There’s a graveyard somewhere on the internet for the websites we used to love. GeoCities is buried there. Digg. The original Twitter. Most of Yahoo. And for the last two years, every other tech writer has been arguing that Stack Overflow belongs in the same hole.
The numbers, on the surface, agree. Monthly question volume on Stack Overflow peaked at over 200,000 in 2014. By December 2024 it was 25,566 — a 76.5 percent drop from the month ChatGPT launched, and the lowest monthly count since May 2009, ten months after the site went live. Question volume has been below 50,000 a month for over a year. Traffic, depending on whose numbers you trust, is down somewhere between 35 and 60 percent off its peak. A site that once handled the volume of every Fortune 500 engineering team’s daily problem-solving now gets fewer questions per month than it did in its first full year of existence.
You’d be forgiven for assuming the obituary is overdue.
It isn’t. And the people who keep writing it have been ignoring something genuinely interesting that’s happened over the last twelve months: Stack Overflow has stopped pretending the world hasn’t changed, and started shipping like it’s 2009 again.
A site we forgot we needed
If you’re under thirty, you have to be told this. If you’re over thirty and write code for a living, you remember.
There was a moment — call it 2010 to about 2020 — when Stack Overflow was the always-open tab. You didn’t bookmark it because you didn’t need to. It was where you went when something broke, when a stack trace didn’t make sense, when you needed the regex you couldn’t quite remember, when you wanted to know if anyone else had seen this weird ConcurrentModificationException at 3 AM on a Tuesday. And nine times out of ten, someone had. The accepted answer was usually two years old, written by someone with a profile picture of a cat, and it just worked.
That wasn’t an accident. Stack Overflow was a designed solution to a hard problem: how do you let strangers help strangers without it devolving into a flame war? They invented the answer — reputation systems, structured Q&A, ruthless duplicate-closing, voting that surfaced the best response over the loudest one — and then exported that pattern to dozens of sister sites covering everything from cooking to academia. For roughly a decade, Stack Overflow was the closest thing software development had to a public utility.
Then ChatGPT shipped, and the utility started getting bypassed.
The quiet, awkward truth about what killed it
ChatGPT didn’t kill Stack Overflow. ChatGPT just made it embarrassing to need Stack Overflow.
You can argue about the data — and people have. The collapse in question volume is uncontested. But Stack Overflow’s traffic decline started in mid-2021, eighteen months before ChatGPT existed. The site had already alienated a chunk of its community with high-profile moderation disputes and the 2023 moderator strike. The new-question pipeline had been slowing for years, partly because the back catalog grew so large that almost any common question was a duplicate of something asked in 2014.
What ChatGPT did was remove the social cost of asking. There is no Stack Overflow user who hasn’t had a question closed as off-topic, or needs-more-focus, or duplicate-of-some-link-that-isn’t-actually-a-duplicate. There is no Stack Overflow user who hasn’t been condescended to. ChatGPT will never tell you to read the FAQ. It will never close your question. It will never reply with “this is a duplicate of [unrelated question from 2011].” It is, for better and worse, infinitely patient with bad questions.
That’s the deeper wound. The traffic collapse is a symptom. The disease is that a generation of developers learned they could solve their problem without ever having to ask another human for help. The feedback loop that built the site’s value — questioner, answerer, voter, future Googler — broke quietly. The questions stopped flowing in. The answers stopped getting written. The training data the LLMs depended on stopped getting created. The well that everybody was drinking from started running dry, and the people drinking didn’t notice because the water kept arriving in a different container.
That’s where it gets uncomfortable. Because if you’ve used a coding assistant in the last two years — if you’ve shipped a PR that included a snippet you got from Claude or Copilot or ChatGPT — you should know whose work you were almost certainly building on.
The OSS legacy nobody talks about
While everyone was writing eulogies, somebody pulled up the NuGet stats.
Dapper — the micro-ORM Stack Overflow built because Entity Framework wasn’t fast enough for their own site — has been downloaded over 350 million times. It is, in real terms, one of the most-deployed pieces of data-access code in the entire .NET ecosystem. Every time you’ve used connection.Query<T>(...) in a C# project, you were running something Stack Overflow gave away free.
StackExchange.Redis — the Redis client they built because the existing options didn’t handle production load — sits at over 63 million NuGet downloads. If you’ve written a .NET app that talks to Redis in the last decade, there’s a better-than-even chance you ran their code. It is the de facto standard.
MiniProfiler — the lightweight profiler that gave a generation of .NET web apps that little query-timing badge in the corner of every page — runs in production at companies that have never heard of Stack Overflow’s product team. Opserver, the dashboard they used internally to monitor SQL Server, Redis, Elastic, HAProxy, and exception logs, is open source. So is Bosun, their monitoring and alerting system. So is DNSControl, which now manages DNS for parts of the internet most engineers don’t think about. So is BlackBox, the Git secrets manager. So is Jil, a JSON serializer that for years was the fastest thing in the .NET world, and Sigil, the IL-generation library underneath it. So is the Stack Exchange Data Explorer, which lets anyone with a SQL brain query the entire community Q&A dataset directly. So is MarkdownSharp, which powered the original Markdown rendering pipeline used by half the early .NET web.
This is not a small footprint. It is, by any honest accounting, one of the most consequential open-source contributions any single web property has ever made — measured not in stars on a README, but in production deployments.
The irony writes itself. Stack Overflow, for fifteen years, gave the .NET ecosystem the tools it used to build the web. Then Stack Overflow gave the LLM industry the training data it used to replace Stack Overflow. The community did the work, the platform hosted it under Creative Commons, and somewhere between 2022 and 2024 a whole generation of trillion-parameter models learned to code by reading those answers — without anyone ever paying the people who wrote them, and without most users ever realizing where the answers came from.
That’s the bill we haven’t paid yet. And Stack Overflow, for better or worse, is the only entity left that’s still trying to send it.
The reinvention nobody noticed
This is the part of the story that’s actually news, and the part nobody is writing about because “Stack Overflow is dead” is a more clickable headline than “Stack Overflow is shipping again.”
In December 2025, Stack Overflow took AI Assist out of beta and made it generally available. It is not what you’d expect from a forum site that spent 2022 banning AI-generated answers. AI Assist is a conversational search interface that prioritizes community-verified answers first, then falls back to LLMs only when the existing community knowledge doesn’t cover the question. Sources are surfaced inline with citations and direct quotes. Original contributors get credit. If the knowledge base has a gap, the interface routes you toward asking the question — putting questions back into the system instead of letting them die in a private LLM chat.
It is the right architecture. It is the only architecture that makes long-term sense. An LLM-only answer is fast and frictionless and wrong often enough to matter. A community-only answer is slow and gatekept and offline for half the technologies that exist now. A hybrid that prefers the human-verified answer and only synthesizes when it has to is the actual shape of the future. Stack Overflow shipped it before any of its competitors did.
Then there is the MCP server. Stack Overflow’s Model Context Protocol server launched in beta in late 2025 — currently rate-limited to 100 requests per day per user, but live. Any MCP-compatible client — Claude Desktop, Cursor, VS Code with Copilot, your own .NET 10 agent built on Microsoft Agent Framework — can pull the trusted knowledge base directly into the agent’s context window through a standardized protocol. No bespoke API integration. OAuth through your existing Stack Overflow account. The first corporate-knowledge MCP that HP integrated into their developer-experience experiments was Stack Overflow’s, and HP’s distinguished technologist on the project called it “a robust source of truth” and “a very exciting precedent.” There is also an internal version — Stack Internal MCP — that lets enterprise customers pipe their own private Q&A into the same agent workflows.
Translation: instead of fighting the fact that every developer’s primary interface is now an AI agent, Stack Overflow is making sure the agent can ask their database. That is exactly the move. It is the only move that keeps the community contributing while still meeting users where they actually live.
And then there is ProLLM.
ProLLM is the benchmark suite Stack Overflow runs through Prosus Group, evaluating LLMs on actual Stack Overflow questions across two leaderboards: StackEval, with 925 historical samples from 2018 to 2023, and StackUnseen, with 194 questions from after the typical training cutoff. The point of StackUnseen is the point everyone in the LLM industry is currently pretending isn’t a problem: models that aren’t continuously retrained on new community Q&A fall behind on new and emerging content. New library versions, breaking API changes, new framework conventions, new error patterns — none of that exists in a model frozen at its training cutoff. ProLLM measures, with numbers, exactly how far behind the frontier models drift on questions humans actually ask in a given week.
This is, quietly, one of the most useful evals anyone is publishing. It is also a not-very-subtle argument that the LLM industry needs Stack Overflow more than Stack Overflow needs the LLM industry — because the second the community stops generating new answers, every model in the world starts going stale on the same schedule.
Why this matters more than the obituary writers think
There’s a thing the AI maximalists keep getting wrong. They look at Stack Overflow’s question-volume chart, draw a line to zero, and assume the world is fine without it because models have already absorbed everything worth knowing.
This is wrong in two ways.
First, it’s wrong technically. The most-upvoted answers on Stack Overflow are largely from 2012 to 2016. They are aging. The languages, frameworks, and runtimes they describe have all moved several major versions forward. A model trained on that corpus, without a mechanism to ingest new community-verified content, drifts toward giving you 2014’s answer to 2026’s question. That happens slowly enough that most users don’t notice — until they do, usually at 2 AM in a production incident.
Second, it’s wrong culturally. The reason Stack Overflow worked wasn’t the questions or the answers. It was the public-ness of the disagreement. Six people would post six different solutions to the same problem. The community would vote. The accepted answer wasn’t always the highest-voted. Comments would catch errors. Edits would update the answer when the underlying library changed. That whole structure — adversarial, skeptical, transparent — produced something LLMs structurally cannot: a record of how a real community of working engineers reasoned about a problem in public, including their mistakes.
You can replace the lookup. You can’t replace the reasoning trace. And the reasoning trace is exactly what most developers actually need when they’re stuck.
Give them their flowers
Here is what I’d like the rest of the industry to do, before it’s too late.
Stop writing the obituary. Stack Overflow’s question volume is collapsing, but the company has shipped more meaningful product in the last twelve months than it shipped in the previous five. AI Assist is the right architecture. The MCP server is the right integration. ProLLM is the right benchmark. The redesigned site is in beta and rolling out to all users in April 2026. They opened public chat to all registered users. They started supporting opinion-based questions in October 2025 — the exact category developers were taking to ChatGPT because Stack Overflow used to close them. They are, finally, listening.
Pay them. If your AI product was trained on the Stack Overflow data dump — and it was, because everyone’s was — at minimum link to the original answer when you surface its content. Better, license the data the way Google, OpenAI, and GitHub already have. Even better, contribute. The next time your team solves a hard problem with a coding assistant, take ten minutes and write up the question and answer on Stack Overflow itself. The feedback loop only restarts if the people who currently take from it start putting back.
Credit the OSS. Dapper. StackExchange.Redis. MiniProfiler. Opserver. Bosun. DNSControl. Jil. Sigil. MarkdownSharp. The Data Explorer. These are not curiosities. These are infrastructure. They were built by Stack Overflow engineers, given away free, and run in production at companies that wouldn’t exist without them. The next time someone tells you “the community” gave us modern open-source .NET, the correct answer includes a specific organization name.
Hope they make it. I’m going to write this part plainly, because I don’t have the energy to be cynical about it. I want Stack Overflow to come back. I want the question feed to start filling up again. I want the next generation of developers to learn what it’s like to have their first answer roasted in the comments by a stranger who is, annoyingly, correct. I want there to be a place that isn’t a private LLM chat where the canonical answer to a hard problem lives in public, where someone can edit it when it gets out of date, and where the people who wrote it get credit for the next fifteen years.
Stack Overflow had every reason to fold. Question volume in free fall. A parent company under pressure. The entire developer industry pivoting to interfaces that bypass them. They could have shut the lights off in 2024 and most of the noise would have been people complaining about their own moderators on the way out.
Instead they shipped AI Assist, an MCP server, an LLM benchmark, and a beta redesign. They opened up chat. They reopened opinion-based questions. They are doing the actual work of meeting the moment.
That is not the behavior of a dying company. That is the behavior of a company that has finally figured out what it is for in the new world — the curated, human-verified, citable layer underneath the agentic developer workflow — and is building toward it.
Don’t bury Stack Overflow yet. They might just be rebuilding the road we’re all already driving on, one more time, while we’re not looking.
Gal Ratner is the CTO of WhiteStar Labs and Chief Architect at Prana Entertainment. He writes about the working architect’s view of AI, infrastructure, and the platforms developers actually depend on at galratner.com.