Stack Overflow After ChatGPT: Questions Fell 73%, Revenue Rose 12%

Stack Overflow received 108,981 new questions in calendar 2025, down 73% from 398,648 in 2024 and 95% below its 2016 peak. In the year to March 2026 its revenue rose 12% to $129 million. The Stack Overflow traffic decline did not kill the company. It killed the forum as the thing the company sells, and replaced it with data licensing and an enterprise product called Stack Internal.

Stack Overflow question volume: the decline, counted year by year

The cleanest public measure of Stack Overflow's community health is the number of new questions per month. Stack Exchange publishes this through its Data Explorer, a live copy of the site's database that anyone can query. The standard questions-per-month query counts posts of type 1, which are questions, grouped by the month they were created.

PPC Land ran that count in August 2026 and published the annual totals. They are the backbone of this post. I have computed the year-on-year changes myself from those totals.

PPC Land did not publish totals for 2017 or 2021, so those rows are absent rather than estimated. The Data Explorer excludes deleted questions, which understates older years and makes every decline here conservative.

The break came after ChatGPT, the erosion came before it

Questions were already falling slowly from 2016 to 2019. The lockdown year briefly reversed that. Then the rate of decline accelerated every single year: 41% in 2023, 49% in 2024 and 73% in 2025. That is not a one-off shock that settles. It is a compounding loss.

The monthly figures make the point sharper. PPC Land's count shows 17,938 questions in December 2024 and 3,314 in December 2025, an 81.5% fall in twelve months. July 2026 logged 1,442. The site launched in 2008, and its monthly volume is now below what it handled in its first full year.

What the CEO says got lost

Chief executive Prashanth Chandrasekar has argued the loss is concentrated at the easy end. On The Verge's Decoder podcast in December 2025, as quoted by Sherwood News, he said: "pretty much all those declines were with very simple questions. The complex questions still get asked on Stack because there's no other place."

I think that is partly right and strategically dangerous. Simple questions are what brought new users to the site and trained them to answer. A forum that only receives hard questions is a forum with no on-ramp. The wider pattern of community platforms losing their entry traffic is covered in the analysis of why community mentions drive AI visibility, and this post does not repeat it.

Stack Overflow revenue went the other way

Stack Overflow has been owned by Prosus since 2021, so it publishes no accounts of its own. Prosus reports it inside an Edtech segment alongside GoodHabitz, a Dutch corporate learning company. The segment figures are primary. The Stack Overflow figures within it come from Prosus's annual reports as summarised by financial press.

The direction is unambiguous. Prosus's FY26 results booklet states that Stack Overflow and GoodHabitz grew revenue 11% to US$189 million, primarily driven by Stack Overflow, and that both companies reached profitability for the year. The FY25 equivalent was 15% growth to US$170 million.

Inside that, Stack Overflow itself reported $129 million of revenue for the year to March 2026, up 12%. Adjusted EBITDA swung to a $12 million profit from a $9 million loss. On adjusted EBIT, a stricter measure that also deducts depreciation and amortisation, it made $1 million. The prior year was $115 million of revenue with $22 million of trading losses, and Sherwood reports losses were $84 million in FY2023.

What Prosus does not break out

The useful question is which revenue lines did the growing. Prosus does not split Stack Overflow's revenue into advertising, enterprise subscriptions and data licensing in any FY25 or FY26 material I could verify. That gap is the most important limitation of the public record on this company, and the table below separates what is disclosed from what is not.

Stack Overflow data licensing: the deals, and how big they can be

Stack Overflow's licensing product is OverflowAPI. It debuted in June 2023, according to SiliconANGLE's report on the OpenAI deal. Subscribers get the existing archive of questions and answers, new posts as they arrive, and feedback signals such as vote counts that tell a model which answers the community trusted.

Google first, then OpenAI, with no prices attached

Google was announced as a partner on 29 February 2024, bringing Stack Overflow content into Gemini for Google Cloud. OpenAI followed on 6 May 2024, with a commitment to attribute the Stack Overflow community inside ChatGPT answers. Neither company disclosed a price. Neither did Stack Overflow, and Prosus has not itemised either contract.

The deals are a recognisable category: an archive of human-verified content, sold to the companies whose products reduced the flow of new content into it. Across the wider market, the terms that are public are tallied in the ledger of AI content licensing deals. Stack Overflow's are among the least transparent entries in that ledger.

A ceiling on licensing, computed from Prosus's own wording

There is one way to bound the size of licensing without a disclosed number. Prosus's FY25 results booklet said Stack Overflow's API offering was primarily responsible for its revenue growth in FY25. That year, revenue rose 17% in local currency to $115 million.

A 17% rise to $115 million implies a prior year near $98 million, so total growth was roughly $17 million. If the API was "primarily" responsible, it added more than half of that and less than all of it. My working estimate is that licensing added somewhere between $9 million and $17 million of new revenue in FY25. That is my calculation, not a Prosus figure, and currency effects blur it by a point or two.

That number matters because it is small. A business losing most of its community traffic found a new revenue line worth perhaps 10% to 15% of its total. That is a real offset. It is not a replacement for a forum that once received two million questions a year.

Stack Internal: the enterprise product carrying the business

The larger revenue line is enterprise software. Stack Overflow for Teams, the private question-and-answer product, was renamed Stack Internal at Microsoft Ignite in November 2025. Techzine's report on the launch, dated 19 November 2025, describes it as a way to capture internal technical knowledge for both employees and AI systems.

The detail worth noticing is in the reasoning. According to that report, Stack Overflow saw several enterprise customers using its public API to train their own models. Rather than treat that as leakage, it built a product around it. Stack Internal adds metadata such as author, date, tags and a reliability score, and ships with a Model Context Protocol server. MCP is the open standard that lets coding agents such as GitHub Copilot, ChatGPT and Cursor query an outside tool directly.

Why the MCP server is the strategic move

The MCP connection changes what Stack Overflow is selling. The old product was a destination: developers visited a page and the company sold ads and job listings around it. The new product is a supplier: the agent visits, and the developer may never see the page at all. How MCP became the default plumbing for this is covered in the explainer on MCP as the agent interoperability standard.

Techzine names Mercado Libre, Xerox and Eli Lilly among beta users. Sherwood reports that Stack Internal is used by 25,000 companies. That count likely includes free and small team plans, and no revenue per customer has been disclosed. I would treat the 25,000 as a reach figure, not a revenue figure.

What Prosus's write-downs say about the Stack Overflow pivot

Revenue growth and a profitable year sound like a successful turnaround. The owner's balance sheet says something harsher. Prosus bought Stack Overflow for $1.8 billion in June 2021, five months before the question count began its steepest fall.

The write-downs followed in stages. Prosus's FY24 results recorded impairments of US$374 million, primarily related to Stack Overflow, and blamed the broad adoption of generative AI tools. In FY26 Prosus booked US$482 million of impairments on goodwill and other assets, mainly at Stack Overflow and GoodHabitz. Moneyweb reports the Stack Overflow share as a $280 million goodwill charge.

Moneyweb reported in July 2026 that Prosus had carried Stack Overflow at $564 million at the start of FY26, and at about $230 million after the year's impairments. That is a 59% cut in one year, and roughly 87% of the purchase price written off in total. The chart below reconstructs the path from those reported figures.

Two ratios follow from those numbers. At $230 million, Prosus values Stack Overflow at about 1.8 times its FY26 revenue. At $1.8 billion, it paid roughly 14 times that same revenue figure. A software business growing 12% and just breaking even on adjusted EBIT would normally trade above 1.8 times sales. The low mark says Prosus does not believe the growth is durable.

My view is that Prosus is right to be sceptical. Licensing revenue depends on a shrinking supply of new content, and enterprise knowledge tools are a crowded market. The question is whether Stack Internal can become a standalone software business before the archive's value to AI labs decays. Nothing in the disclosed numbers settles that yet.

Where this argument is weakest

Three parts of this analysis rest on less than I would like. Each one could move the conclusion, so each is stated here rather than buried.

The question counts do not agree with each other

PPC Land's Data Explorer count gives 3,314 questions for December 2025. Sherwood News, writing in January 2026, said Stack Overflow recorded 6,866 questions "last month". Wikipedia cites a 78% year-on-year fall where the Data Explorer implies 81.5%. The gap is likely a mix of deletions, snapshot timing and whether closed questions count. The direction is not in doubt. The precise monthly level is, and the annual totals in this post carry that same uncertainty.

There is also a measurement problem underneath. Questions asked are not page views. Developers may still read old Stack Overflow answers through search, or through an AI assistant that cites them. The question count measures the supply of new content, and that is exactly what a licensing business depends on, but it is not a traffic measure in the analytics sense.

Fiscal years and calendar years do not line up

Prosus's fiscal year ends on 31 March. The question totals are calendar years. When this post sets 2025's 108,981 questions beside FY26 revenue of $129 million, the periods overlap by nine months, not twelve. On that loose basis, revenue per question asked rose from about $288 in the 2024 pairing to about $1,184 in the 2025 pairing. That is a vivid number and a rough one. I would not quote it without the period caveat attached.

The case that this is a good pivot

The strongest counter-argument is simple. A business that lost 95% of its core activity and still grew revenue and reached adjusted profitability has done something most media and community companies failed to do. Ad-funded publishers that lost search traffic mostly shrank. Stack Overflow found enterprise buyers and AI labs willing to pay, and it cut costs hard, including about 28% of staff in October 2023. If Stack Internal compounds, the write-downs will look like an accounting reset on an overpriced 2021 deal, not a verdict on the strategy.

I find that case credible on operations and weak on durability. Both can be true.

What the Stack Overflow case means if you own a content archive

You may not run a developer forum. If your business owns a body of expert content, such as support answers, reviews, documentation or a community, the Stack Overflow sequence is the most complete public case of what happens next.

Three lessons are worth taking. First, the archive's licensing value is real but bounded, and it peaked when the content was freshest. Stack Overflow licensed at the moment its supply was collapsing, which is the weakest negotiating position available. The underlying legal fights over who owns that training value are tracked in the review of AI copyright cases on training data.

Second, the defensible product was not the data. It was the workflow around the data: structured internal knowledge, with metadata, delivered to the agent a developer already uses. That is the distinction drawn in the comparison of data moats and workflow moats, and Stack Overflow is a live test of it.

Third, the consensus framing, that AI killed Stack Overflow, is wrong about the company and right about the forum. If I owned a content archive today, I would price a licence on the assumption that its value halves every year new contributions decline. I would spend the proceeds building the enterprise product rather than defending page views.

Frequently asked questions

Where to start if your content is someone's training data

Pull your own version of Stack Overflow's chart this week. Plot new contributions per month for the last 36 months, whatever contributions means in your business: answers, reviews, tickets resolved, posts written. If the line bends after November 2022, you have the same problem on a smaller scale, and your archive's licensing value is falling while you watch.

Then check who already reads it. Look in your server logs for AI crawler and API traffic, and compare it with human visits. If machines read more than people, you already have a licensing negotiation. You just have not started it.

Related reading