Stack Overflow’s $115 Million Question

Stack Overflow received 6,866 questions in December 2025, a monthly count comparable to 2008. Sherwood News reported $115 million in annual revenue, double its level since ChatGPT’s debut.

GitHub lowered the price of asking for coding help inside the editor. Developers still bear the effort of turning a verified fix into an answer strangers can find. The question count and revenue describe different activities; neither measures AI’s effect on questions.

Public questions turned individual fixes into shared knowledge

In 2016 coverage, Stack Overflow was the community site known for answering programmers’ coding questions. A developer described a problem in public; other developers could answer, dispute, and find that exchange later. The person needing help and the person adding to the archive participated in the same process.

Stack Overflow’s 2021 copy-and-paste analysis counted 40.6 million copies across 7.3 million posts during a two-week measurement period. Copying does not establish that a snippet was correct. It does show why an old answer could remain useful to people who never joined the original conversation: the archive let one public exchange serve many later problems.

The December question count measures new questions, not archive readership, answer quality, or software production. Developers can also take a problem to documentation, a GitHub issue, Discord, or a private company channel. Old answers remain available to copy even as fewer new problems enter the public record.

GitHub and Apple put the first answer beside the code

GitHub made Copilot free in VS Code in December 2024, with 2,000 completions and 50 chat messages per month. GitHub also reported 150 million developers, up from 100 million in early 2023. That growth makes a quiet public Q&A venue a poor proxy for the number of people writing software. The free tier lets a developer ask for help without composing a question for a public audience.

Stack Overflow’s 2025 developer survey found that 84% used or planned to use AI tools in their workflow, up from 76% in 2024. “Use or plan to use” is broader than active use. In February 2026, Apple added Anthropic’s Claude Agent, OpenAI’s Codex, and MCP support to Xcode 26.3. Cursor subsequently introduced automations that could start coding agents from codebase changes, Slack messages, or timers.

An editor can supply the current file; a repository can supply related code; an integration can retrieve information or initiate work. GitHub, Apple, and Cursor each reduced the effort required to get help where a developer was already working. None of those workflows requires a public account of the resolution.

Stack Overflow found buyers for the archive

Stack Overflow had an enterprise business before its AI data deals. In 2021, it reported 1,500 customers for Teams, its private Q&A product, including Microsoft and Box. Prosus acquired the company for $1.8 billion that year. Selling knowledge for internal use was therefore not an abrupt departure from a purely public business.

AI companies joined its customer list in 2024. Stack Overflow launched OverflowAPI to give AI companies access to its knowledge base, with Google using the data for Gemini for Google Cloud. It then agreed to supply data to OpenAI to improve programming-related models, with platform integrations planned. Enterprise tools and AI licensing drove the revenue increase reported in January 2026. Developers could still visit the public archive; AI companies could also use its contents in their tools.

Stack Overflow can use that revenue to maintain or improve its knowledge base; falling question volume does not prove that the archive is degrading. But licensing changes whose needs pay the bills. A model buyer wants dependable access to useful answers. A contributor wants their work handled in ways they consider fair. In May 2024, one user alleged that Stack Overflow suspended his account after he altered posts to protest the OpenAI partnership. That allegation does not establish the company’s broader treatment of contributors. It does raise a governance question a data contract alone cannot settle.

Other publishers have begun setting explicit terms for AI scraping, including through the Really Simple Licensing standard. The broader community-data licensing market can price access to an archive; the next correction still depends on a contributor willing to publish it.

Private context makes answers more relevant and harder to inspect

Ai2 designed its SERA coding agents to adapt to private codebases. For a developer diagnosing a proprietary service, that repository may contain the crucial information a generic public answer lacks. Enterprise retrieval vendors make a similar bet on controlled organizational information: Contextual AI raised an $80 million Series A in 2024 for tools that ground model responses in enterprise material.

Repository instructions are becoming part of this working context. In September 2026, Anthropic added support for AGENTS.md to Claude Code, adopting an instructions specification OpenAI had contributed to the Agentic AI Foundation. Rival tool makers converged on a way for developers to tell agents how to work inside a repository. That helps a team carry its own conventions into an agent’s task. It does not make those conventions, or a fix produced under them, available to anyone outside the repository.

An agent with tools, retrieval, and memory can work against current local information instead of relying only on its pretrained account of a language or framework. A public Q&A page leaves the question, competing answers, and later corrections open to inspection. A fix assembled from a private prompt, repository instructions, retrieved documents, and logs may serve its owner better while leaving outsiders no comparable record to challenge or reuse.

Trust gives the public record a job agents cannot skip

Stack Overflow’s survey found that 33% of developers trusted AI accuracy in 2025, down from 43% in 2024, even as reported use or planned use rose. That combination argues against treating agent adoption as proof that developers no longer need attributable sources or review.

An analysis of GPT-3.5-based ChatGPT answers to 517 Stack Overflow programming questions found incorrect information in 52% of answers. Stack Overflow also temporarily banned ChatGPT-generated responses after moderators said most were wrong. Neither measures today’s coding agents; both show why a plausible answer still needs checking.

Teams can check their own fixes with tests, maintainers, and accountable human review. Sharing a verified, nonproprietary fix takes another decision. At $115 million in annual revenue, Stack Overflow has buyers for its archive. The 6,866 questions in December mark a narrower path for the next fix that might spare a stranger from solving the same problem again.