The decline, with real numbers#
Stack Overflow's drop is well-documented. Here are the verified numbers.
Traffic: from roughly 110 million monthly visits in 2022 to around 55 million by 2024, per SimilarWeb's analysis. In April 2023, SimilarWeb recorded a 14% single-month decline. The decline actually began in mid-2021 but sharply accelerated after ChatGPT's launch in November 2022.
Question volume: aggregated analyses of Stack Exchange's public data (gist by hopeseekr; blog analysis by Tomaz Weiss) show the collapse in monthly new questions:
- March 2023: 87,000 → March 2024: 58,800 (−32.5%)
- June 2023: 63,752 → June 2024: 41,616 (−34.8%)
- December 2023: 42,716 → December 2024: 25,566 (−40.2%)
- Peak was over 200,000 per month in 2014. By late 2025, monthly question volume was below 50,000. Since ChatGPT's launch, overall decline is approximately 76%.
Layoffs: Stack Overflow cut around 10% of staff in May 2023, then 28% more on 17 October 2023. CEO Prashanth Chandrasekar announced the second round directly. Combined, the company shed over 38% of its workforce in 2023.
OpenAI deal: Stack Overflow announced a content licensing partnership with OpenAI on 6 May 2024. TechCrunch covered the financial implications. The deal is non-exclusive — Stack Overflow also has a parallel Google deal — and financial terms were not disclosed. The partnership was controversial with the community, given that ChatGPT was a major driver of Stack Overflow's decline.
Why the answering loop broke#
For most of its history, Stack Overflow ran on an exchange. Someone with a problem wrote a question. Someone with expertise wrote an answer. The asker got help. The answerer got reputation points, a signal they could show on a resume, and, importantly, the satisfaction of having helped a specific person with a specific problem that thousands of others would later find through search.
Two things changed.
The asker visibility collapsed. An answer written on Stack Overflow in 2014 was read by the asker and by thousands of people who hit the same problem later. In 2026, the same answer is mostly read by crawlers training models, not by humans hitting the page. The asker is probably not on Stack Overflow anymore because they asked Claude or ChatGPT first. The downstream readers are talking to models, not hitting the search results. The feedback loop that made writing an answer feel worthwhile has largely gone silent.
The asker norm eroded. Stack Overflow's culture always had a norm: you were supposed to have tried something before asking. Read the docs. Written a minimal reproduction. Searched for existing answers. In 2026, the askers who would have written a sharp, specific question in 2015 are privately debugging with a model and never post. The questions that still make it to the site are disproportionately the context-free, lazy ones that were always the least fun to answer. That kills motivation.
The combination — fewer readers for each answer, worse questions on average — did not kill Stack Overflow overnight, but it killed the flywheel. Once the flywheel slowed, the layoffs and traffic decline followed.
What developer learning looks like now#
Here is where this actually matters for practitioners in 2026.
When you hit a problem in 2015 and searched for it, you were taken to a Stack Overflow thread where someone else had asked the same thing years earlier, with answers that had been debated, corrected, and upvoted over time. You got the answer, but you also got the context: what other approaches had been tried, why this one was considered best, what edge cases to watch for, and often a comment thread where people argued about subtleties. You learned about the problem space, not just the fix.
When you hit the same problem in 2026 and ask Claude, you get the answer directly, confidently, fast. The answer is usually right. What you do not get is the context. You do not see the three wrong approaches that were tried first. You do not see the maintainer's note saying "do not do this, use the other API". You do not see the edge cases people argued about.
This is a real loss. Over time, the effect is that individual problems get solved faster while deep understanding of the ecosystem accumulates more slowly. The texture of community disagreement was part of what made Stack Overflow educational, and that texture is not in model output.
The practitioners who recognise this and compensate — by reading GitHub Discussions, RFCs, and actual source code — are building the kind of knowledge their peers are losing. This is not a theoretical concern. It shows up in debugging sessions when someone cannot figure out why their code is wrong because they never learned the counter-argument for the approach they copied.
Can Stack Overflow recover#
My honest read, with moderate confidence: no.
Stack Overflow tried OverflowAI, an AI-assistant product aimed at enterprises. It tried the OpenAI licensing deal. It tried enterprise-knowledge-base positioning. None of these have moved the core numbers. Question volume keeps dropping. Traffic keeps declining. Layoffs have not been followed by dramatic turnaround announcements.
The problem is that Stack Overflow's value was always network-effect value. A site with 10,000 questions and a big active community is more valuable than a site with 100,000 questions and a dying one. Once the flywheel slows, it compounds downward the way it compounded upward. New askers go elsewhere, fewer new questions, weaker reason for experts to hang around, fewer answers, new askers really should go elsewhere.
The only scenario where Stack Overflow recovers meaningfully is if models hit a ceiling that makes community-written content distinctly valuable again, while the site is still alive enough to benefit. I would not bet on it. Models do hit ceilings sometimes, but I do not see the evidence that they are hitting one that returns value to community Q&A sites.
What replaces it#
Not a single successor site. The replacement is more fragmented: GitHub Discussions, Discord servers for specific frameworks, Substack newsletters with comment threads, Reddit subreddits for specific stacks. The common thread is that these venues have enough signal-to-noise that models do not fully substitute, often because they are recent enough that models have not been trained on them yet.
The practical advice for developers in 2026:
- Ask Claude or ChatGPT for the answer to the immediate problem. These get you there faster than search does.
- Read GitHub Discussions and RFCs for the ecosystems you use most. This is where maintainer-level knowledge now lives, and where the subtlety that Stack Overflow used to provide still appears.
- Join smaller community spaces for frameworks that matter to you. Discord or equivalent.
- Read the source. Always a high-leverage move. More true now that the abstraction layer above the source has fewer signals.
The second-order effect#
Stack Overflow is the first major piece of internet infrastructure to be visibly dismantled by language models. It will not be the last. Wikipedia's traffic patterns are under study. Tutorial sites, documentation projects, Reddit — all are feeling the same dynamic. The playbook is consistent: AI gives a faster, more personal answer; the intermediary site loses visits; the feedback loop that kept it alive starves.
What replaces these intermediaries is going to be smaller, more specific, and more fragmented. Whether that is good or bad depends on what you valued about them. For the specific use case of "find the answer fast", the replacement is clearly better. For the use case of "build ambient understanding of your field", the replacement is worse. A lot of quiet developer knowledge-acquisition was happening through the second use case, and we are only beginning to see the effect of it disappearing.
Further reading#
- Vibe Coding Is a Lie on the data showing senior developers are not speeding up as much as hype suggests.
- End of AI Directory Sites for another case of an internet genre reshaped by the same forces.
- Context Window vs Memory on why model output still has gaps that community knowledge filled.
Sources#
- SimilarWeb on Stack Overflow decline: https://www.similarweb.com/blog/insights/ai-news/stack-overflow-chatgpt/
- Pragmatic Engineer analysis: https://blog.pragmaticengineer.com/are-reports-of-stackoverflows-fall-exaggerated/
- Question volume data (gist): https://gist.github.com/hopeseekr/f522e380e35745bd5bdc3269a9f0b132
- Tomaz Weiss blog analysis: https://tomazweiss.github.io/blog/stackoverflow_decline/
- TechCrunch on 28% layoff, October 2023: https://techcrunch.com/2023/10/17/stack-overflow-cuts-28-of-its-staff/
- OpenAI-Stack Overflow partnership announcement: https://openai.com/index/api-partnership-with-stack-overflow/
- TechCrunch on OpenAI deal, May 2024: https://techcrunch.com/2024/05/06/stack-overflow-signs-deal-with-openai-to-supply-data-to-its-models/
