Journal
Wikipedia Pageviews as a Leading Indicator of Public Interest in AI
Wikipedia pageviews are a free, high-frequency, publicly auditable measure of what people actually care about. Here is how to read them — and what five years of the 'Artificial Intelligence' article show.
Surveys ask people what they think. Wikipedia pageviews show what people actually do — millions of times a day, with timestamps, in the open. When the “Artificial Intelligence” article spikes, it is usually because something real happened in the AI world; when it trends upward for months, that is public interest compounding.
This article explains how Wikipedia pageviews work as a data source, how to use them as an interest indicator, and what they say about AI’s trajectory. The monthly series is tracked in our Wikipedia Pageviews: Artificial Intelligence dataset.
Why pageviews are a genuinely good dataset
Almost every “public interest” metric is flawed in ways you cannot audit. Pageviews has three properties that make it unusually trustworthy:
- It is a census, not a sample. Every view that hits Wikipedia’s REST API counters is counted. No panel, no weighting, no margins of error.
- It is auditable. Anyone can reproduce the numbers for any article, any date range, any language — the raw counters are public.
- It is high frequency. Daily counts (monthly in our dataset) react to news cycles, product launches and cultural moments in near-real time.
Compared to Google Trends, pageviews win on reproducibility (Trends refuses to publish absolute numbers) and lose on granularity (Trends covers broader query behavior across the web, not just Wikipedia).
How the numbers are produced
The pageviews pipeline (the one behind our Wikipedia-ai series) is:
- Source: Wikimedia REST API v1 (
pageviews/metrics), user agents only — bot traffic is excluded at the source. - Granularity: monthly sums per calendar month; every access type (desktop, mobile web, app) counted.
- Definition: views are not deduplicated per visitor, and one human opening the article twice counts twice.
That last point matters: pageviews measure interest events, not unique people. A viral moment can be driven by the same audience refreshing repeatedly. For trend work this is usually fine — but it is the first caveat to keep in mind.
Reading the AI interest curve
Look at the chart and you can reconstruct the modern AI era from pageviews alone:
- Late 2022: ChatGPT launches. Monthly pageviews of the “Artificial Intelligence” article multiply several times over in a single month — the clearest single spike in the series.
- 2023–2024: the spike cools but never returns to pre-ChatGPT levels. The plateau sits far above the old baseline: a permanent step-change in public interest.
- 2025–2026: a slow drift upward, with small bumps around major model releases and industry events.
That shape — a step-change followed by higher plateau — is the classic signature of a new mainstream technology entering the culture: it is exactly how you would draw “the internet, 1994” or “the smartphone, 2008.”
How to use pageviews as an indicator
Pageviews work best as a leading, sentiment-neutral complement to other signals:
| Use case | How pageviews help |
|---|---|
| Product timing | A sustained rise in a topic’s interest suggests demand for tools/news on that topic |
| Event verification | Spikes tell you what actually captured attention versus what the press thought would |
| Baseline comparison | Track interest before/after a launch to measure its cultural footprint |
| Cross-language | Compare the same article across language editions to find geographic divergence |
The main weakness to respect: volume tells you how many eyeballs, not how positive. A spike caused by bad news looks identical to a spike caused by good news. Pair pageviews with sentiment signals when you need the valence, not just the volume.
More public-interest signals on the matrix
Interest is a domain-spanning signal. The same “engine” (hard-data time series from public sources) powers several cells of the matrix:
- Wikipedia Pageviews: Artificial Intelligence — public interest in AI
- Framework Weekly Downloads — developer adoption of JS frameworks
- U.S. Inflation Rate (CPI, year-over-year) — economic attention, monthly
And if you want the methodology behind finding and vetting sources like these:
- How to Find Reliable Data Online: Sources, Methodology and Red Flags
- LLM Benchmarks Explained: MMLU, HumanEval and How to Read Them
FAQ
Are Wikipedia pageviews adjusted for bots? The Wikimedia API excludes known bots and crawlers at the source for user-agent counts. Sloppy scraping of other sites’ “views” is far less clean — another reason to prefer the official API.
Why use the English article, not a broader query set? The English edition is the largest and most stable, and the “Artificial Intelligence” article acts as a hub — most people landing on AI content from search or news pass through it. Cross-language comparisons are possible and remain part of the methodology notes.
Can pageviews predict anything? They are a coincident-to-leading indicator of attention. They cannot tell you what people think, but they are often the earliest and cleanest public signal that a topic has crossed from niche to mainstream.