Click here to get this post in PDF
Something is going quiet on the web, and for once it is measurable. Places people used to visit by the million — to ask a question, hire a freelancer, license a photograph — are emptying out. Not because they were badly run, but because a machine now does the small job people used to show up for. There is a name for the broad version of this idea, the “dead internet.” Most of that talk is speculation. Some of it is not, and the part that is not should concern anyone who runs a business.
In a new six-part study, HarperFlow took six well-known platforms and measured what actually happened to them, using only public data anyone can pull and recompute: stock prices from Yahoo Finance, developer activity from the Stack Exchange API, and public attention from Wikimedia’s pageviews. No estimates, no forecasts. The Dead Internet research series makes uncomfortable reading, and not only for the companies named in it, because the same force reshaping them is quietly changing how your own customers find you.
Six businesses, one pattern
Start with the answer sites. Stack Overflow was where the world’s developers went to ask questions. New questions ran at 109,301 in the month ChatGPT launched in November 2022. This past month there were 1,624 — a fall of 98.5%, and a fifteen-year low. The site was already sliding gently before 2022; what changed was the slope. A slow fade became a freefall the moment a free assistant could answer the same questions instantly.
Chegg sold homework help by subscription. On 2 May 2023, the day after it named ChatGPT on an earnings call, its stock fell 48% in a single session. Measured from its 2021 peak it is now down roughly 99%, from about $96 a share to under a dollar. When your product is answers, and answers become abundant and free, there is no moat left to defend.
Quora was built on people answering each other’s questions. Public interest in it — tracked through Wikipedia pageviews, a proxy the researchers are careful not to confuse with actual site traffic — has fallen about 43% between 2022 and 2025. The kind of open-ended, “what do you think” question Quora existed for is exactly what a language model now handles in one turn.
The creative marketplaces tell the same story from a different angle. Shutterstock sold the right to use a photograph. Its shares fell from about $121 to around $5 as generative image tools made “good enough” pictures free to produce. The study is precise about it: a stock-market bubble explains part of the early drop, but not the further 39% slide through 2024, a stretch when the wider market was setting new highs.
Getty Images tried the other route — it fought. It took Stability AI to court in both the US and the UK over images allegedly used to train an image model without permission. In the UK, Getty abandoned its central copyright claims mid-trial, and in November 2025 the High Court found the model was not itself an “infringing copy.” The US case continues. By the time any of that was decided, Getty’s shares had already fallen around 90%. The report’s conclusion is blunt: you cannot sue your way out of a capability shift.
Fiverr is the most instructive of the six, because at first nothing happened. Its stock held near $30 for two full years after ChatGPT launched. Then, as the models got good enough to write the copy and translate the text that made up so much of its high-volume work, it broke — down about 71% by mid-2026. The researchers call it a two-year fuse.
Why the fuse matters more than the explosion
Read together, the six add up to a single lesson. In every case, the company sat in the middle, between a person with a need and the answer or asset that met it. For years, being that middle layer was the entire business. Once a machine could produce the answer or the asset directly, the middle layer became optional — and optional layers get skipped.
Fiverr’s delay is the part that should stay with any operator. The danger did not arrive with the headline in November 2022. It arrived quietly, a year and a half later, when the technology crossed a quality line few people were watching for. By the time the market repriced the company, the real decision had already been made thousands of times a day by ordinary users quietly choosing the faster option. Disruption rarely knocks. It is usually well inside the house before the share price notices.
The quieter collapse coming for everyone else
This matters even if you sell something a model will never generate. The same shift is changing the first step of almost every buying journey: discovery.
More of your potential customers now open an AI assistant and ask it directly — “what’s the best accounting platform for a small agency,” “who should I hire to rebuild our website,” “which suppliers do this properly.” What comes back is not ten blue links to work through. It is a short list of named companies, chosen by the model. Your business is on that list, or it is not mentioned at all. Not ranked tenth — absent.
That is a quieter collapse than a share price, and it will never appear on an earnings call. It shows up as inbound enquiries that slowly thin out for reasons you cannot quite name, because the customers who were never shown your name never knew to go looking for it.
What actually works now
There is no trick for this, and anyone selling one is worth avoiding. Language models tend to name sources that are genuinely useful, specific, and well organised: content that answers real questions plainly, carries original data or a clear point of view, and sits somewhere the model can actually read it. The old game was ranking a page for a keyword. The new one is being the kind of source a machine is comfortable repeating to a stranger who is about to spend money. That is closer to earning a citation than winning a lottery — built to be cited, never guaranteed to be.
The honest place to start is simply to look. Ask the major AI assistants the ten questions a good customer would ask right before choosing a company like yours, and read the answers closely. Note who gets named, who does not, and what the named ones did to deserve it. Most businesses have never run that check even once; it tends to be uncomfortable and clarifying in equal measure.
The six platforms in the research did not fail because they got worse at their jobs. They failed because the job itself stopped needing a middleman. Discovery is now heading through the same door. The businesses that come out the other side intact will be the ones that made themselves genuinely worth citing — before their own fuse runs out.
If you want to see how AI engines currently answer the questions that matter in your market, you can run a free GEO audit on your own site and category.
You may also like:
The Evolution of Online Search: How AI is Transforming Information Retrieval
Improve Your Brand Visibility and Reach Online with Digital Marketing
Image source: elements.envato.com
