For fifteen years, developers brought Stack Overflow their questions — at the peak, more than 200,000 a month. Four years after ChatGPT, they bring 1,624. This is what it looks like when AI quietly drains the well that helped train it.
127 monthly data points · Jan 2016 → Jul 2026 · new questions by creation date · official Stack Exchange API · every figure recomputable
From 2016 to 2022, Stack Overflow's monthly question volume drifted down about 8% a year as search shifted and the easy questions ran out — steady, unremarkable. Then ChatGPT shipped in November 2022, and the gentle fade became a freefall. Below is every calendar month since 2016; the inflection needs no interpretation.
An 8%-a-year drift does not compound into a 98% collapse in four years. A step-change does. Year over year, the decline didn't merely continue — it accelerated.
| Year | Questions | Year over year |
|---|---|---|
| 2016 (peak) | 2,183,926 | — |
| 2022 · ChatGPT launches | 1,336,266 | −13% |
| 2023 | 788,512 | −41% |
| 2024 | 398,649 | −49% |
| 2025 | 111,098 | −72% |
| 2026 · annualized | ~32,000 | −71% |
A site that once took two million questions a year now takes fewer than one a minute.
Here is the part that should unsettle anyone who builds on the open web. The archive AI is emptying is also barricading itself against AI. In 2024, Stack Overflow restricted its long-standing public data dump — the very corpus that helped train the models now replacing it — moving it behind a login and, briefly, watermarking it. The site is going quiet and pulling up the drawbridge at once. The human-authored knowledge that made the last decade of software possible isn't only shrinking; it is being withdrawn.
They're now typed into ChatGPT, Claude, and Perplexity — which answer from models trained on the very archive going quiet. For anyone who publishes to be found, three things follow.
1 · Assume discovery runs through AI answers, not blue links. Write to be quoted by a model, not merely ranked by a crawler.
2 · Structure beats volume. Pages an engine can parse cleanly — clear headings, FAQs, tables, cited sources — get pulled into answers; walls of prose don't.
3 · Measure continuously. What AI cites shifts week to week. A snapshot is a photograph; you need the film.
| Metric | Value |
|---|---|
| Peak monthly questions (Mar 2017) | 201,429 |
| Questions, Nov 2022 (ChatGPT launch) | 109,301 |
| Questions, Jul 2026 (latest) | 1,624 |
| Change since ChatGPT | −98.5% |
| Change vs all-time peak | −99.2% |
| 2023 / 2024 / 2025 year-over-year | −41% / −49% / −72% |
| Months measured | 127 (Jan 2016 – Jul 2026) |
| Source | Stack Exchange API |
Is Stack Overflow actually dying, or just changing? By its core metric — people asking questions — it has lost 98.5% of its volume in four years, and traffic and answers track the same curve. Independent analyses (The Pragmatic Engineer; Slashdot) report the same near-total decline.
Couldn't this be Google updates or spam filters, not AI? Those pressures ran throughout the gentle 2016–2022 slide. The 40–70%-a-year freefall begins precisely at ChatGPT's launch — a step-change, not a trend.
Can I reproduce these numbers? Yes — one query per month against the public Stack Exchange API. That's the whole point.
The same shift that emptied Stack Overflow decides whether AI answers cite you. Run a free GEO audit and see where you stand.
The lesson isn't 'don't use AI' — it's to use it the way we describe in our guide to AI for Webflow content: grounded in sources, not generated by the thousand.
Run the free audit