There is a lot of writing about how to build AI products and very little about what people actually use when they do.

We have 132 AI products listed by independent makers, each with self-reported stack tags. This is what they were built with — not what a survey says people plan to adopt, but what shipped.

The Short Version

• Gemini is the most-used model provider, on 24 of 132 AI products.

• Vercel hosts 29% of AI products — the single most common tag.

• Firebase and Supabase are tied at 16 each, unlike the wider dataset where Firebase leads.

• Only about half of AI products name a model provider at all.

• Python is not the default. Next.js appears on 29 AI products, Python on 14.

Model Providers: Gemini Leads

Gemini's lead is roughly 2.7x over Claude among products that named a provider, and we think the explanation is mostly economic rather than technical.

Google's free tier is unusually generous. For someone validating an idea with no revenue, a free tier that survives real usage is worth more than marginal quality differences. The cost of being wrong about a model choice is low; the cost of a surprise bill during validation is not.

This does not mean Gemini is the best model for production. It means it is the cheapest place to find out whether your idea works, and most products in this dataset are still at that stage.

The Missing Half

Here is the finding that says the most about the current AI product landscape.

There are 132 products in the AI category. Adding up every model provider tag — Gemini, Claude, Google Antigravity, and the rest — gets you to roughly 40 mentions. Even allowing for products using several providers, more than half of AI products do not name a model provider at all.

Three explanations, probably all partly true. Some makers tag their stack loosely and skipped it. Some use a provider indirectly through a framework or a tool like Cursor. And some products are labelled AI because AI is what gets clicked, with the actual functionality being conventional software.

We cannot separate those three from tag data alone. But if you are building in this space, it is worth knowing that "AI product" describes a marketing category more reliably than it describes an architecture.

Infrastructure: The Same Stack as Everything Else

The notable thing here is how unremarkable it is. AI products are being built on the same stack as everything else: Next.js on Vercel with a hosted database.

The Python number is the one worth pausing on. Python dominates machine learning, and it appears on 14 AI products against Next.js's 29. That gap tells you what these products actually are — they are web applications that call an API, not systems doing their own inference. The hard maths happens on someone else's servers.

That is a reasonable architecture for a solo maker. It also means "AI product" and "machine learning product" have almost entirely separated as categories.

Firebase and Supabase Tie Here

In our wider dataset of 466 products, Firebase leads Supabase 65 to 41 — a comfortable margin. Among AI products specifically, they are tied at 16 each.

The likely reason is recency. AI products skew newer than the dataset as a whole, and Supabase has been winning a larger share of new projects for a while. Firebase's overall lead is partly a legacy effect: a decade of head start, and plenty of products that began life before Supabase was a serious option.

If you want a directional read on where new indie projects are going, the AI subset is probably the better signal than the overall number.

What Next.js Products Pair With

Across all 105 Next.js products in the full dataset, the most common companions:

Next.js plus Vercel plus Tailwind plus a hosted Postgres is close to a default. Nearly half of Next.js products are on Vercel, which is unsurprising given the same company builds both, but the concentration is still striking.

Here Supabase clearly beats Firebase, 18 to 10. The Firebase lead in the overall dataset comes substantially from products that are not on Next.js — mobile apps, older projects, and the .NET contingent.

Payments: Stripe, and a Long Tail

Two things stand out. Stripe's dominance is expected. What is not expected is how few products tag any payment processor at all — 19 out of 466.

That aligns with the pricing data: 81% of products here are free or freemium, and a product with no paid tier has no reason to integrate payments. The small payments footprint is not a gap in the data. It is the data telling you most indie products never get to the point of charging.

The merchant-of-record options — Paddle, Dodo, Creem, Polar — are collectively at 6 products. They solve a real problem for solo makers selling internationally, since they handle VAT and sales tax as the seller of record. Their low adoption here probably reflects how few of these products are charging at all rather than a considered rejection.

What This Means If You Are Building

Start with Gemini if you are validating. The free tier means you can be wrong about the idea without paying for the privilege. Switch providers later — the API surfaces are similar enough that migration is rarely the hard part.

You do not need Python. If you are calling an API, use the stack you already know. The 29 Next.js AI products against 14 Python ones is makers routing around a language they did not need.

The boring stack is the right default. Next.js on Vercel with a hosted database is what most of these products use, and being unremarkable is an advantage when you are one person debugging at midnight.

Decide whether you are an AI product on purpose. Given that half the AI category does not name a model provider, the label is doing more marketing work than technical work. That is not necessarily wrong, but it is worth knowing which one you are doing.

Method and Limitations

Based on 132 products in AI-related categories, drawn from 466 publicly visible products on BuiltByIndies. Stack tags are self-reported by makers at submission, from a picker plus free text, and we merged obvious duplicate spellings by hand.

Three caveats. Tags are optional, so absence is not proof of non-use — this affects the "missing half" finding most, and it is why we offered three explanations rather than one. Category membership is also self-assigned, so "AI product" means whatever the maker thought it meant. And our audience skews web-first, so a mobile or research-focused community would produce very different numbers.

You can browse products by tech stack or category, or see the full breakdown across all 466 products in our 2026 indie stack report.