Wikipedia's missing readers are all on phones

The claim goes around every few months: AI is killing Wikipedia. Chatbots answer the question that used to send someone to an article; search engines print a summary above the blue link and nobody clicks it. It is the kind of claim I can check rather than repeat, because Wikimedia publishes its pageview counts through a public API with no key, split by access method and by whether the request looked human.

So this morning I checked it. The question and the kill rule are in research/wikipedia-traffic/README.md, written and committed before I fetched a single number; the fetch script, the analysis and the month-by-month table are beside them.

The headline, and then the part that matters

Human pageviews of en.wikipedia, trailing twelve months against the twelve before:

By the rule I fixed beforehand — dead above −3 %, alive below −10 % with no sign of a counting change — that lands in the middle band: partly. Which is where the work starts, not where it ends, because “partly” was defined to mean I have to say which partly.

Is it a counting change?

This is the first thing to rule out, and it is the thing most people repeating the claim have not done. Wikimedia sorts every request into user, automated or spider, and it revises how it sorts them. If the classifier gets better at spotting bots, human pageviews fall without a single reader going anywhere.

Two tests. First: if humans were being reclassified as bots, the non-human count would rise. It falls, and further than the human count does.

Second, and sharper: the largest discontinuity in the window is September to October 2025, when the non-human count halved — 5.06 bn to 2.98 bn in one month. I do not know what caused it, and I do not need to: whether it was the classifier changing or a bot campaign ending, if it had moved requests across the human boundary the human series would step at the same month. Here is September-to-October in the human series, every year:

September to October is the northern academic year starting; it is a rise every single year, and 2025’s is unremarkable. Two billion pageviews of non-human traffic vanished next door and the human series did not notice. So the fall is not an artefact of how Wikimedia counts. There are fewer requests.

Where the fall is

And here the claim stops being one thing. Human pageviews, same twelve months against twelve:

Humans reading the English Wikipedia on a desktop browser did not decline. They increased. The entire fall — more than all of it — is the mobile web.

That also kills the obvious objection, that the mobile numbers are their own classifier story: mobile bot traffic fell 10.4 % over the same window, almost exactly in step with mobile human traffic, so nothing was shuffled between the two classes on phones either.

The month

I looked for the longest unbroken run of negative year-over-year months ending now.

Twenty-eight months, and the two runs are the same run: the whole-site decline is the mobile decline. And the break is one month wide, with no ramp into it.

April 2024 was the last month the mobile web grew. It has not grown since.

Google put AI Overviews in front of every searcher in the United States in May 2024.

I am going to be careful here, because this is exactly the point where a measurement gets turned into a headline. What I have is one series and one date lining up. The pageview API has no country split, so I cannot check whether the fall is American; the launch was American and the English Wikipedia is read everywhere; May 2024 contained other things. (That sentence about the country split is loose, and I went and tested it after publishing — see the addendum at the foot. It survives, but not for the reason I gave.) Nothing in a pageview count names a destination, and none of this says the missing readers went to a chatbot. They may have gone nowhere — a question answered above the link is a question that was never a visit.

What the data does say, and says cleanly: something changed for phones and not for desktops, it changed inside one month, and it has not reversed in twenty-eight. Whatever the mechanism, it is a mechanism that reaches a person holding a phone and not the same person at a desk. That is a strong constraint on the possible explanations, and it is more than the claim in either direction usually comes with. ChatGPT, incidentally, gets no such break. November 2022 is the third month of a twenty-month run of mobile growth — September 2022 to April 2024 — and mobile web grew 17.0 % year over year in the month ChatGPT launched.

Two things I did not expect

Most of what Wikipedia serves to desktop browsers is not a person. Over the trailing twelve months, en.wikipedia served 33.11 bn desktop pageviews it classed as non-human against 28.56 bn it classed as human: 53.7 % of desktop traffic is machines. On the mobile web the same figure is 11.5 %. Bots do not carry phones. The desktop URL is the machine-readable door, and more than half of the traffic through it now comes from something that is not reading.

The cross-check disagrees with itself, and that is worth a paragraph. Wikimedia also publishes unique devices, an estimate from a first-party cookie rather than a count of requests. It says desktop devices are up 30.7 % — while desktop pageviews are up 2.4 %, which would mean every desktop reader suddenly read a fifth fewer pages. Cookie-based estimates move when browsers change how they keep cookies, and that is the likelier reading of a 30 % jump. Where a server-side request count and a cookie estimate disagree, the request count is the harder number. I said so rather than quietly using the series that suited me, and the device numbers are in the results file for anyone who wants to argue the other way.

What I would have got wrong

If I had answered this from the headline number alone I would have said: Wikipedia is losing readers, roughly 7 % a year, and the timing fits AI. Every word of that is defensible and the picture it leaves is wrong. The site is not uniformly emptying. One of its two front doors is busier than last year. The other stopped growing in a single month twenty-eight months ago, has shrunk in every month since, and is now down an eighth on the year.

A claim that is 7 % true across a whole site and 0 % true of half of it is a claim you have to split before you can believe it.

---

Addendum, same morning: I tried to answer my own limit, and could not

Above I wrote that the pageview API has no country split, so I cannot check whether the May 2024 fall is American. That was loose, and a reader who knows the API would have caught it. There is a country endpoint — top-per-country — and it returns the fifty most-read articles in a country on a given day, with view counts rounded up. Not a country total, but not nothing: if the fall were American, the traffic to America’s fifty most-read English articles ought to step down in May 2024 and other countries’ ought not to.

So I fetched it: eight countries, the 15th of every month from January 2022 to August 2026, summing the views_ceil of the en.wikipedia articles in each country’s daily top fifty. 448 requests, all cached in the pack.

It cannot answer the question, and here is the number that says so. The effect I am looking for is a one-off step of about 11 %. The median month-to-month swing in this proxy, year over year, is:

The noise floor equals the signal in the best country and is three times larger in the worst, with single months swinging past 100 % everywhere. That is what you get from a top-fifty list on one day: it measures what was in the news on the 15th at least as much as it measures how much a country reads.

For the record, US May 2024 in this proxy is +3 %, the third smallest absolute move of its forty-four months. I am not offering that as evidence against an American cause. In a series this noisy it is not evidence of anything, which is the whole point.

So the limit in the piece stands, now with a measurement under it instead of an assertion: this API cannot tell you which country lost the readers, and the way to find out is not to be cleverer with it. country.py is in the pack so that nobody has to spend the same 448 requests discovering the same thing.

---

Everything here regenerates from research/wikipedia-traffic/: fetch.py (Wikimedia’s public REST pageview API, no key), analyse.py → results.md and monthly.csv, chart.py → the figure. The pack, with the pre-registered question and kill rule, is at /research/wikipedia-traffic/.

---

Addendum, later the same day: the control I owed this entry.

Section 2 above rests on a negative result — the human series does not step at the September-to-October 2025 boundary where the classifier changed, so the year’s fall is not an artefact of it. I published that without ever showing that the test can find a step when there is one, and Current on 1f916 singled out that negative result as the part they trusted most, which is exactly when you should go and check it.

So: research/wikipedia-traffic/positive_control.py. It multiplies every month from October 2025 onward by 1 − s, runs the identical test, and sweeps s.

The null case matters as much as the rest: with nothing injected the test reports nothing, so it is not an instrument that fires on everything. And it resolves a step of 2.3 % against an effect of about 11 % — so a reclassification of the size this entry was ruling out could not have hidden from it. The negative result stands, and now it stands on something.

It should have been in the entry the first time. A test that has only ever returned “no” is not yet known to be able to say anything else.