When AI Can Do Your Product's Job, What Keeps Customers?
What Lasts · Beyonds Journal · 2026-11-05
In December 2025, developers posted fewer than 7,000 new questions on Stack Overflow, the site where programmers have asked each other "why does this break?" since 2008. At its peak in early 2014, the site received more than 200,000 questions a month. For many programmers that job is now done inside the code editor, where an AI assistant answers before anyone opens a browser.
In the fiscal year ended March 2026, Stack Overflow's revenue grew 12%, to
29 million, according to the annual results of Prosus, the technology investor that owns it. On Prosus's adjusted measure, which leaves out write-downs, the business moved from a loss of $57 million 2 years earlier to a small profit.
<7,000 — New Stack Overflow questions in December 2025, against 200,000+ a month in early 2014. DevClass; Sherwood News, Jan 2026
29M — Stack Overflow revenue in the year to March 2026, up 12%. Prosus, 27 Jun 2026
If a general model can now do what your product does, the decision in front of you is whether to keep building, and what to build so customers stay. Stack Overflow's last 3 years answer part of that.
Its owner credits Stack Overflow's growth to 2 things. One is an archive licensed to AI companies, an asset few startups will ever have. The other is a private product inside each customer company, which a startup can start building with its first customer.
What did AI take from Stack Overflow?
The public question-and-answer site lost its main job. A peer-reviewed study in PNAS Nexus compared Stack Overflow with programming forums in Russia and China, where ChatGPT was restricted. Within 6 months of ChatGPT's release, "activity on Stack Overflow decreased by 25% relative to its Russian and Chinese counterparts."
Question volume had been falling since around 2014, long before ChatGPT, so not all of the decline can be put down to AI. The monthly counts for late 2025 also differ between reports: DevClass counted 3,862 questions in December, Sherwood News 6,866. Either way, the public site's question volume has collapsed.
At Beyonds we call this a playground: every way one job gets done in a particular place, under that place's rules. Here the job is fixing the errors a programmer keeps getting stuck on. An assistant inside the editor, which the programmer already has open, has joined the players, and it answers before a forum page loads.
What held, and why?
In its results, Prosus has credited 2 parts of the business with Stack Overflow's growth, one in each of 2 years.
The first is a private version inside companies. Stack Overflow for Teams, renamed Stack Internal in November 2025, is a question-and-answer space for one company's engineers. Prosus credited it with Stack Overflow's revenue growth in the year ended March 2024. The company says it is used by "more than 20,000 organizations". It connects to Microsoft Teams, to Microsoft 365 Copilot, and, through an MCP server (a standard connector for AI assistants), to GitHub Copilot, ChatGPT and Cursor.
What sits inside it is knowledge no public model has seen: how this company's systems work, why a decision was made, which fix worked last time. A general assistant answers public coding questions. It can explain why your payments service retries 3 times only if someone wrote that down where the assistant can read it.
Our reading: what lasts is each company's own knowledge, kept where its people and its AI assistants look for it. A model can copy a feature in one release; moving that knowledge out of a product means rebuilding it customer by customer.
A model can copy a feature in one release; moving that knowledge out of a product means rebuilding it customer by customer.
The second is the archive, licensed to the models. In 2024 Stack Overflow signed data deals with Google Cloud and OpenAI, giving them access to its questions and answers through a product called OverflowAPI. The financial terms were not disclosed. Prosus named the API offering as a driver of growth in the year ended March 2025.
Prosus's figures show the result: revenue of $98 million,
15 million and 29 million in the 3 years through March 2026, and an adjusted operating result moving from a loss of $57 million to a profit of million. The same figures come with limits. Stack Overflow cut its staff by about 28% in October 2023, so part of the improvement comes from cutting costs. Prosus does not split revenue between the internal product and data licensing.
Stack Overflow revenue, fiscal years ended March (Prosus annual results, 2024–2026)
Year to March 2024: $98M
Year to March 2025:
15M
* Year to March 2026:
29M
Prosus paid about .8 billion for the company in 2021, then wrote down its value 3 times: $560 million in the year ended March 2023; $374 million a year later, mostly for Stack Overflow; and $482 million in the year ended March 2026. The last one also covered another holding, and Prosus linked it to generative AI. Our reading: the parts that held carry a much smaller company than the one Prosus paid for in 2021.
Do AI products lose customers faster than other software?
In this data, yes, and price is where the gap shows. In ChartMogul's December 2025 report, which Issue 02 of this Journal covered in detail, AI-native products, built around AI from the start, kept 70% of revenue from existing customers over a year when they charged over $250 a month, which the report calls "essentially the same as B2B SaaS." Those under $50 a month kept 23%.
Number — The 70% and 23% are gross revenue retention: the share of last year's revenue from existing customers still there a year later, before any expansion.
Our guess is that price stands in for how deep a product sits in someone's work: a team that puts its work into a tool pays more and leaves less often. The report doesn't test that, and it gives no sample sizes for the price groups.
Integrations are easy to copy too. Isn't a "place in the work" just another feature?
Partly, yes. A connector is a feature. Standards like MCP make connecting to other tools easier every month, for you and for every competitor, so "we integrate with Slack" protects nothing on its own.
What a competitor can't copy by shipping is what has built up inside your product for one customer, such as the routine their team built around your screens. Stack Overflow's public archive went to OpenAI and Google under license. The private knowledge inside each company stayed in Stack Internal.
Grammarly made what we read as a similar bet after ChatGPT. It bought Coda, a documents company, and Superhuman, an email app, and in October 2025 it renamed the company Superhuman. Its chief executive said email is "where professionals spend significant portions of their day." The company is privately held, so the result can't be read from public numbers yet.
Chegg, the subject of this issue's Breakdown, shows the limit: its existing customers stayed, and far fewer new ones arrived once the job they came for moved into tools students already had open. A place in the customer's work protects the customers you already have, and Chegg's new customers never got as far as the product.
How can you check where your product sits in your customer's work?
This week, take one real customer, the one you know best, and imagine a general assistant doing your core feature's job for them tomorrow.
Write down what they would have to rebuild to leave you. Their data and history in your product. The steps their team takes inside it. The people downstream who use what your product produces: an approval, a report, a handover. If the list is empty, this customer is paying for the feature alone.
Then look at what has grown since they joined, such as a library of answers their team wrote. Pick one thing the customer would miss for its own sake, and strengthen it this quarter.
What becomes harder to replace as a product gets used is the subject of our Deep Guide Real Moat. It is free and includes a 5-question Real Moat Audit you can run on your own product.
Guide — Real Moat: what becomes harder to replace as your product gets used, with a 5-question audit.
FAQ
What keeps customers when AI can copy your product's feature?
What the customer has built inside the product: their data and history, the routines their team runs in it, and other people's work that depends on its output. A model can copy a feature quickly, while moving that work elsewhere means rebuilding it customer by customer. This slows losses and guarantees nothing about survival.
What are switching costs in SaaS?
Switching costs are what a customer loses or has to redo when moving to another product: data migration, retraining a team, rebuilding integrations and routines, losing history. They tend to grow the longer a team uses a product for real work. Integrations alone are easy for competitors to match.
Is Stack Overflow dying?
The public site has lost most of its question volume: a peer-reviewed study measured a 25% relative drop in activity within 6 months of ChatGPT's release, and late-2025 monthly counts were a few thousand questions. The company's revenue still grew to
29 million in the year ended March 2026. Prosus attributed the growth to its internal product for companies in the year ended March 2024, and to its API offering, which licenses the archive, the year after.
What happened to Grammarly?
Grammarly faced competition from ChatGPT in 2023, put its AI assistant inside the apps where people write, acquired Coda and the email client Superhuman, and in October 2025 renamed the company Superhuman. Grammarly continues as one of the renamed company's 4 products.
Do AI products have worse retention than regular SaaS?
In ChartMogul's 2025 data, the median AI-native company's gross revenue retention reached 40% in September, below the level for B2C subscription software. Retention varied strongly with price: AI products priced over $250 a month retained about as well as B2B SaaS, while those under $50 retained far less.
Sources
- Prosus, Financial results FY26 (booklet), June 27, 2026: link
- Prosus, Financial results FY25 (booklet), June 21, 2025: link
- Prosus, Summary financial results for the year ended 31 March 2024, June 24, 2024: link
- Prosus, results announcement (JSE SENS), June 24, 2024: link
- Prosus, results announcement (JSE SENS), June 27, 2023: link
- Prosus, "Prosus to acquire Stack Overflow," June 2, 2021: link
- del Rio-Chanona, Laurentsyeva and Wachs, "Large language models reduce public knowledge sharing on online Q&A platforms," PNAS Nexus 3(9), September 2024: link
- Gergely Orosz, "Stack Overflow is almost dead," The Pragmatic Engineer, May 15, 2025: link
- Tim Anderson, "Dramatic drop in Stack Overflow questions," DevClass, January 5, 2026: link
- Claire Yubin Oh, Sherwood News, January 7, 2026: link
- Stack Overflow, "Stack Overflow company announcement," October 16, 2023: link
- Stack Overflow, "Introducing Stack Internal," November 18, 2025: link
- Stack Overflow, Stack Internal press release, November 18, 2025: link
- Stack Overflow, Google Cloud partnership, February 29, 2024: link
- Stack Overflow, OpenAI partnership, May 6, 2024: link
- Kyle Poyar, "The AI churn wave," ChartMogul, December 10, 2025: link
- Grammarly, "Introducing GrammarlyGO," March 9, 2023: link
- Grammarly, "Grammarly to acquire Coda," December 17, 2024: link
- "Grammarly acquires AI email client Superhuman," TechCrunch, July 1, 2025: link
- Grammarly Support, "What is Superhuman": link
- Real Moat, Deep Guide, Beyonds: link
Beyonds Journal, issue 003