The AI Tool Death Rate in 2026: 12% of the 17,603 Tools We Scan Are Already Offline
11 min read
Updated
Original data from a continuously-run site scanner: how many catalogued AI tools are dead, how fast working tools disappear, and which categories churn hardest.
Of the 17,603 AI tools catalogued on NeedAnAI, 2,141 — 12.2% — no longer respond at all, and only 63.5% still serve a working site at their own address. Those figures come from our own scanner, which re-checks every listed tool on a rolling 30-day cycle; the snapshot below was taken on 20 August 2026.
Directory counts of "10,000+ AI tools" are published everywhere. What almost nobody publishes is how many of those listings still work. We check, so we can say. This article is the raw output of that checking, with the queries and the caveats attached.
The headline snapshot
Every tool in the catalogue carries a site_status set by the scanner, not by an editor. As of 20 August 2026:
| Status | Tools | Share | What it means |
|---|---|---|---|
| alive | 11,183 | 63.5% | Responds normally at its own URL |
| redirect | 2,876 | 16.3% | Resolves, but somewhere other than the URL we hold |
| dead | 2,141 | 12.2% | Three or more consecutive failed checks |
| unreachable | 1,353 | 7.7% | Failing checks, below the "dead" threshold |
| unknown | 50 | 0.3% | Not yet classified |
| Total | 17,603 | 100% |
Combining the two failure buckets, 3,494 tools — 19.9%, one in five — cannot currently be reached. We treat only the dead bucket as a confirmed death, for reasons set out under Limitations.
The real death rate: tools that worked, then stopped
A point-in-time count of broken links is not a death rate. A dead listing might have been dead before we ever imported it. The stronger measure is the set of tools our scanner has personally watched respond, and then watched stop.
Each tool carries a verified_alive_at timestamp — the last moment the scanner got a good response from it. 14,280 tools have one, meaning we saw them working at least once since the catalogue opened on 9 May 2026.
| Observed cohort | Count |
|---|---|
| Tools seen responding at least once | 14,280 |
| Have since stopped responding | 942 (6.6%) |
| — of those, now confirmed dead | 321 |
| — of those, now unreachable | 621 |
| Observation window | 103 days |
6.6% of the AI tools we personally verified as working stopped working within 103 days. That is the number to quote, because it has a denominator of confirmed-live tools and a measured window rather than an assumed one.
If that rate held steady for a full year it would work out to roughly 21% annually — but that is an extrapolation from a single 103-day window, not a measurement, and we would not defend it as one.
The month each casualty was last seen alive shows the attrition is continuous rather than a single import artefact: 798 last responded in May, 58 in June, 9 in July, and 77 in August. The May figure is inflated because tools checked early and never recovered stay pinned to their first month; the later months are the steadier signal.
Where the deaths concentrate
The catalogue was assembled in two very different ways, and they behave nothing alike.
| Intake month | Tools | Dead | % dead | % dead or unreachable |
|---|---|---|---|---|
| 2026-05 (bulk import) | 16,138 | 2,129 | 13.2% | 21.4% |
| 2026-06 (daily discovery) | 791 | 12 | 1.5% | 3.0% |
| 2026-07 (daily discovery) | 169 | 0 | 0.0% | 1.2% |
| 2026-08 (daily discovery) | 505 | 0 | 0.0% | 2.2% |
The May figure is a bulk import of tools that were already being listed across the existing AI-directory ecosystem. When we inherited 16,138 tools that other directories were actively publishing, 21.4% of them did not resolve. That is a statement about the state of AI tool directories in 2026, and it is the single most useful thing in this dataset.
The later months are tools we found ourselves, at or near launch, through daily discovery across Product Hunt, Hacker News, Reddit and GitHub. Their near-zero death rate is not evidence that new tools survive better — it is mostly that they have had two months of exposure instead of four, and that the dead classification needs three consecutive failures to trigger. Do not read those rows as a survival curve.
Which categories churn hardest
To compare categories fairly, this table is restricted to the May import cohort, so every tool has had the same amount of time to fail. (Across the whole catalogue, newer categories such as AI agents look artificially healthy: only 63.4% of that category came in with the May batch, against roughly 90% for everything else.)
| Category | Tools | Dead | % dead | % dead or unreachable |
|---|---|---|---|---|
| Image generation | 1,095 | 185 | 16.9% | 26.9% |
| Customer support | 1,323 | 193 | 14.6% | 22.7% |
| Writing & content | 1,857 | 267 | 14.4% | 23.5% |
| Marketing & sales | 1,235 | 178 | 14.4% | 23.2% |
| Productivity | 1,812 | 224 | 12.4% | 19.1% |
| Education & research | 1,164 | 140 | 12.0% | 19.8% |
| Code & development | 1,250 | 149 | 11.9% | 17.9% |
| Chatbots & assistants | 1,255 | 148 | 11.8% | 20.5% |
| Data & analytics | 1,115 | 118 | 10.6% | 17.8% |
| Audio & music | 780 | 80 | 10.3% | 16.5% |
The spread runs wider than the top ten shows. HR & recruitment, at 557 tools, matches image generation on 16.9% dead. At the other end, AI infrastructure (686 tools) sits at 6.9%, foundation models (385) at 4.9%, and no-code/low-code (301) at 4.7%.
The pattern is consistent: the categories with the lowest barrier to entry churn hardest. Image generation, content writing and customer-support chat were the three easiest wrappers to ship in 2023–2024, and they are the three carrying the most broken links now. Infrastructure, where shipping anything requires actual engineering, holds up roughly two and a half times better.
Free tools die most often
Pricing model is recorded for every row. Restricted, again, to the May cohort:
| Pricing | Tools | Dead | % dead | % dead or unreachable |
|---|---|---|---|---|
| Freemium | 8,124 | 822 | 10.1% | 16.8% |
| Free | 5,822 | 960 | 16.5% | 26.6% |
| Paid | 2,192 | 347 | 15.8% | 24.8% |
Free tools die at 16.5% against 10.1% for freemium — a gap of more than half again. Fully paid tools sit almost as high at 15.8%, which is the more surprising half of the result: charging money did not protect a tool nearly as much as running a free tier that converts. Freemium being the healthiest bucket is consistent with it also being the largest, and with it being the model that a serious, funded product tends to pick.
Redirects: the quiet exits
The 2,876 redirects are worth separating out, because they are not deaths. 2,437 of them have both the original and the resolved URL recorded. Of those, 1,371 (56.3%) redirect within the same host — an http-to-https or www canonicalisation, entirely healthy. The other 1,066 (43.7%) land on a different domain.
Those off-host redirects are a mix: rebrands, acquisitions, and products folded into a parent site. The most common destination by a wide margin is the Chrome Web Store (81 tools), where the "website" turns out to have been a browser extension listing all along, followed by GitHub (11). We count none of these as dead, but if you are auditing a directory, an off-host redirect is worth a look — it is often where a product quietly stopped being its own company.
How current these numbers are
Freshness is the whole basis of this dataset, so it is measured too. Of the 15,462 tools not currently classified dead:
| Re-checked within | Tools | Share |
|---|---|---|
| 7 days | 7,312 | 47.3% |
| 30 days | 15,315 | 99.0% |
| 60 days | 15,412 | 99.7% |
The median tool was last checked 7.3 days ago; the 90th percentile is 28.4 days, and the oldest check in the live catalogue is exactly 30.0 days old. 53 tools have never been checked. The scanner processed between roughly 1,000 and 1,400 tools per day across the fortnight before this measurement.
Methodology
- Source. The
ai_toolstable of the NeedAnAI production database, read on 20 August 2026. All figures are single SQL aggregates over that table; the headline counts were cross-checked against the public read-only API, which returns the same 17,603 total and the same 2,141 dead. - How a tool is checked. An HTTP request to the tool's recorded URL. Failures are recorded with a reason (DNS failure, connection error, timeout, or an HTTP status code) and increment a per-tool failure counter, which resets on any success. A second, independent scraping engine is used as a cross-check on ambiguous results.
- How a tool becomes "dead." Three or more consecutive failed checks on separate scan runs. Every one of the 2,141 dead rows carries a failure count of at least three; none was classified on a single bad response.
- These classifications were automated; new ones require a person. The dead set measured here was classified by the scanner's failure rule alone, and 1,758 of the 2,141 dead rows are still queued for review at the time of measurement. The advantage is that no human judgement, and no commercial relationship, influenced whether a listing was marked dead; the cost is that a small number of misclassifications will be in these figures. Under the current pipeline the rule is stricter: automation can only flag a tool as a dead suspect, and a person confirms the verdict before it sticks — so the review queue drains over time rather than growing.
- Cohorts. Where a rate is compared across categories or pricing models, the comparison is restricted to tools imported before 1 June 2026, so every tool in the comparison has had the same observation time.
- Measurement date. 20 August 2026. Re-running these queries later will produce different numbers; that is the point of a continuously scanned catalogue.
Limitations
"Unreachable" is a softer signal than "dead," and we keep it separate. A DNS failure marks a tool unreachable on the first occurrence, because a domain that no longer resolves is usually definitive — which is why tools in that bucket average only 1.89 failed checks, and why just 325 of the 1,353 have reached three. Less defensibly, 116 of them carry a failure reason naming our own scraping service timing out; that is our infrastructure failing, not the tool. Treat the 19.9% combined figure as an upper bound and the 12.2% dead figure as the defensible one.
The catalogue is not the AI tool population. It is a large sample of what the directory ecosystem was publishing, plus what our discovery pipeline has found since. Tools that never got listed anywhere are absent, and tools that died before May 2026 were probably never imported at all — which means these figures, if anything, understate historical mortality.
There is no tool-age data. A launch date is recorded for 2 of 17,603 rows, so we cannot say anything about how long a tool survives after launch, or whether 2023 vintages died faster than 2024 ones. Any article claiming a death rate by launch year is not getting it from data like this.
The observation window is 103 days. Everything longitudinal here rests on that window. It is long enough to be a real measurement and short enough that the annualised figure is a projection rather than a result.
A dead site is not always a dead company. A domain migration handled badly, an expired certificate, or an aggressive bot filter can all produce three failed checks. We reduce that with a second scraping engine and a reset-on-success counter, but we do not eliminate it.
What to do with this
If you are choosing a tool: the 12.2% figure is a reason to check when a listing was last verified before trusting it, on any directory including this one. If a directory cannot tell you when it last checked, assume it has not.
If you are building one: the churn is concentrated in exactly the categories where the product is a thin layer over someone else's model. Image generation and content writing are 16.9% and 14.4% dead respectively; infrastructure is 6.9%. That gap is the clearest signal in the dataset.
If you are citing this: the two numbers that carry their own denominator and window are 12.2% of 17,603 catalogued tools are confirmed dead, and 6.6% of 14,280 tools verified as working stopped working within 103 days, both measured on 20 August 2026.