34%: pageviews from Google Search, December 2024 to December 2025, in Chartbeat's Q1 report.
• 33% globally, 38% in the United States: Google organic search traffic to over 2,500 sites, November 2024 to November 2025, Chartbeat data inside a Reuters Institute report.
• 18.8 percentage points, or 14.9, depending on what counts as a search: a click-through effect measured on 714 people assigned to a search setting in March 2026.
• 8.8 percentage points: a second effect from a different arm of the same experiment, 742 people in the comparison, 615 in the LATE estimate (a local average treatment effect). The same arm's intention-to-treat result is 2.7 points up and not significant; 8.8 is significant only under an assumption the authors flag.
• 39.8%: a figure in circulation attributed to a paper I could not read.
• 31.1 billion against about 32 billion: average monthly pageviews across Chartbeat's network, January to July 2026, against the 2025 average.
• The Wikipedia citation above: one quantity in the early versions, two different ones in the current abstract.

Nine figures, and the label is the part of each that is cheap to count: short, quotable, and silent about which of the nine it is.

One figure, three renderings

Take the 40.2%. Search Engine Journal published it on 24 September 2026, reporting Chartbeat's Playbook: "Chartbeat reports a 40.2% year-over-year decline for Google Search across its publisher network from July 2025 to July 2026." The Measure reported the figure on 23 September, a day earlier, and rendered it differently: "But over the past year, pageviews generated by Google search are down 40.2%." AFP, via TechXplore on 16 October, put it at a 40% drop.

One figure, three sentences, and they are not interchangeable. Search Engine Journal says "decline for Google Search" and leaves the unit where the Playbook left it. The Measure says pageviews. AFP rounds, and relabels unit and population.

That is not a complaint about any of the three, because the thing being rendered does not state its own unit, and Search Engine Journal says so in the same article that carries the figure: "The playbook doesn't specify whether its year-over-year declines refer to the number of pageviews from each channel or each channel's proportion of pageviews." And: "The report doesn't include a site count for either year." And, on who the network is, that it "skews toward news and media publishers".

Those three caveats are Search Engine Journal's own, not Chartbeat's. They matter because of where they are: printed in the same piece as the number they qualify. The objection does not have to be constructed from outside. It shipped attached.

I have not read the Playbook. This site has argued before that teams report the number everyone knows is wrong, but that argument is about which number an organisation chooses to report about itself, and this is a different question: what an outside figure is before anyone chooses anything. What I can say about the Playbook is what two trade articles say it says, plus one sentence Search Engine Journal quotes from it, about "AI-generated search results and a shifting referral landscape" putting "real pressure on audience acquisition". The Playbook itself I did not read.

Neighbours that are not the same quantity

Three more figures sit close enough to the 40.2% to be mistaken for it. All four run on one vendor's data.

Chartbeat's own resource page gives a second: "In July 2024, Google Search accounted for roughly 9% of pageviews. By July 2026, that share had dropped to 5%, representing a nearly 46% decline over two years." Two-year window, a share rather than a count. Search Engine Journal reports that the two-year figures represent shares, which is more than the Playbook's own figure says of itself.

Chartbeat's Q1 report gives a third: "Pageviews from Google Search fell 34% from December 2024 to December 2025." December to December, pageviews, stated outright.

The Reuters Institute's trends report carries Chartbeat data over a fourth window: "Aggregate data, sourced from the analytics company Chartbeat for this report, show that Google traffic from organic search to over 2,500 sites was down by a third (33%) globally between Nov 2024 and Nov 2025 and by 38% in the United States, but it is not clear how much of this is down to AI overviews." That last clause is doing narrow work. It qualifies the November-to-November figure, not search traffic in general.

Four windows, at least two units, and no arithmetic that gets you from any one of them to any other. I am not going to supply that arithmetic. Multiplying rounded second-hand numbers together produces something that looks like a measurement and is not one. The totals have the same problem: both trade articles carry an average of 31.1 billion monthly pageviews across the network for January to July 2026, against about 32 billion for the 2025 average, which is a fifth window, with no 2024 total to extend it.

The experiments

Two of the nine come out of a randomised experiment rather than an analytics panel. That changes what they are, not how carefully they need reading.

The August paper recruited US-based participants between 17 and 19 March 2026 and analysed 1,100 people across three arms, each assigned a condition that ran for seven days. The best-known figure is not about all 1,100. The comparison that produces it, assignment to AI Mode against current Google Search, rests on 714. The result sentence reads: "significantly reduced click-through rate by -18.8 percentage points (pp)", with a 95% interval of −22.2 to −15.3 points, below zero throughout.

That figure carries a counting convention, and the convention is load-bearing. The paper's supplementary table gives two estimates for the same comparison: 18.8 points down if each prompt in an AI Mode conversation counts as a search, and 14.9 points down if the whole conversation counts as one. The authors state the rule they used: "Each turn of a conversation in AI Mode is counted as a separate search." One experiment, one decision about the denominator, two numbers. Neither is a percentage fall in clicks, and I did not find a control-arm baseline in the paper's text that would turn either into one.

The authors are plain about the setting: "our AI Mode condition reflects a forced, full-adoption scenario rather than the gradual, opt-in uptake most users would experience in practice", and "Our estimates reflect an early snapshot of Google's AI Mode". The third arm, in which a browser extension hid AI Overviews, returns an intention-to-treat effect on click-through of 2.7 points up, interval −0.7 to 6.2, p = 0.124, which is not significant, on 742 people. The larger 8.8-point figure from that arm is a local average treatment effect, and it rests on how well the hiding held: "Compliance in the No AI Search condition degraded over the study period due to a Google interface change, falling to roughly 50% overall."

A paper that states its own limits this plainly has done the reader's work already, which is also what makes quoting one of its numbers without them so easy. Reading the headline sentence and stopping is the research equivalent of signing off without opening the file, and cursory approval is not review.

Then the one I could not check. A figure of 39.8% is in circulation attributed to another paper, reported by ppc.land on 1 and 4 July 2026 as outbound organic clicks when an AI Overview was shown, from 1,065 US desktop Chrome users, data collected between 7 January and 10 February 2026. I could not read the paper. SSRN returned 403 on three routes on 31 October. So I cannot tell you what it reports. I can tell you that a figure attributed to it is in circulation, which is a different kind of statement, and the difference is the whole point of printing it here.

What the labels cannot tell you

Line the nine up and the pattern is not that any one of them is wrong. It is that the label each travels under carries no unit, no window and no population. A share and a count are different quantities. July to July and November to November are different windows. An unnamed network of publisher sites and 714 recruited US participants using Chrome are different populations. Nothing here converts into anything else.

None of this argues the decline away. The decline is in the record as an observation, from several directions at once, and Chartbeat's own page advises publishers, among its takeaways, to "look for more ways to reach readers through channels you can influence directly", which is its advice to give and not this piece's subject. What is not in the record is a single quantity that all nine of these figures are measurements of.

What I did not read also bears on every paragraph above: the Chartbeat Playbook; the paper behind the 39.8%; Google's 15 October announcement about links in AI search content, which reached me only through an AFP report; and Adweek, Digiday, Press Gazette and Axios on the same figures. Every query I ran was in English.

I read the nine as the ordinary condition of a figure in circulation rather than a failing of anyone who printed one. This site has written about the one number you can't buy; the harder problem with an outside figure is not that it can be bought but that it cannot be converted.

Forty percent, a third, 18.8 points. The labels are short and they convert cleanly into one another in conversation, which is the only place they convert. Underneath are nine figures, each with its own unit, its own window and its own population, and not one of the labels says which. The same label is not the same measurement.