# Results — is AI killing Wikipedia's traffic?

Generated by `analyse.py` from Wikimedia's public REST pageview API. Complete calendar months; the last is **2026-08**. The question and the kill rule are in README.md and were fixed before the data was fetched.

## 1. The kill rule's answer

| | human pageviews |
|---|---|
| trailing 12 months (2025-09..2026-08) | **81.08 bn** |
| the same 12 months a year earlier | 87.17 bn |
| change | **-7.0 %** |

The pre-registered reclassification detector fires on **4 month(s)** of the window:
  - `user` fell 196 m into 2025-09, non-human rose 538 m into 2025-08
  - `user` fell 228 m into 2025-11, non-human rose 385 m into 2025-11
  - `user` fell 102 m into 2025-12, non-human rose 385 m into 2025-11
  - `user` fell 239 m into 2026-06, non-human rose 412 m into 2026-06

**Verdict by the rule fixed before the data: PARTLY** — the fall is inside the −3 % to −10 % band. Section 2 decides which kind of "partly" it is.

## 2. Is it a counting change?

The detector above compares month-on-month moves and so fires on ordinary seasonality. Two harder tests:

**(a) If humans were being reclassified as bots, the non-human series would rise.** It falls:

| series | trailing 12 (bn) | previous 12 (bn) | change |
|---|---|---|---|
| human (`user`) | 81.08 | 87.17 | -7.0 % |
| non-human (`automated`+`spider`) | 39.67 | 45.08 | -12.0 % |
| everything counted | 120.76 | 132.25 | -8.7 % |

Total counted traffic fell **-8.7 %**, more than the human part. Traffic did not move between the classes; there is less of both.

**(b) The largest discontinuity in the window is 2025-09→2025-10, when the non-human count halved (5.06 bn → 2.98 bn). Whatever caused it — the classifier changing, a bot campaign ending — if it had moved requests into or out of the human class, the human series would step at the same month.** It does not — September to October is a seasonal rise every year, and 2025's is inside the usual range:

| year | Sep → Oct, human pageviews |
|---|---|
| 2016 | +4.6 % |
| 2017 | +7.1 % |
| 2018 | +4.1 % |
| 2019 | +7.9 % |
| 2020 | +8.7 % |
| 2021 | +1.5 % |
| 2022 | +1.1 % |
| 2023 | +6.5 % |
| 2024 | +3.8 % |
| 2025 | +3.4 % |

So: the fall in human pageviews is not an artefact of how Wikimedia sorts requests. It is fewer requests.

## 3. Where the fall is: one access method, not all of them

| access method | human, trailing 12 (bn) | previous 12 (bn) | change |
|---|---|---|---|
| desktop | 28.56 | 27.90 | **+2.4 %** |
| mobile-web | 50.43 | 57.04 | **-11.6 %** |
| mobile-app | 2.09 | 2.23 | **-6.2 %** |

Desktop reading of en.wikipedia by humans **grew**. The whole of the decline is the mobile web, plus a smaller fall in the app.

The same split for the bots, which rules out a mobile-only classifier change as the explanation — mobile bot traffic falls by about as much as mobile human traffic, so nothing was moved between the two:

| access method | non-human, trailing 12 (bn) | previous 12 (bn) | change |
|---|---|---|---|
| desktop | 33.11 | 37.77 | -12.3 % |
| mobile-web | 6.54 | 7.30 | -10.4 % |

Worth stopping on: on **desktop**, 33.11 bn of the last twelve months' pageviews were declared non-human against 28.56 bn human — **53.7 % of what en.wikipedia served to desktop URLs was not a person.** On the mobile web it is 11.5 %.

## 4. When it started

| series | unbroken run of negative year-over-year months ending 2026-08 | starts |
|---|---|---|
| all human pageviews | 28 months | 2024-05 |
| mobile web, human | 28 months | 2024-05 |
| desktop, human | 1 months | 2026-08 |
| mobile app, human | 16 months | 2025-05 |

The break is one month wide. Mobile web, human pageviews, year over year:

| month | mobile web | desktop |
|---|---|---|
| 2024-02 | +12.1 % | +4.3 % |
| 2024-03 | +6.2 % | -0.8 % |
| 2024-04 | +7.9 % | +2.6 % |
| 2024-05 | -3.1 % | +0.9 % |
| 2024-06 | -5.1 % | -1.9 % |
| 2024-07 | -6.0 % | +2.4 % |

April 2024 was the last month mobile web grew. It has not grown since. Desktop shows no such break.

Google put AI Overviews in front of every searcher in the United States in May 2024. That is an alignment of one series with one date, and this data cannot turn it into a cause: the pageview API has no country split, the launch was American and en.wikipedia is read everywhere, and other things happened that spring. What the data does say is that whatever changed, changed for phones and not for desktops, in a single month, and has not reversed in 28 months.

## 5. The cross-check that disagrees, and why I trust it less

Wikimedia also publishes **unique devices**, which is not a count of requests but an estimate from a first-party cookie:

| series | trailing 12 (bn) | previous 12 (bn) | change |
|---|---|---|---|
| all-sites | 10.13 | 11.95 | -15.3 % |
| desktop-site | 3.62 | 2.77 | +30.7 % |
| mobile-site | 6.50 | 9.18 | -29.2 % |

Desktop **devices** are up +30.7 % while desktop **pageviews** are up only +2.4 %, which would mean every desktop reader suddenly read a fifth fewer pages. A cookie-based estimate moves when browsers change how they keep cookies, and that is the likelier reading. Where a server-side request count and a cookie estimate disagree, the request count is the harder number, and it is the one section 1 uses.

## 6. Context: the longer series

| year | human pageviews (bn) | change |
|---|---|---|
| 2016 | 93.07 |  |
| 2017 | 91.41 | -1.8 % |
| 2018 | 92.04 | +0.7 % |
| 2019 | 92.15 | +0.1 % |
| 2020 | 94.31 | +2.4 % |
| 2021 | 88.96 | -5.7 % |
| 2022 | 86.59 | -2.7 % |
| 2023 | 91.50 | +5.7 % |
| 2024 | 91.53 | +0.0 % |
| 2025 | 85.54 | -6.5 % |
| 2026 (2026-01..2026-08) | 53.27 | -7.7 % vs the same months a year earlier |

The series has fallen before — 2021 and 2022 both lost ground and then 2023 took it back. What is new is the length: 28 consecutive negative months.

All Wikimedia projects together, human pageviews: 164.00 bn in the trailing twelve months against 179.46 bn, **-8.6 %**.

## 7. What this does not say

Where the readers went. Nothing in a pageview count names a destination, and no number here should be read as "they asked a chatbot instead". The findings are: the fall is real and not a counting artefact; it is a phone fall; it began in May 2024 and has not stopped.

## 8. The country proxy, and why it fails

Section 4 says this API has no country split. That is loose: `top-per-country` exists, returning the fifty most-read articles in a country on a day with `views_ceil` rounded up. `country.py` sums the en.wikipedia articles in that top fifty, the 15th of each month, 2022-01..2026-08, for eight countries (448 requests, cached).

Median absolute year-over-year change per country:

| US | GB | IN | CA | AU | DE | FR | JP |
|---|---|---|---|---|---|---|---|
| 10.5 % | 12.5 % | 15.3 % | 21.4 % | 26.6 % | 19.2 % | 23.5 % | 33.1 % |

The effect sought is a one-off step of about 11 %. The noise floor equals it in the quietest country (US, 10.5 %) and more than triples it in the loudest (JP, 33.1 %); single months exceed 100 % in every country. US 2024-05 reads +3 %, the third smallest absolute move of its 44 months — **not** evidence against an American cause, because a series this noisy is not evidence of anything.

**Verdict: the proxy cannot resolve the question.** The limit stated in section 4 stands, now measured rather than asserted. Kept in the pack so nobody repeats it.