How this site works
01What qualifies
The archive records dead or abandoned products. A product that survived a public meltdown goes to Incidents; a living product with source-backed concerns goes to the Watchlist. An acquisition where the product lives on is not a fail.
02What moves an entry to the Archive
From the Watchlist
Two independent sources must establish a shutdown, discontinuation or abandonment. We update the fail type and outcome, record the change and score the Archive entry. When concerns resolve, we retire the watchlist entry with its reason visible.
From Incidents
If a product later dies, its entry moves to the Archive with the sourced outcome and a Fail Score. A resolved incident alone is not a death.
03The evidence ladder
Primary record. The source is an original record such as a court filing, regulator order, company statement, or securities filing.
Major outlet. The source is independent reporting by a major news outlet.
Specialist reporting. The source is credible trade press, specialist reporting, or a newsletter.
Individual reporting. The source is credible individual reporting, a video teardown, or an archived post.
Every key claim needs two independent sources, including at least one E1 primary record or E2 major outlet. Every money figure needs an E1 or E2 source. Syndicated copies and outlets quoting the same press release count as one reporting group. Every source records a supporting quote, access date, archive URL and the claims it supports. E4 alone never supports a factual claim about a company.
04How similar fails and Idea Check match
Similarity = 0.6 × embedding cosine + 0.4 × structured overlap. Overlap adds category +1, primary cause +2, form factor +1, user type +1, era +0.5 and shared idea cluster +2, divided by 7.5. Missing dimensions add nothing. A match percentage is this similarity score expressed as a percentage, not a probability of failure.
Idea Check retrieves candidates using full-text and embedding matches, then applies this formula against the idea’s inferred dimensions. A published idea cluster matches above 0.65 cosine similarity; the highest matching cluster supplies up to three most recent recorded deaths and their primary causes for the timing question. Timing verdict, sentiment and Fail Score never filter or score these matches. Incident and watchlist matches keep their section and outcome labels. Warning signs in an idea must quote the supplied text; signs recorded in neighbouring entries are shown separately.
05The Fail ScoreOur take
Fail Score is an editorial notability measure: how big a story the fail was. It adds four 0–25 scores for hype, money, spectacle and lesson value. Severity measures harm separately; evidence confidence measures sourcing. Research results rank by relevance and evidence confidence. Incidents are unranked; watchlist entries are unscored.
| Sub-score | 0–5 | 6–12 | 13–19 | 20–25 |
|---|---|---|---|---|
| Hype | Quiet launch | Trade press | Mainstream press, keynote, celebrity | ‘iPhone moment’ claims, Super Bowl ad, presidential mention |
| Money | Under $1M | $1M to $25M | $25M to $250M | $250M+ or public-company write-down |
| Spectacle | Quiet shutdown notice | Trade press post-mortem | Mainstream coverage, memes | Global news cycle, congressional or regulatory attention |
| Lesson | Idiosyncratic, nothing to learn | One clear lesson | A pattern others repeat | A canonical case study |
06The money counter
Only disclosed raised and write-down figures for archive entries with outcomes shut down, acquihired and killed, or product discontinued where spend was disclosed. Incidents and watchlist items never count. Settlements, refunds, fines, market-cap drops and estimates may appear on cards with their method labels but are never summed. Reported currencies are kept separate.
07Warning signs
A fixed checklist of 19 observable signs, each recorded with the date observed and a source. This records what was visible; it is not a prediction engine.
- Demo Only Launch
- The launch showed a demonstration without making a usable product available.
- Waitlist Over 6 Months
- Customers remained on a waiting list for more than six months before gaining access.
- Announce To Ship Over 12 Months
- More than twelve months passed between the product announcement and shipping.
- Hardware Requires Subscription
- The device requires a paid subscription for its essential functions.
- Features Marked Coming Soon At Launch
- Promoted features were still labelled as future additions when the product launched.
- No Public Pricing
- Potential customers could not find a public price for the product.
- Humans In The Loop Undisclosed
- People helped produce supposedly automated results without clear disclosure to customers.
- Benchmark Claims Not Reproduced
- Independent testers could not reproduce the performance claimed in benchmarks.
- Exec Departures Pre Death
- Senior leaders left before the product or company ended.
- Restated Metrics
- The company materially revised previously reported business or performance figures.
- Down Round
- The company raised funding at a lower valuation than its previous round.
- Pivot Count 2 Plus
- The company changed its core product or business direction at least twice.
- Refund Complaints At Scale
- Many customers publicly reported difficulty obtaining promised refunds.
- AI in the company name
- The company explicitly included AI in its public name.
- Founder Hype Statements
- A founder publicly described the product with unusually sweeping promises.
- Regulatory Warning Received
- An authority warned the operator about the product before its failure or incident ended.
- Single Platform Dependency
- The product relied on one external platform for an essential capability or distribution channel.
- Celebrity Or Influencer Launch
- The launch relied on a celebrity or influencer to attract initial attention.
- Raised Over 100m Pre Revenue
- The company raised more than $100 million before earning product revenue.
08Taxonomy
Every entry has one tier, fail type, outcome and category, with one primary cause leaf and up to two secondary causes. The primary cause is the one that would have killed it alone.
tier
- Titan
- The product comes from a major AI lab or a public company.
- Funded
- The company raised at least $10 million or received major mainstream press coverage.
- Indie
- The product launched publicly with little funding and has two independent sources establishing its notability.
fail type
- Killed
- The product shipped and was then shut down or abandoned.
- Vaporware
- The product was announced but never shipped.
- Faked Demo
- A demo or claim was shown to be staged, misleading, or dependent on undisclosed human work.
- Meltdown
- The product caused a live public embarrassment.
- Collapse
- The company failed through insolvency, a fire sale, or an acquisition that killed its product.
- Pivot
- The company survived while abandoning the original product or use case.
- Watchlist
- The product is still active but has source-backed signs of trouble.
category
- AI Hardware
- The product is a physical device sold around an AI capability.
- Chatbots Assistants
- The product helps users through conversation or personal assistance.
- AI Agents
- The product uses AI to carry out tasks across tools or workflows.
- Coding Agents
- The product uses AI to write, change, or operate software.
- Autonomous Vehicles
- The product uses AI to drive a vehicle with reduced human control.
- Robotics
- The product uses AI to control a robot that performs physical tasks.
- Enterprise AI
- The product applies AI to an organisation's business operations.
- AI Content Media
- The product uses AI to create, select, or distribute media.
- Consumer Apps
- The product is an AI application for everyday personal use.
- Public Sector
- The product applies AI to government services or public administration.
- Healthcare AI
- The product uses AI to support health advice, diagnosis, or care.
- Adjacent Hype
- The product is not strictly AI but shares the archive's pattern of technology hype and disappointment.
See the cause glossary and warning-sign checklist below for the full reference vocabulary.
09Cause glossary
No Demand
Too few customers wanted the product enough to sustain it.
Tech Not Ready
The available technology could not reliably deliver the product's promised job.
Unit Economics
The cost of serving customers could not be sustained by the revenue they generated.
Trust Safety
The product lost viability because it could not protect users or earn their trust.
Regulation Legal
Legal obligations or regulatory intervention made the product unviable or forced a material change.
Timing
The product reached the market at a time when the opportunity could not sustain it.
Execution
Decisions or delivery failures inside the company prevented the product from succeeding.
Overpromise
Public promises exceeded what the product could demonstrably deliver.
Faked
The product's apparent capabilities or business results depended on material misrepresentation.
Ethics Backlash
Objections to the product's social impact damaged its acceptance or operation.
Platform Risk
Dependence on another company's platform exposed the product to changes it could not control.
10Status vocabulary
- Shut Down
- Both the product and the company are gone.
- Product Discontinued
- The product is gone while its company remains active.
- Acquihired Killed
- An acquirer took the team and the product was shut down.
- Absorbed Into Platform
- The product's function now lives inside an incumbent's platform.
- Pivoted Away
- The company remains active but no longer pursues the original product or use case.
- Zombie
- The product is technically alive but has no meaningful activity, as supported by two sources.
- Incident Resolved
- The product survived a public meltdown and the incident has been resolved.
- Ongoing
- The product remains active and belongs only on the watchlist.
11Right of reply & corrections
Named companies and founders can respond; replies are published on the entry, unedited, above the fold. Report a factual error or request a reply through corrections. Requests are reviewed within 48 hours; credible complaints can lead to temporary unpublication or correction, with dated public changes.
12Open data
Card-level fields and sources are available under CC BY 4.0. Narratives stay on-site. Download the data → or build with the API →.
The joke is free. The facts cost us a methodology page.