Stack Overflow Sells the Reservoir

Stack Overflow recorded 6,866 questions in December 2025, a monthly count comparable to 2008. Its annual revenue reached $115 million, double its level at ChatGPT’s debut, with enterprise tools and AI licensing identified as drivers in Sherwood News’s account of the divergence. The place where developers ask questions is quieter; the company that sells access to their answers is making more money.

Stack Overflow has found a way to monetize its accumulated knowledge while coding assistants interrupt the public question-and-correction process that produced it. The business can succeed on the archive for longer than the archive can stay useful without renewal. That is a maintenance problem, not an imminent revenue forecast.

A public question used to create more than one answer

Stack Overflow’s original design asked a developer to expose a problem in public. Another developer could answer, someone else could challenge the answer, and a later reader could find the exchange without knowing any of its participants. The person asking got help; the site got a searchable record that could be corrected and reused. Each solved problem also gave a knowledgeable contributor a visible reason to participate in the next one.

GitHub changed the first step when it put assistance inside the editor. In December 2024, GitHub offered Copilot free in VS Code, with 2,000 code completions and 50 chat messages a month. A developer who gets a workable suggestion there need not stop to compose a public question, wait for an answer, or return to mark which advice survived testing. The suggestion may be useful. It simply leaves a different artifact.

Stack Overflow’s 2025 survey found that 84% of developers used or planned to use AI tools in their workflow, up from 76% in 2024. That survey and one December question count do not establish that Copilot caused the decline in public questions. Stack Overflow gives us a December question count here, but no comparable series for answers, active contributors, referrals, or changes in moderation and search. It does show why the old route to an answer now has a readily available private alternative.

Code agents extend that route beyond a developer’s explicit request. Cursor’s Automations can launch an agent after a codebase change, a Slack message, or a timer. Ai2 designed its SERA coding-agent family to adapt to private codebases. Those tools can address problems within the environment where the problems arise; an organization may have little reason to publish the resulting exchange. Stack Overflow cannot collect a reusable public question from a conversation that never leaves the repository.

An answer can travel farther than its correction

Developers have not stopped needing judgment. In the same Stack Overflow survey, only 33% said they trusted AI accuracy, down from 43% in 2024. An assistant can spare someone a search while still leaving that person to check whether the suggested behavior applies to the library version and codebase in front of them.

An analysis of GPT-3.5-based ChatGPT answers to 517 Stack Overflow programming questions found incorrect information in 52% of them. GPT-3.5 is not a benchmark for today’s coding agents. The study matters here for a narrower reason: plausible wording did not reliably tell users when an answer was wrong. Better models can reduce errors without making provenance, version context, or a route to correction unnecessary.

Stack Overflow promoted a GitHub Copilot extension intended to surface trusted technical answers with attribution to their source. A citation can let a developer inspect the underlying post rather than accept an unsupported sentence. It cannot, on its own, show that an old answer still describes the current API or bring a developer’s newly discovered exception back into the public thread. The extension addresses the handoff from archive to assistant; contributors still have to perform the correction at the other end.

OverflowAPI gives the archive a buyer

Stack Overflow had an enterprise business before the generative-AI boom. In 2021, it reported 1,500 customers for Teams, including Microsoft and Box. That product served organizations seeking a place for their own technical knowledge. By 2023, Stack Overflow’s CEO was discussing charges to large AI developers for access to the service’s 50 million questions and answers: a market for knowledge the public site had already accumulated.

The company made that market concrete in 2024. OverflowAPI launched with Google using Stack Overflow data for Gemini for Google Cloud. Stack Overflow then signed a deal to supply content and data to OpenAI’s models, alongside planned developer-tool integrations. Instead of requiring every developer to visit a question page, Stack Overflow could supply an AI company or enterprise product that met the developer elsewhere.

Those deals help explain why shrinking question creation has not meant reported revenue contraction. The $115 million figure is evidence of a successful commercial adaptation, not evidence that the knowledge base has stopped aging. Coverage does not establish how much revenue comes from each product line, whether licensing proceeds support moderation, or whether they reach the people who wrote and corrected the posts. Without those details, revenue cannot serve as a measure of the archive’s rate of renewal.

A license can set access terms without securing consent

Publishers have begun to formalize what an AI company may take. Reddit, Yahoo, Medium, Quora, O’Reilly, wikiHow, Ziff Davis, and others adopted Really Simple Licensing to establish terms for AI scraping. That work belongs to a broader community-data licensing market: an operator can negotiate access rather than leave every automated collector to decide what “public” permits.

Stack Overflow faces a separate question because developers supplied the underlying expertise. After its OpenAI partnership, one user alleged a week-long suspension after altering posts in protest of their use for AI training. The allegation does not establish a general contributor response. It does expose the difference between a platform’s ability to license an archive and a contributor’s willingness to keep improving it. A contract with an AI company governs access to existing posts; it does not answer what motivates the next expert to revise one.

Other knowledge businesses show why traffic alone cannot settle the issue. Chegg reported first-quarter subscribers down 31% year over year to 3.2 million and cited students’ increased AI use as part of the pressure. The Wikimedia Foundation reported a significant decline in human traffic to Wikipedia as people obtained information through chatbots without clicking through. Chegg’s paid subscriber loss and Wikipedia’s reduced human visits are different measures from Stack Overflow’s question count. All three leave an operator needing to distinguish consumption of existing knowledge from participation in making new knowledge.

Stack Overflow can sell an answer for delivery inside an editor, and a citation may bring a developer back to inspect it. The empty question box on the public site is where that answer’s replacement would have begun.