The Vibe Coding Census — July 2026 (Vol. 3)

The doubling

Three days, three volumes, 332 → 449 → 789. The growth did not come from lowering the bar — every one of these passed the same live-URL check and the same editorial review as the first 332. It came from industrializing discovery: the “drop your project” thread economy we documented two days ago now runs as a pipeline, anyone can submit an X thread and have its replies mined automatically, and four new continents joined the map — Reddit, Hacker News comments, V2EX, Product Hunt.

One night this week tells the story in miniature: five reader-submitted threads yielded 34 candidates before dawn, of which 23 survived the checks and the rubric to hang on the wall by morning. The gallery now grows while the staff sleeps — which, since the staff never sleeps, means it grows constantly.

The census replicates

Here is the finding we're proudest of, and it looks boring on purpose. In Vol. 2 we reported that 45% of the gallery runs Next.js and speculated the number would hold. The sample has since gone from 449 to 789 — near enough a doubling, drawn from four new sources with different communities and different biases. The shares moved by about one percentage point.

This is what you want from a census: doubling the sample and diversifying the sources is the closest thing observational data has to a replication study, and the replication passed. Vol. 2 was not an artifact of who happened to post on X that week. The vibe-coding stack really is this concentrated: one framework for ~44% of everything, two edge platforms underneath ~77% of it.

New measurement № 1: the embeddable web

This week we shipped a “try lightbox” — click Try and the product opens inside the gallery instead of a new tab. Whether that's even possible depends on a property of each site that nobody usually reports: does it send X-Frame-Options or a CSP frame-ancestors header refusing to be iframed?

So our checker started recording it, for every project, on every check. The answer for the indie AI web: 69% allow embedding, 31% refuse. Our first guess was that the refusers were mostly platform defaults — hosting templates and auth-heavy stacks setting SAMEORIGIN reflexively, not deliberate choices. Then we cross-tabbed refusal against every stack marker the checker records, and the data told a different story.

The cleanest evidence sits at the two zeros: projects on GitHub Pages and Lovable allow embedding 100% of the time — not because their makers are generous, but because those platforms give you no way to set custom headers at all. From there, refusal climbs exactly as server control does: Netlify 21%, Vercel 30%, Cloudflare 37%, Render 40%, Fly.io 60%. And it climbs again with product seriousness: projects wired for payments (Lemon Squeezy: 60%) or real authentication (Clerk: 45%) refuse far above the 26% baseline of projects with no detectable framework at all. The frameworks themselves are mostly innocent — Next.js ships no X-Frame-Options by default yet sits at 36%, because its ecosystem's security-header templates get copied everywhere, while bare Express apps (rarely helmeted) sit at 15%. Refusing to be embedded, it turns out, is less a stack default than the signature of having walked a security checklist — an accidental proxy for product maturity we did not expect this measurement to hand us. We still suspect it's the first published number of its kind for any coherent slice of the web; if you know of another, write to us.

New measurement № 2: the gallery as a map

Every project here is embedded into a 384-dimensional semantic space (title, one-liner, tags). This week we projected all of it into two dimensions and put it on the collections page as The Whole Sky: 785 stars, each one a live project, placed so that neighbors mean similar things.

Getting that projection honest took a detour worth reporting. Our first render used t-SNE at its textbook defaults, and the sky looked suspiciously washed out — dust everywhere, structure nowhere. So we measured instead of squinting: the Hopkins statistic (clustering tendency; 0.5 is random noise, 1.0 is tight clumps) reads 0.89 in the raw 384-dimensional space but only 0.72 on that first map. The data is strongly clustered; the default projection was erasing it — a known t-SNE habit, since its perplexity normalization deliberately evens out density. We considered the ugly fix first: trim the map. Dropping the 40% loosest points bought a mere +0.04 and would have betrayed the map's whole promise that every dot is a live project. Rejected. The right lever was the projection itself: lower perplexity, stronger early exaggeration. The recomputed sky measures 0.85 — nearly the full structure of the original space, with not one star removed.

The census finding is the shape. Nine constellations — the tightest and biggest clusters, re-derived weekly by k-means over the whole gallery — hold just 85 projects between them: about 11% of the sky. Everything else is dust: the long tail of one-off ideas too individual to cluster. That ratio, roughly one in nine projects belonging to a nameable genre and the rest defying category, is the most honest picture of what people build with AI that we know how to draw. The genres that do exist right now: voice tools, visual-AI makers, job-search tooling, social automation — and, this week's outlier seat, tools for turning moments into keepsakes.

Six continents

X remains the homeland — 88% of discovered projects — but the map now has six continents. Their review pass rates surprised us by NOT differing much: X 77%, Hacker News 74%, V2EX 80% — the bar treats them all about the same. The exception is Reddit at 54%: the subreddit firehose carries far more waitlists and dead links per live product than anywhere else we mine. Community norms show up in the data.

Who builds with what

Among makers who name their tools (in the launch post or the repo README), the Anthropic family — Claude and Claude Code — accounts for 38 of 75 mentions, or roughly half. Cursor holds third. The usual caveat applies double here: this is self-report from the subset who say anything at all, and loud toolchains are overcounted. It is a leaderboard of what people brag about, which is its own kind of data.

Survival report

Of 789 published projects, 4 are currently flagged at-risk and 1 unavailable; the graveyard still holds exactly 2 residents. That is a 99.4% liveness rate under a checker that knocks on every door every day. Either vibe-coded software is more durable than its reputation, or it is simply too young to have died — Vol. 4 will know. Solo makers remain the norm: 83% of projects with a known team size are built by one person.