Hype Index: 'AI Search Cut Wikipedia's Human Traffic by 8%'

TL;DR: Wikipedia's human pageviews fell roughly 8% in May and June 2025 against the same months in 2024, and the figure is credible because the Wikimedia Foundation published it along with the caveat that it only became visible after Wikimedia updated its bot detection and reclassified a surge of traffic from Brazil that had been counted as human. I fetched 30 articles citing it: 18 of 28 carry that caveat, only 15 link the source that contains it, and 16 of 22 headlines name AI as the sole cause when Wikimedia named several. We score it 35% noise / 65% signal: the measurement is sound, the citation practice is where it decays.

This is the stat that gets used to prove AI is eating the open web. It is one of the better stats in that argument. It is also the clearest example I have found of a source doing the honest thing and the citations quietly dropping the honest part.

What is the claim?

Stated the way it circulates, in headlines from my own sample: "AI Is Killing Wikipedia's Human Traffic." "AI Search Cuts Wikipedia Traffic 8%." "Wikipedia says AI is to blame." "Wikipedia's Human Traffic Drops 8% As AI Takes The Wheel."

What did Wikimedia actually say?

Marshall Miller, Senior Director of Product for Core Experiences, published New User Trends on Wikipedia on the Wikimedia Diff blog on October 17, 2025. The number is exact and modest: "a decrease of roughly 8% as compared to the same months in 2024."

The caveat is not buried. It is in the first few paragraphs, and it is the reason the post exists. Around May 2025, Wikimedia observed unusually high amounts of apparently human traffic, most of it originating from Brazil. That prompted an investigation and an update to its bot detection systems, and the reclassified data for March through August 2025 showed that much of the unusual May and June traffic was coming from bots built to evade detection.

Read that sequence carefully, because the order matters. Traffic that had been counted as human was reclassified as bot. The comparison against 2024 became possible only after that correction. Part of what reads as "decline" is the removal of readers who were never readers.

How much of the 8% is correction and how much is real decline? Wikimedia does not decompose it, so neither can I, and neither can anyone citing it. "Part of it" is the honest phrasing. Any number attached to that split is invented. What is not in question is that Wikimedia found and published this itself, at cost to the tidiness of its own headline.

Wikimedia's attribution is also plural. It names search engines using generative AI to answer directly from Wikipedia content, and younger people going to social video platforms instead of reference sites. Two named forces, not one.

What did 30 citations do with it?

I collected 32 articles published between October 17, 2025 and August 15, 2026 that cite the Wikipedia decline figure, and fetched each one. Thirty resolved: one returned 404, one was a duplicate of another entry. Tally run August 15, 2026.

Twenty-eight state the 8%. On the details, the sample does better than I expected going in:

• 27 of 28 correctly specify human pageviews rather than traffic generally. That distinction is the one I assumed would evaporate first, and it mostly held.
• 21 of 28 mention the bot detection update in some form. Applying a stricter test, where the reclassification has to be tied to how the figure came to light rather than mentioned in passing, 18 of 28 carry it. TechCrunch, Engadget, Search Engine Journal, Decrypt, gHacks and Dataconomy all state it plainly. Seven never mention it in any form.
• 26 of 28 keep at least one non-AI cause in the body.

So the bodies are largely fine. The decay is somewhere else, and it is easy to locate.

Only 15 of 28 link to the Wikimedia Diff post. Slightly more than half. The rest cite a blog post they do not point at, which means the caveat is one unlinked hop away for every second reader, and the reader who wants to check has to go find it.

Then the headlines. Of the 22 news-style headlines in my sample that name a cause at all, 16 name AI alone. Six preserve Wikimedia's plural framing, and the ones that do are near-verbatim reuses of the source's own language: TechCrunch's "due to AI search summaries and social video," the AI Insider's "Amid Rise of Generative AI and Social Media Search." None of the 22 mentions the measurement correction. Not one.

That is the whole finding. The caveat survives the paragraph and dies in the headline, and the link that would let you check sits in barely half the articles. A reader who reads only headlines, which is most readers, gets a clean causal story that the source never told.

One more failure mode worth naming, because it is new. In my sample, an automatically generated summary box on a syndicated copy of the story rendered the finding as an 8% decrease "due to sophisticated bots disguising themselves as human visitors." That inverts the caveat completely: the bots inflated the earlier baseline, they did not cause the decline. The machines summarizing this material are already getting the causality backwards, which is a fitting problem for a story about machines summarizing material.

Is AI search actually killing Wikipedia?

Not on this evidence, and Wikimedia does not claim it is. What the primary supports is that human pageviews fell roughly 8% year over year in a specific window, that Wikimedia believes generative AI answers and social video both contributed, and that the number emerged from a data correction that removed bot traffic from the human count.

There is a real counterargument, and it comes from inside Wikipedia. A Wikipedia Signpost op-ed published June 21, 2026 argues that more than half of Wikipedia's organic search decline happened before ChatGPT reached the public in late 2022, which points at search algorithm changes and shifting attention rather than at AI answers. If that holds, AI is accelerating a slide that was already underway, which is a smaller and more interesting claim than the headlines make.

What would change my mind?

A decomposition from Wikimedia separating the reclassification effect from the underlying trend, or a longer series showing the 8% holding across windows that do not contain the bot correction. Wikimedia publishes pageview data openly, so this is answerable rather than rhetorical. If the decline persists at that scale into windows unaffected by the May 2025 correction, the strong version of the claim gets much stronger, and I will say so.

Verdict

Wikimedia measured carefully, published a number that made its own situation look worse, and volunteered the caveat that complicates it. That is exactly what you want from a source. What happened next is the failure: half the citations never linked back, three quarters of the headlines deleted every cause but AI, and not one headline mentioned that the figure came out of a bot reclassification. The same drift is visible in how the AI Overviews click studies get quoted, and it rhymes with the 95% pilot-failure stat, where the research was fine and the sentence carrying it was not. If you want your own content read by machines that summarize this badly, start with what they can actually fetch.

Verdict: 35% noise / 65% signal. The number is good. Cite the caveat that came with it.