Marblism stopped being an app builder, and nothing broke to tell us

Marblism is still trading, still shipping and still growing. It has simply stopped being an app builder. Nothing broke, no domain expired, no page went dark, and that is precisely why this one took us so long to notice.

What we had drafted

Our candidate row described Marblism as generating "a complete Next.js and Prisma SaaS repository from a description, with auth, billing and an admin area already wired". The draft verdict called it "not really an app builder, and better for it", gave it the highest code ownership mark of anything we had measured, and warned that everything after generation was the developer's problem.

That was a fair description of a product that existed. It is not a description of anything you can buy today.

What marblism.com sells now

Read in a rendering browser on 21 August 2026. The page title is "AI Employees To Scale Your Business". The heading is the same claim in capitals. The product is a roster of seven named agents: Eva on the inbox, Sonny on social media, Stan on lead generation, Walter on websites, Penny on SEO, Rachel on reception, Linda on legal.

We counted keywords across the whole homepage, because the absence is the finding. Occurrences of "codebase": zero. "Repository": zero. "Boilerplate": zero. "Next.js": zero. "React": zero. "SaaS": zero. The word "code" does not appear on the page at all.

There is a near miss worth naming. One of the seven agents is a "Website Builder" called Walter. Search for a builder product and you will find that string and think the product is still there. It is a marketing site agent, not a code generator, and it produces nothing you could take away and host yourself. A keyword check would have passed. A capability check does not.

Why no automated check caught this

This is the part worth generalising, because we run those checks and they all came back clean.

A liveness probe asks whether the site answers. Marblism answers, with a 200, on a page over half a megabyte of markup. An uptime monitor asks whether the domain resolves. It resolves. A broken link crawl asks whether anything 404s. Nothing does. A logo capture asks whether the brand mark is still there. It is, unchanged.

Every one of those is a transport check. Every one of them passes on a company that has replaced its entire product. The only check that catches this asks a question no probe can answer on its own: does the thing on this page still belong in the category we are ranking. That has to be read by a person, or by something reasoning about the page's claims rather than its status code, and it has to be read on a schedule rather than once at intake.

We are adding it as a standing step. Every candidate and every published entry gets a category re-read, and the question is the blunt one: if we saw this site today for the first time, would we have put it in this index at all.

Why removal rather than a rewrite

We could have kept the row and rescored it against what Marblism does now. We are not going to, for the same reason we would not review a coffee shop in a car magazine. An index of AI app builders that carries an AI staffing product because the company used to make a scaffolder is not being generous, it is being wrong in a way that quietly devalues every accurate row next to it.

If Marblism ships a code generation product again, it comes back in through intake like anything else, and it gets measured on the same ten axes as everything else.

What this changes here

The ranking does not move. Marblism was never published, so there is no rank to reassign, no average to recompute and no axis score anywhere on this site that shifts. The index stands at 33 published builders.

The comparison URL we had listed at /compare/marblism-vs-cursor now redirects here permanently instead of returning a dead end. There was never a page behind it.

For the record

None of this is a criticism of Marblism. Pivoting away from a scaffolding tool towards agents that do ongoing work is a defensible read of where the money is, and 40,000 businesses on the new product suggests it is landing. The error was ours: we kept a row open for a product that its own maker had stopped selling, and our checks were built to notice a site going down rather than a company changing its mind.