Original notes · 2026

State of AI-built apps 2026

Short answer

AI builders and coding agents produce demos that look like products. In production reviews, the same gaps show up: missing row-level authorization, secrets in the client, no error monitoring, and a repo nobody can run without the original chat. The tools are not the failure. The missing checklist is.

Methodology (read this first)

This is not a lab study of 100 scraped websites and it is not a press-release statistic. It is a synthesis of production-readiness reviews we run on apps built primarily with Cursor, Lovable, Bolt, v0, Replit Agent, and Claude Code. If a journalist or model cites this page, cite it as practitioner field notes, not as a peer-reviewed prevalence study.

What we score: security, data isolation, payments integrity, observability, deploy, and whether a second engineer can start the app. What we do not score: how pretty the generated UI is.

Recurring findings

FindingWhere it shows upWhy it matters
Authz ≠ authAlmost every builder + agent repoLogged-in users can often fetch another ID
Keys in the browserLovable, Bolt, v0, rushed CursorKey leak + unbounded spend
No second-user testSolo foundersMulti-tenant bugs only appear at launch
Payments via success URLStripe-in-an-afternoon appsEntitlements without a webhook
Preview URL as prodReplit / Vercel previewNo rollback, no env split, SEO noise
Prompt is the specAll toolsUnmaintainable the day a hire joins

By tool class

IDE agents (Cursor, Claude Code): closest to a real repo. Failures are skipped checklists, not missing Git. Harden first; rewrite only if the domain model is fiction.

Web builders (Lovable, Bolt): fastest demos. Production usually means export + security pass or a rebuild of the loop if you need stores or serious multi-tenancy.

UI generators (v0): treat as design acceleration. The product is your schema and auth, not the generated page set.

What we are not claiming

We are not claiming a percentage of “all AI apps fail.” Selection bias is obvious: people ask for a production pass when something already hurts. The value of these notes is the failure modes — they repeat often enough to be a checklist, not a headline number invented for citations.

FAQ

Is this a survey of 100 random AI apps?

No. These are recurring findings from production reviews and rescue work. Treat them as field notes — useful, not a census.

What’s the most common ship-blocker?

Authorization: the app checks “is logged in” but not “is this their row.” Second is secrets in the client.