State of AI App Building 2026: Data from 11,355 Generated Apps

AI app builders have moved from novelty to workflow: people describe an app in a sentence, and a working codebase — frontend, database schema, API — comes back in minutes. But most of what is written about this category is anecdote. This report looks at what people actually build when an AI writes the code, based on 11,355 projects, 124,755 generated code files, and 80,189 builder–AI chat messages on the VULK platform (November 2025 – July 2026).

Key findings at a glance

• The median builder ships their first AI-generated app 47 seconds after signing up. Time-to-first-app is now measured in seconds, not days.
• The median app prompt is just 364 characters — roughly 60 words. People describe apps in a short paragraph, then refine through conversation.
• The median build conversation is only 4 messages long; the top 10% of projects run past 25 messages, and the longest single project conversation reached 588 messages.
• 80.5% of AI-generated apps are React + Vite web apps; Flutter is the #2 platform at 5.6%.
• 62% of AI-generated apps include a SQL database schema — most "vibe-coded" apps are not static pages, they are data applications.
• 27% of apps get a provisioned backend (API + hosted database), but only 3.7% are deployed to a public URL — the biggest drop-off in the funnel is the last step.
• Only 1.6% of AI-generated apps are to-do lists. The tutorial-era cliché is dead: dashboards and admin panels (12.5%) are the #1 category.
• The average AI-generated project contains 20.7 files and roughly 1,500 lines of code; the largest single project reached 677 files.
• 65% of all generated code files are TypeScript (including TSX); SQL is the second most-generated language at 6.1%.
• The AI does most of the talking: 3.4 assistant messages for every 1 human message in build conversations.
• South Asia produces 47% of new AI app builders but 8% of revenue; North America produces 3.5% of builders and 25% of revenue — a 7× monetization gap between where AI app building is learned and where it is paid for.
• 30.5% of people who sign up generate at least one app; 29% of those come back to build a second, and the average builder creates 2.5 apps.
• Peak building happens 12:00–15:00 UTC on weekdays (Tuesday is the busiest day); only 26% of apps are created on weekends — AI app building is work-hours behavior, not a weekend hobby.
• 3.5% of all signups convert to paying customers.

What do people actually build with AI?

The stereotype of AI-generated software — to-do apps and toy demos — does not survive contact with the data. Classifying the natural-language prompts behind 10,994 active projects by topic:

Two things stand out. First, the #1 use case is internal-tool-shaped software: dashboards, admin panels, booking systems, finance trackers — the unglamorous applications that businesses previously bought or built by hand. Second, to-do apps rank 13th. When generating an app costs one paragraph of text, people skip the practice project and go straight to the thing they actually need.

A notable meta-trend: 3.1% of AI-generated apps are themselves AI apps — people are using AI to build AI assistants, chatbots, and agent frontends.

How much code does an AI app builder actually write?

Across 11,355 projects, the platform generated 124,755 code files totaling 769 MB of source code — approximately 9.5 million lines (based on a random sample of ~6,000 files averaging 76 lines each).

The average project:

• 20.7 files (median 20, max 677)
• ~127 KB of source code, roughly 1,500 lines
• 6.2 KB per file on average — AI generators produce many small, focused files rather than monoliths

The language distribution of generated files tells you what the modern AI-generated stack looks like:

TypeScript has effectively won the AI codegen era. Two-thirds of everything the models emit is typed. And SQL being the #3 output language reinforces the database point below.

Do AI-generated apps have real backends?

Yes — and this is probably the most underreported shift in the category.

• 61.7% of projects include a SQL database schema among their generated files.
• 27.1% of apps have a provisioned backend — a live API worker plus a hosted database, not just frontend code.
• The remaining apps are genuinely frontend-only (portfolios, landing pages, games).

The popular framing of AI app builders as "UI toys" is outdated: the majority of generated apps model data, and more than a quarter run server-side infrastructure.

The funnel narrows sharply at the end, however:

Only ~1 in 27 AI-generated apps ships to a public URL. Generation is no longer the bottleneck — the last mile (deployment, domains, going live) is where most projects stop. For the ecosystem, this is the clearest product gap the data reveals.

What platforms do people build for?

Detecting the platform from each project's actual files (build configs, manifests):

The web still dominates output — but demand runs ahead of delivery. Looking at what users ask for in prompts (all-time platform-demand mentions on the same platform): Flutter leads with 381 requests, followed by Node backends (310), Next.js (265), Python (246), React Native (213), and PHP (174). Mobile and backend-first demand significantly exceeds what currently gets generated, which suggests the next phase of AI app building is multi-runtime: mobile binaries, Python services, and PHP apps, not just React SPAs.

How do people talk to an AI that builds apps?

• The median initial prompt is 364 characters — about 60 words, or 3–4 sentences. The mean is far higher (1,983 characters) because a long tail of users paste entire specifications, briefs, or existing code; the single longest prompt was over 500,000 characters.
• The median project conversation is 4 messages. Half of all apps are essentially "one shot plus a tweak."
• The top 10% of projects have 25+ messages — these are the users treating the AI as an iterative pair-programmer, not a vending machine. The record: 588 messages on a single project.
• Across all conversations, the AI sends 3.4 messages for every 1 human message (61,977 assistant vs 18,212 user messages) — reflecting the streaming, multi-step nature of code generation.

The pattern that emerges: short intent in, long iteration loop out for the projects that matter. Prompt length is not where the skill lives; knowing what to ask for next is.

Who is building AI-generated apps — and who is paying for them?

Geography is where the data gets most interesting. Among signups with a known country:

• South Asia (India, Pakistan, Bangladesh, Nepal, Sri Lanka): 47.4% of builders — India alone is 25% of all signups, the single largest country.
• Africa: 15.9% (Nigeria, Algeria, Egypt, South Africa, Kenya, Morocco all in the top 15 countries)
• Europe: 7.9%
• North America: 3.5%

Revenue distributes almost inversely:

• North America: 24.8% of revenue (from 3.5% of signups — a ~7× overindex)
• Europe: 24.4%
• Gulf states: 12.5%
• South Asia: 7.7% (from 47.4% of signups)

The interpretation: AI app building is being adopted as a skill in emerging markets and purchased as a tool in developed ones. The largest builder populations are in markets where $20/month is a meaningful price point; the paying customers cluster in the US, Germany, the UK, France, the Netherlands, Switzerland, and the Gulf. Any AI dev-tool company will recognize this shape — it is the defining go-to-market tension of the category.

Overall, 3.5% of all signups become paying customers — notable for a product category where the free output (a working codebase) is itself the value.

When do people build?

Excluding automated/batch activity:

• Peak creation window: 12:00–15:00 UTC, squarely inside the European/African/South Asian workday and the US morning.
• Tuesday is the busiest day; Monday and Sunday are the quietest.
• Weekends account for only 26% of app creation.

AI app building shows a work-tool usage curve, not an entertainment curve. People generate apps during working hours, on working days — consistent with the dashboard/admin-panel dominance in the "what people build" data.

How fast is the first app?

The median time from account creation to first generated app is 47 seconds. This is arguably the category's defining number: the distance between "I have an idea" and "I have running code" has collapsed to under a minute.

Activation and retention:

• 30.5% of signups generate at least one app (4,622 of 15,165 users)
• 29.5% of builders return to create a second app (1,362 users)
• 181 power users have built 5+ apps; the average builder creates 2.5 apps
• Platform activity is spiky: half of all projects were created in the December 2025 – January 2026 launch window, with organic monthly creation re-accelerating from May 2026 onward (June 2026 was the strongest organic month since January, up 2.2× over May).

Methodology

This report is based on aggregated, anonymized production data from the VULK platform (vulk.dev), an AI-powered application builder.

• Dataset: N = 11,355 generated projects (10,994 active), 124,755 generated code files, 80,189 chat messages, 15,165 registered users, 576 payments from 533 customers.
• Period: November 2025 – 17 July 2026.
• Platform detection is derived from files actually present in each project (e.g., vite.config.*, pubspec.yaml, requirements.txt, next.config.*, composer.json), computed over the 6,025 projects with retained file trees.
• Category classification uses keyword matching on project prompts; categories are not mutually exclusive.
• Lines-of-code figures are estimated from a random sample of ~6,000 files (TABLESAMPLE); byte totals are exact.
• Time-of-day and day-of-week figures exclude identified batch/automated creation events (days where a single account created >200 projects — one such event, 1,569 projects on 2026-07-16, was excluded from behavioral statistics; headline totals include all projects).
• Geographic figures use country at signup; signup-share percentages are computed over users with a known country (78.4% of accounts). Revenue shares are computed over completed payments.
• No individual user data, prompts, emails, or personally identifiable information were accessed or reported; all statistics are aggregates.

This report may be cited freely with attribution to VULK (vulk.dev). For questions about this data or the methodology, contact the VULK team at vulk.dev.