Most AI-platform adoption numbers are inflated. We counted the real ones.
Platform footprints are usually tallied by keyword mentions, which overstate genuine built-with adoption by anywhere from a few times to five hundred. We fingerprinted 505,406public repositories for the real markers, a lovable-tagger dependency, a .replit file, a langgraph import, and counted what is actually built with each AI app platform. Because those confirmed counts are conservative floors that miss private and deleted apps, we lead with the relative picture: each platform's share of new built-with apps, and its momentum against GitHub's own growth, month by month.
The clearest case of why keyword counts mislead. The name “openclaw” collides with a well-known open-source video game, so 25,940 repositories matched the term in March 2026 while only 51 carried a real build signature. Pisama covers OpenClaw as a framework, and the honest number is 51.
Inflation tracks name distinctiveness. Platforms with a unique fingerprint (Streamlit, the .replit file) sit near their real count; platforms with common names or noisy keywords run far ahead of what is actually built with them.
Shares are robust where absolute counts are not. Streamlit still leads but its share is falling as Replit and LangGraph climb. “Now” is the latest complete month; “a year ago” is the first. The multiple is each platform measured against its own start.
Measuring each platform against itself sidesteps the undercounting, but GitHub itself nearly doubled over the year, so the dashed baseline is the real bar to clear. Replit cleared it about nine times over and LangGraph nearly twice; n8n and Lovable merely kept pace with GitHub, while Streamlit and Gradio grew slower than the platform at large. Across every tracked platform, confirmed built-with still grew about 2.3 times against GitHub's 1.8, a slim but real gain in share of everything created. Platforms with too small a starting base to index are omitted.
Raw discovery counts, shown for scale and context rather than ranking. The per-platform adoption story is in the relative exhibits above.
Discovery. De-floored, date-windowed GitHub search across about 90 platform queries, 505,406 public repositories captured with their creation dates.
Built-with.Each candidate repository's files were fetched from the raw content CDN and confirmed only on a real fingerprint: a lovable-tagger dependency, a .replit file, a .bolt directory, a streamlit or langgraph import, or a platform project backlink.
Why relative. Confirmed counts are survivor lower bounds. Deleted or private repositories cannot be checked, and platforms whose apps are hosted off GitHub are undercounted. Shares and momentum are far more robust than absolute counts, so the charts lead with those.
The GitHub baseline. Total public repository creation on GitHub grew about 1.8 times over the year, from 4.9 million new repositories in June 2025 to 8.8 million in May 2026 by the same dated search. Momentum is read against that line, so outpacing the baseline is what counts, not simply rising with it.
The keyword trap. openclaw is the worked example: 25,940 keyword matches, 51 confirmed built-with, because the name collides with a video game.
This report exists because Pisama measures what software actually does, not what it claims. We build failure detection for multi-agent LLM systems, and we publish per-detector F1 scores and an open calibration dataset in a category that mostly publishes round numbers. The same discipline produced these counts. A fingerprint in the code, or it does not count.
See the per-detector F1 scoresCite as: Pisama Built-With Report, June 3, 2026 (pisama.ai/stats). Press and data requests: [email protected].
Source data: ai-discovery-production · signature-confirmed built-with + de-capped catalogue. Aggregate, anonymized counts only. No individual app, author, or contact is identified.