Reviews stop being written. They start being overheard.
Stack Overflow went from 200,000 questions a month to a few hundred. If people stop publishing for humans, how does AI find out which products are any good? Five answers, and the one nobody is discussing.
Read the argumentIn January 2026 a chart went round showing Stack Overflow’s monthly question volume across its entire history. The shape is brutal. A peak of roughly 200,000 questions a month in 2014, a slow slide through the late 2010s, then a cliff after ChatGPT launched in November 2022, ending at a few hundred questions a month by early 2026. Either way you filter the Data Explorer query, it is a rounding error against the peak, and the lowest the site has been since it opened in 2009.
The obvious reading is that developers stopped asking questions. That is not what happened. Developers ask more than ever. They just ask the model in their editor, and nobody else ever sees the exchange. The question got asked, the answer got given, and the public artefact that used to be produced as a byproduct was not.
That byproduct was the training data. It was also the thing the next developer found on Google two years later. Stack Overflow is the cleanest available example of a loop closing on itself, which is why it gets cited every time somebody raises the Dead Internet Theory.
The interesting version of the question is not about Stack Overflow. It is about reviews. Product knowledge is the part of the open web with the most direct commercial logic behind it, and therefore the part where the incentives break most cleanly. If nobody has a reason to publish “I bought this dishwasher and here is what is wrong with it,” how does an agent buying a dishwasher on your behalf know anything at all?
Is the open web dying, or just changing shape?
Both halves of the Dead Internet Theory are now measurable, and the numbers are worse than most people assume. Different studies, different methodologies, shown together to indicate direction.
The direction of these numbers is no longer a contrarian read. On 9 August 2026 Elon Musk, replying to a thread on global bandwidth capacity, wrote that “AI agentic Internet traffic will obviously VASTLY exceed human usage” and called Cloudflare's forecast accurate, adding that it was not a close call. The post passed four million views within a day. Whatever you make of the messenger, the operator of the largest satellite internet constellation planning capacity around agent-majority traffic is a signal in itself. Source
What all of these describe is one specific thing: the open, ad-funded, search-indexed, publicly scrapeable web. That is the layer going quiet. It is not the same as human activity going quiet.
People did not stop talking about products in 2023. They moved. Into WhatsApp groups and Discord servers, into private subreddits and paid communities, into YouTube where a twelve-minute teardown is still worth making because YouTube pays for attention directly, and into group chats where the recommendation that actually drives a purchase has always lived. Those spaces are harder to measure, which is why they are missing from the charts, and harder to scrape, which is why they are increasingly valuable.
The internet is not dying, it is bifurcating: a thinning public layer optimised for machines, and a thickening private layer where the humans went. The Dead Internet Theory measures the first and concludes the second does not exist.
Why does anyone write a product review in the first place?
The Dead Internet conversation tends to treat web content as though it were produced out of civic goodwill, and its decline as a tragedy of lost generosity. Some of it was. Most of it was not.
The overwhelming majority of product content on the open web existed because of a specific, boring, four-step loop:
- 1Someone publishes a review of a product, optimised for a search query.
- 2Google sends people who searched that query to the page.
- 3Those people click an affiliate link, see an ad, or become a lead.
- 4The revenue from step three pays for step one, and the loop repeats.
Every part of that loop depends on a human arriving at a page. Not on the review being read, particularly, and certainly not on it being good. On arrival. That is what was being monetised.
An agent researching a dishwasher on your behalf performs the first two steps and then stops. It reads the page. It does not click the affiliate link, because it is not buying through the publisher’s tracking cookie. It does not see the ad, or rather it sees it and does not care. It does not become a lead. The publisher’s content was consumed, its bandwidth was used, and its revenue was zero.
So the incentive does not weaken gradually. It inverts. Publishing becomes a cost with no return, and the rational move for a commercial publisher is to stop, paywall, or block the crawlers. All three are happening.
The loop was never “write a review so people find it useful.” It was “write a review so a search engine sends you someone who might click something.” Remove the person, and nothing about the sentence survives.
Worth being precise about causation, because “AI killed the review site” is too neat. Google was killing review sites first. The Helpful Content Update, launched in 2022 and folded into core ranking in March 2024, did enormous damage to exactly the sites that produced the product content the models trained on. Some of that was deserved. A great deal of what ranked in 2021 was thin, unfelt, written-to-rank content by people who had never touched the product. But the update did not distinguish well between that and genuine independent publishing, and the economic base under open-web reviewing was substantially removed two years before agentic browsing became a real threat to it.
If nobody writes reviews, how will AI know what is good?
This is the strongest objection in the debate and it deserves stating at full strength. A model’s knowledge of whether a dishwasher is quiet, or whether a brand of milk sours early, is not reasoning. It is memorised human testimony. If people stop writing it down, the model’s product knowledge freezes at the moment the writing stopped, and every product released after that is a gap. Worse, the gap is invisible: the model will answer confidently about the 2028 dishwasher using patterns from the 2024 one.
Compounding it, the models are approaching a measured ceiling on the raw material. Epoch AI puts the effective stock of quality-adjusted human public text at around 300 trillion tokens, with an 80% confidence interval for that stock being fully used between 2026 and 2032, since revised towards 2028 at the front. The era of simply finding more text is ending.
But the objection assumes written public text is the only channel through which the physical world can report on itself. That was true for about twenty years and is becoming untrue quickly. Text was never the signal. Text was the cheapest available encoding of the signal, in an era when the only practical way for a person to tell a machine something was to type it into a box a crawler could reach. Millions of people form an opinion about a dishwasher every day. What is disappearing is one pipe that used to carry those opinions to a crawler.
Five places product truth comes from next
Not equally mature. Each is graded by how much is observable today rather than projected.
Licensed and structured feeds
Platforms stopped giving data away and started selling it. Reddit is the reference case: roughly $60 million a year from Google signed in early 2024, roughly $70 million a year from OpenAI that May, and an aggregate contract value of $203 million disclosed in its IPO prospectus. That is around 10% of Reddit’s revenue from selling human conversation to model builders. In parallel, manufacturer catalogues and retailer inventory APIs are becoming the primary source for hard product facts, because they are authoritative, real-time, and need no inference from prose.
Gated human signal
The high-trust opinion is moving behind walls: paid communities, private Discords and WhatsApp groups, verified-purchase-only review systems, and video. Video matters most here. A twelve-minute teardown is expensive to fake, carries visual evidence text cannot, and is monetised by attention directly rather than through a search intermediary that agents bypass. That makes YouTube one of the few places where honest product reviewing still pays, which is exactly why it is becoming a licensing target.
Paid real-world capture
The clearest evidence the industry already accepts the free text is running out. Companies now pay people to generate physical-world data directly. Micro1 reports roughly 4,000 contributors across 71 countries producing more than 160,000 hours of video a month. DoorDash has paid delivery drivers to film household chores at rates reported up to $25 an hour. Instawork opened a robotics lab paying for two to fifteen minute chore recordings. This is aimed at robotics, not reviewing, but the precedent is the point: when the free text ran out, the response was to pay humans for non-text data at industrial scale.
Agent-run testing
If an agent can order a product, it can evaluate one. Automated benchmarking, controlled purchase-and-return testing, and aggregated outcome tracking, meaning whether the customer kept it, returned it or claimed on the warranty, produce product knowledge with no human writing at any stage. Currently rare and mostly confined to software and electronics, where testing is cheap and instrumentable. It is hard to see how it works for milk.
Overheard speech
The one this series has been building towards. Hundreds of millions of always-listening devices are already installed in homes, cars and offices, and they are being upgraded from keyword matchers to LLM-grade assistants right now. Those devices sit inside the exact moments when people form and voice product opinions. Nothing about this is technically hard. All of the difficulty is consent.
Note the pattern across the first four. In every case the response to the drying-up of free human text has been to pay for a denser signal in a format that was previously too expensive to collect. The fifth is what happens when the cost of collecting the densest format of all, everyday speech, falls to roughly zero because the microphone is already there and already powered.
What does an overheard review actually look like?
Not like a review. That is the entire point.
Don’t buy that milk again, it went off two days early.
Kitchen, unprompted, three days after purchase
This thing is useless in low light, the old one was better.
Living room, during use, with a direct comparison
I can’t believe how quiet the new dishwasher is.
Kitchen, unsolicited, positive, nobody asked
Third time this month the app has logged me out.
Car, in frustration, timestamped and repeatable
None of these would ever be typed into a review box. The milk one is the instructive case: nobody in the history of the internet has gone to a website to leave a considered two-paragraph review of milk. But nearly everybody has said that sentence out loud in a kitchen. It is real product information about a real purchase and it has, until now, evaporated the instant it was spoken.
Why unprompted speech carries different information
| Property | Written online review | Overheard remark |
|---|---|---|
| Trigger | Solicited by an email prompt, or self-selected by someone unusually delighted or unusually angry | Triggered by the product itself, in the moment it succeeded or failed |
| Timing | Written once, days or weeks after purchase, from memory | Continuous across the whole ownership period, including month nine when the fault appears |
| Selection bias | Severe. Reviews cluster at one and five stars because moderate experiences do not motivate a trip to a website | Low on motivation, but heavily skewed by who owns a listening device and who leaves it on |
| Comparison | Rare, and usually to a competitor the reviewer has not owned | Natural and constant, because people reference the thing they actually replaced |
| Context | Stripped. The review does not know the room, the use case, or who else was present | Rich and automatic. The device already knows time, place, and often the purchase history |
| Incentive to fake | High, and industrialised. Paid review farms are a mature market | Low per instance, and expensive to fake at scale |
| Detail and reasoning | Strong. A good written review explains why, at length, with photographs | Weak. A spoken fragment gives a verdict with almost no supporting argument |
| Consent and legality | Unambiguous. The reviewer chose to publish | The central unsolved problem |
Read the last two rows before getting excited about the first six. Overheard signal is not a better review. It is a worse review and a better sensor. It gives you a very large number of very thin, very honest, very well-contextualised data points instead of a small number of thick, considered, heavily biased ones. Those are different instruments, and for the job of keeping a model’s product knowledge current, the thin honest stream is arguably the more useful.
It is also harder to fake. Written review fraud is a mature industry because the attack surface is tiny: text, an account, and a way to bypass detection. Faking overheard signal means producing spontaneous-sounding speech in physically distinct acoustic environments, on devices with distinct hardware fingerprints, tied to accounts with plausible purchase histories, across enough households to move the aggregate, sustained without falling into detectable patterns. Not impossible, but several orders of magnitude more expensive. Signal that is a byproduct of doing something else is harder to forge than signal produced deliberately, because forging it means faking the underlying activity too.
What stops this from happening?
Quite a lot, and this section is longer than the optimistic ones for a reason. The technical path is the easy part. Everything difficult about ambient review capture is human, legal and political.
Continuously capturing household conversation and using it to build a commercial dataset runs straight into GDPR’s consent and purpose-limitation requirements, the EU AI Act’s transparency obligations, and national wiretapping laws written long before any of this existed. It also runs into memory. In 2019 Amazon, Google and Apple were each found using human contractors to review voice recordings, including ones captured accidentally. As covered in The Home, that episode is still the reference point for public trust in these devices, and it was seven years ago.
Then the simplest counter-argument to this entire report: people just turn it off. A meaningful share of smart speaker owners already mute the microphone or never enabled the assistant. If ambient product capture is presented as the feature, most households will decline, and the ones that accept will not be a representative sample of buyers. A signal drawn from the subset of people relaxed about always-on microphones is biased, and the bias correlates with age, income and technical confidence in ways that would distort product knowledge badly.
What a workable version would require
On-device processing
Raw audio never leaves the building. What leaves, if anything, is a structured anonymised assertion: category, sentiment, attribute, comparison. Not a recording.
Explicit, granular, revocable opt-in
Not buried in a terms update. A separate, plainly worded choice that can be withdrawn without losing the rest of the device’s function.
Provenance metadata
Downstream models need to know a signal came from genuine spontaneous use rather than staged or synthetic speech, or the channel becomes as poisoned as the open web it replaced.
Benefit that returns to the household
“Your assistant warns you this brand sours early in your fridge” is a proposition. “We are collecting your kitchen conversation to improve our models” is not.
An answer on who owns it
If three companies control the microphones, they control the product-truth layer, and every brand’s visibility becomes a licensing negotiation. That is a worse concentration of power than Google has over search today.
Honest confidence rating. That written open-web reviewing collapses as a commercial activity: high confidence, largely already happened. That product knowledge re-routes into licensed feeds, gated communities and paid capture: high confidence, observable today. That ambient devices become a meaningful source of product signal: moderate confidence at best, and dependent on a consent settlement nobody has yet designed.
On timing, the likeliest first appearance is narrow and opted-in, in vehicles and workplaces where the device is company-owned and the consent conversation is simpler, somewhere in 2027 to 2029. Broad household adoption is a 2030-plus question and may never fully arrive. Anyone giving you a firmer date is guessing.
What should a business do about this now?
The strategic error would be waiting to see how the review economy resolves. These four are worth doing whichever channel wins, which is what makes them worth doing now.
Publish machine-readable product truth
If an agent is the reader, prose is a lossy format. Structured specifications, real-time pricing and stock, warranty terms, compatibility. The brands that win agent recommendations are the ones an agent can answer questions about without guessing.
The agentic web stackExpose services as agent-callable tools
An MCP endpoint and an A2A agent card cost very little and decide whether an agent can transact with you or merely read about you. This is the SEO of the ambient era: invisible to humans, decisive for machines.
Check where a site standsOwn your first-party outcome data
Returns, warranty claims, support tickets, repeat purchase rates, usage telemetry. The one category of product truth competitors cannot license from a platform, because only you have it. Behavioural, continuous, unsolicited, and much harder to fake than a testimonial.
Attach provenance to everything
In a market where roughly three quarters of new pages contain machine-generated text, the scarce asset is not content. It is verifiable origin. Signed data, named authors with real track records, timestamps, evidence of first-hand testing.
The old game was ranking for a query a human typed. The new game is being the source an agent trusts when nobody typed anything at all.
Frequently asked questions
Sources: Stack Overflow Data Explorer question volumes as circulated January 2026. Cloudflare Radar bot traffic share, June 2026. Imperva Bad Bot Report 2026. Ahrefs study of 900,000 newly published pages, April 2025. Epoch AI on the limits of LLM scaling based on human-generated data. Reddit S-1 prospectus and reported Google and OpenAI licensing terms, 2024. Published analysis of 671 travel publishers following the Helpful Content Update. MIT Technology Review and Washington Post reporting on egocentric video data collection, 2026.
Ambient AI, parts one to three
Is your product legible to an agent?
We build the machine-readable layer: structured product data, MCP endpoints, A2A agent cards, and first-party data pipelines. Start with a free AgentReady scan, or talk to us about the build.