Stack Overflow's $115M and the Missing Answer Ledger

In December 2025, Stack Overflow logged 6,866 new questions, roughly its monthly count in 2008, while annual revenue reached $115 million, Sherwood News reported in January. The company was earning more as fewer new questions arrived on its public site.

Stack Overflow doubled annual revenue since ChatGPT’s debut, with enterprise tools and AI licensing among its drivers, according to the report. Developers can now get help in their editor without visiting a Q&A page. That shift makes the difference between selling visits and selling knowledge unusually visible.

Developers were extracting answers before assistants arrived

Stack Overflow reported in 2021 that 25% of visitors copied text within five minutes of arriving. Over two weeks, visitors made 40.6 million copies across 7.3 million posts. Copying does not prove that a visitor ignored the surrounding explanation. It does show that moving an answer into another work surface was already a routine use of the site.

GitHub changed where that movement begins. In December 2024, it offered a free Copilot tier in VS Code with 2,000 code completions and 50 chat messages a month. In Stack Overflow’s 2021 account, extraction began on a Q&A page. In GitHub’s product, the developer begins in the editor. A pageview counts the first route and misses the second.

Stack Overflow sells machine access, not proof of use

Stack Overflow launched OverflowAPI in 2024 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 tasks. Those agreements establish commercial demand for the corpus without requiring each developer to navigate to the site.

The two uses pose different record-keeping problems. A model trained on licensed data does not thereby reveal which post, if any, influenced a particular output. A system that retrieves an answer during a coding task can record the post it fetched, but retrieval alone cannot prove that the answer caused a particular line of code. Stack Overflow can sell access under either arrangement; a reliable account of answer-level use requires more than the access contract.

A source record has to survive the code review

Stack Overflow’s 2025 developer survey found that 84% of respondents use or plan to use AI tools, up from 76% in 2024. Only 33% said they trusted AI accuracy, down from 43%. Stack Overflow can still serve developers who consult experts directly or ask agents for help. A code reviewer needs a source record to check an agent’s work, not just to credit the publisher.

One study found incorrect information in 52% of GPT-3.5 answers to 517 Stack Overflow programming questions. That is a finding about those answers, not an error rate for current coding agents. By April 2026, companies were also struggling to review and secure growing volumes of AI-generated code. In that setting, a useful retrieval record would identify the answer and its version, when it was fetched, the permission governing its use, the workflow that received it, and the human who approved the resulting change. The agentic workflow control plane is where an enterprise could attach those separate records to a code review. A citation displayed beside a generated answer cannot do that job by itself.

Publisher payments are testing contribution, not proving it

Reddit, Quora, and other publishers adopted the Really Simple Licensing standard to set terms for AI scraping. That addresses permission to access material. A reported Google pilot with roughly 100 digital publishers asks a harder commercial question: how much each publisher’s content contributes to AI Overviews, AI Mode, and Gemini. Reported payments vary widely.

Neither a scraping term nor a publisher payment establishes a dependable way to assign value to one programming answer used in one software change. A corpus license identifies an authorized supplier. Answer-level settlement would additionally require the supplier and agent operator to agree on what counts as use—and whether a retrieved passage, a model’s learned behavior, or a reviewer’s direct consultation earns credit. Those are different events, even when they produce similar code.

The archive still needs people to keep it useful

Stack Overflow’s revenue growth shows that it can earn more while receiving fewer public questions. The December count does not measure edits to existing answers or prove that the archive has gone stale. It does narrow one visible route by which knowledge about new tools enters the public corpus.

The contributor bargain is less settled than the customer bargain. After the OpenAI partnership, one user said Stack Overflow suspended his account for a week when he altered posts to protest AI-training use. One dispute does not establish how contributors generally feel, but it exposes a governance question a bulk license cannot resolve: who gets a say when volunteer answers become a commercial input?

Stack Overflow can earn $115 million from a business built partly on licensed knowledge. Unless contributors keep that knowledge current on terms they accept, an agent can document every source and permission—and still deliver an obsolete answer.