State of AI PRDs 2026: what 988 real PRDs tell us about how teams ship with AI
This report is built on aggregate analysis of 988 completed PRDs generated through MakeMyPRD between launch and 2026-05-15. The data is anonymized: no business-idea text, user identifiers, or PRD bodies are persisted in the underlying dataset. What we kept is the shape of the work: what categories of products teams are building, which AI builders they mention, and how the patterns differ between indexable (high-quality) PRDs and the long tail.
What people are actually building
The category mix in 2026 tilts heavily toward SaaS and mobile apps, with AI agents and EdTech showing earlier than expected. About half of submitted ideas don't pattern-match cleanly into the 10 categories we classified for — they're niche products, internal tools for specific teams, or experiments without a tidy label. That long tail matters: most "Other" PRDs are products you've never heard of, which is exactly what indie builders are supposed to be making.
The strong showing for SaaS (23%) tracks with the broader market — most builders are still building for businesses. Mobile (8%) is healthier than three years ago, when no-code mobile was a dead category. AI agent products at 3% surprised us; we'd have guessed double that based on Twitter. The actual share is lower because most "AI agent" descriptions get captured as SaaS instead (an AI customer support tool is technically both, and our classifier votes SaaS).
Which AI builders show up in PRDs
Builders mentioned by name in the business-idea text. Note that absence doesn't mean unused. Most people describe what they're building without naming the tool, especially early in the funnel. The ones who do name a tool tilt toward Lovable and Cursor.
Lovable's lead in name-checks aligns with its CRUD-app strength: people building dashboards and SaaS tools often start with "I want to build a Lovable app for…" Cursor's mentions trend in a different direction: longer, more technical PRDs where the user has already decided on the stack and is asking for a spec the agent will execute file-by-file.
The top public PRDs
The 20 most-viewed indexable PRDs, anonymized to slug + view count. These are the queries that bring people to MakeMyPRD via Google and AI search citations.
The top-20 cluster around a handful of evergreen patterns: time tracking, task management, learning apps, and small B2B tools. None of them are novel categories. The signal isn't "what's new" — it's "what do solo developers and small teams keep wanting to build." Boring categories that work.
What we noticed in the data
Three patterns surfaced as we built the report:
Short ideas produce short PRDs. The median idea length is short (77 words). The corresponding PRDs land in the 800-to-1,500 word range that the quality pipeline targets. Longer ideas (300+ words) generate longer PRDs, but the marginal value of the extra length tapers around 1,800 words. The structural cap matters more than the input length.
SaaS and EdTech over-index on quality. Of the indexable PRDs, the SaaS and EdTech categories have higher view counts per PRD than the average. We think this is because the SaaS and EdTech audience is more likely to share PRDs publicly (founder communities, university Slack groups) than the marketing-site audience, who tend to keep them internal.
The "Other" bucket is the future signal. Half of all PRDs don't fit a tidy category. That's not noise. It's where indie builders are exploring before category lines have formed. A year from now, several "Other" subclusters will probably become their own categories.
A real example of what the data is built from
Snapshot of one PRD entry in the dataset (anonymized to what we actually keep):
- Category classification: "SaaS"
- Builder mentions: ["Lovable", "Supabase"]
- Idea word count: 142
- Generated PRD word count: 1,287
- Quality score: 0.91
- Indexable: true
- View count: 18
What we do NOT keep, by design:
- The business idea text
- The generated PRD content
- The user ID
- The user's email
- Any free-form output
Why this matters for citations:
Researchers and analysts citing this report can verify the methodology in the open-source repo (the ETL script is checked in) and confirm no private information powers any of the numbers. The privacy gate runs at script-write time and would have failed loudly if any forbidden field made it into the output JSON.
Aggregate guarantees:
- All counts are direct database aggregates, not modeled estimates
- All percentages are computed from those counts at write time
- The slug+shortId pairs in the top-20 list are already-public URLs (visible on the indexable PRD pages)
- Re-running the ETL produces a fresh JSON; the methodology is the contract, the values are point-in-time
Methodology
The full extraction script is at scripts/research/state-of-ai-prd-2026.ts in the MakeMyPRD repo. Anyone can run it against the production database to reproduce these numbers. The privacy assertions are enforced at script-write time, so any change that would leak protected content fails the build.
What we'll watch for the 2027 report
Three measurements we'd like to add next year:
- Builder mention shifts. If Lovable's name-check share grows, drops, or holds, that's a signal about category maturity.
- The shape of "Other". A year from now we should be able to identify 2 to 3 new categories that have crystallized.
- Quality score vs view count correlation. Right now the correlation is noisy. With more data we'd be able to say whether the quality pipeline actually predicts real-world engagement.
FAQ
How accurate is the category classification?
Coarse. We use rule-based pattern matching (regex on common keywords), which is fast and reproducible but misses nuance. About 5 to 10% of classifications are arguable — for example, an AI tool for customer support gets tagged "SaaS" rather than "AI agent." The aggregate shape is reliable; individual PRD tags are not. This is a deliberate trade-off: rule-based is auditable, LLM-based isn't.
Why no per-user data?
Privacy by design. The MakeMyPRD product lets users generate PRDs without sharing aggregate behavior. The report uses only counts, percentages, and slugs that are already public. No row-level user behavior is part of the input.
Can I cite this report?
Yes. The data was aggregated on 2026-05-15 from 988 PRDs in the MakeMyPRD production database. Cite as "MakeMyPRD, State of AI PRDs 2026, accessed [date], https://makemyprd.com/research/state-of-ai-prd-2026." The numbers will update when we re-run the ETL.
Will the numbers change after publication?
Yes. The ETL script can be re-run quarterly. When we update, we'll bump the updated date in the page metadata and rev the JSON. Older numbers are recoverable from the git history of the data file.
How do these numbers compare to ChatPRD or other PRD generators?
We don't have public data from competitors to compare. ChatPRD has cited "thousands of PRDs" generated; specific aggregate stats aren't published. The point of this report isn't a head-to-head — it's giving the community a measured view of how the AI-PRD audience is actually using the tool, on the basis of a real corpus.