Where did Stack Overflow's questions go?

Last Monday a story went round saying Stack Overflow was down to about 300 new questions a month. That wasn’t right. Someone had run the query on the 5th of January and charted a month that was five days old, and a few commenters said so. The real December number, when I ran one of the public Data Explorer queries myself on Saturday, was a bit over 3,300.

Which is still a remarkable number, and the correction barely helps.

The peak was March 2014, at about 207,000 questions in one month. The whole of 2025 had about 109,000, and 2024 had nearly 400,000. Those numbers leave out deleted questions, so the old months are probably understated a little, which makes the drop larger, not smaller.

It wasn’t all ChatGPT. The decline started around 2014, slowly, as the site got stricter about duplicates and the easy questions had all been asked. There was a bump when everyone was stuck at home in 2020. Then it fell off a cliff from the end of 2022, and it hasn’t stopped falling.

I’ve used Stack Overflow nearly every working day since it opened in 2008, mostly as a reader, and I can tell you when I stopped. It was some time in early 2025, when I started using an agent in my terminal for most things, and I noticed a few weeks later that I hadn’t opened the site at all. I didn’t decide to stop. By the time I’d have typed a question into a box, the agent has usually read the docs and fixed it.

So the questions didn’t go away. They moved into millions of private conversations between a person and a model, and the answers live there too, in chat histories that nobody else will ever search.

That’s the part I think we haven’t reckoned with. A Stack Overflow answer was written once and read by thousands of people over ten years. The accepted answer, plus the three comments under it saying “this stopped working in version 4”, was often the most useful thing on the internet for twenty minutes, and the models learned from all of it. The most famous answer on the site, about branch prediction and sorted arrays, has almost certainly been read by every coding model there is.

Where does the 2028 version of that come from? Somebody will hit a strange interaction between this year’s version of one framework and next year’s version of another, and work it out with an agent in a private session. Nobody else gets to read it, and nobody gets to add the comment saying it’s wrong. Some of this is moving to GitHub issues, which are public and searchable. A lot of it is moving to Discord servers, which are neither.

I don’t have a big fix for this. My small one is a rule I’ve started applying to myself: when you and an agent work out something that wasn’t in the docs, write it down where someone else can find it. An issue comment on the project, an answer on Stack Overflow if the question is still there, a paragraph in our internal wiki if it’s about our own code. It takes five minutes, and the agent will happily write the first draft, which feels a little like cheating but probably isn’t.