What died, and where it went

Catalogs of AI tools grow in one direction. Products get added, launches get covered, and almost nobody goes back to check whether the thing they listed eighteen months ago still answers. In September we did — every tool, one at a time. Nearly a tenth of the catalog was not there any more.

How we counted

In September 2026 we requested the homepage of every tool in the catalog that had one on record — 6,702 of them — and classified what came back. A tool is marked dead when its domain fails to resolve, returns a 404 for its own homepage, or lands on a registrar parking page. Redirects were followed: 566 tools now answer on a different domain than the one we had, and every one of those counts as alive.

Where the answer was inconclusive we did not guess. 510 responses came back as bot protection, rate limiting, or a transport-level error, and all 510 are treated as alive. That makes every figure in this report a floor rather than an estimate: the true failure rate is higher than what follows, never lower. The catalog records 589 tools as shut down in total — 579 from the sweep itself, the rest confirmed individually afterwards.

Acquisition is tracked separately. A product that was bought and folded into a larger platform is not a failure in the sense this report is measuring, and lumping the two together is the most common way this kind of analysis goes wrong. We report them side by side and never in one number.

Which markets buried the most

The headline rate of 8.4% is close to meaningless on its own, because no builder is exposed to the whole catalog. What matters is the market you are actually in, and there the variation is severe.

Gaming, Avatars, Music and 3D & AR/VR occupy the top four places, and they have an obvious thing in common: all four sell a generated artefact to a consumer or a hobbyist. That business has two structural problems. Inference is a recurring cost that scales with usage rather than with revenue, and the buyer churns as soon as novelty fades. A tool can be genuinely good and still fail that arithmetic.

It is worth holding this alongside the opportunity scores. Gaming and 3D & AR/VR are two of the six most open markets we track — high opportunity, low saturation, plenty of room. Low saturation is not always an unclaimed opportunity. Sometimes it is a body count, and the honest reading of these two markets is that both things are true at once: the demand is real, and the people who went first mostly did not survive it.

Security at 2.7% and Learning at 3.0% anchor the other end. Both sell to a budget holder who renews on a contract rather than a subscriber who cancels in an app store, and both address a problem their buyer already had a line item for. Data (4.6%), Ecommerce (5.3%) and Marketing (5.4%) follow the same logic. The pattern across the whole table is that survival tracks who signs the invoice far more closely than it tracks how good the technology is.

Where in the funding cycle they stopped

Every one of the 589 dead tools carried a funding stage in our records. The distribution across them is not the one most people expect.

Series A leads at 141. That runs against the instinct that seed-stage companies are the fragile ones, and the explanation is about timing rather than resilience. A seed company that fails quietly often never entered a catalog like ours in the first place. A Series A company has launched, been written about, been indexed — and then has eighteen to twenty-four months to find the growth rate its next round requires. When it does not, the shutdown is public. Series A is where failure becomes visible, which is not the same as where it becomes likely.

The 135 tools recorded at public stage are a different phenomenon entirely. A listed company does not vanish quietly, so these are products retired by firms that are still trading: a feature sunset, a standalone app folded back into a suite, a brand consolidated after an earlier purchase. The product died; the company did not.

The 2022 and 2023 cohorts

Sorting the dead by founding year puts the concentration in plain view. 116 were founded in 2022 and 102 in 2023: between them, 218 of 589 deaths, more than every cohort founded before 2021 combined.

Those are the two years immediately after ChatGPT’s launch, when founding an AI product required the least conviction it ever has and attracted the most capital. The 2024 cohort shows 42 deaths and 2025 shows 1, but that is a censoring effect rather than an improvement: a company founded last year has not yet reached the point in its life where this sweep would catch it. Ask again in 2028.

What they charged

Pricing model splits the dead into three uneven groups: 320 freemium, 223 paid and 46 free.

It is tempting to read that as an indictment of free tiers, and it is worth resisting. Freemium leads the count because freemium is what most AI products are — it is the largest group in the catalog, so it supplies the most of everything. Measured as a rate against each model’s own population, the order reverses: 9.2% of paid tools are gone, against 8.4% of freemium and 7.1% of free.

So charging from day one did not protect anyone here. If anything it is mildly the other way round, and the most likely explanation is selection rather than causation: a product confident enough to charge immediately is often one that needs revenue immediately. The honest conclusion is the negative one — pricing model, on its own, tells you very little about whether a tool will survive.

Where the value went instead

162 tools were acquired rather than shut down, and they cluster somewhere revealing.

Marketing and Finance lead at 18 each, and both sit in the lower half of the failure table. That is the pattern worth taking away: acquisition concentrates in markets with established incumbents who have budget, a distribution channel, and a reason to buy rather than build. Gaming, the deadliest market in the catalog at 32.9%, recorded zero acquisitions. Nobody was buying.

For a founder, that reframes the category decision. The question is not only whether a market has room — it is whether the market contains anyone who would want to own what you build. A market with no acquirers has exactly one exit, and it is the one this report counted.

Every market, ranked

All 29 markets, sorted by failure rate. Totals include dead and acquired tools, which is why they run slightly ahead of the live counts shown elsewhere on the site.

What this changes on Falcoscan

This sweep is not a one-off study; it feeds the scores. Every dead tool is now excluded from its market’s averages, and the freshness term in each heat score is anchored to the date we last confirmed a product was reachable rather than to the date we last touched the row. The full weighting is on our methodology page. You can see the resulting state of every market in the Market Terminal Dashboard.

Photo: Mikhail Nilov / Pexels. Colour-graded for Falcoscan.

← All reports