Claude Chose Vercel 100% of the Time. That Should Terrify You.

Someone ran an experiment recently that I can’t stop thinking about. They asked Claude Code to build projects — a SaaS app, an API, a data pipeline — without imposing any constraints on which tools to use. Then they tallied up the recommendations across 2,430 responses.

The numbers are staggering.

Vercel100% of JavaScript deployment picks. Not 80%. Not 90%. One hundred percent. 
Stripe gets 91.4% of payment integrations. 
shadcn/ui gets 90.1% of UI component recommendations.
GitHub Actions gets 93.8% of CI/CD. Meanwhile, AWS gets zero primary picks.
Express.js — one of the most downloaded npm packages in history — gets zero. Jest, which is literally in millions of package.json files right now, gets a pathetic 4%.

You might read this and think: “Well, maybe these are just the best tools.” And maybe they are! But that’s exactly the point — and also exactly the problem.

This is not a Recommendation. This is a Monopoly.

I’ve been writing software for a while. I’ve seen technology preferences swing back and forth. I’ve watched Rails go from “the future” to “legacy.” I’ve seen Angular lose to React, seen Redux get dethroned by simpler state management. These transitions happened over the years, through countless blog posts, conference talks, job postings, and developer conversations.

They never happened at 100%.

When a human senior engineer recommends a tool, they might say, “I like Vercel because the DX is great, but honestly, Railway and Render are solid too.” A confident recommendation from a human comes with caveats, personal context, and acknowledged tradeoffs. Nobody who has actually shipped things recommends using a single tool 100% of the time for 100% of projects.

But Claude does. And when you’re a junior developer asking your AI assistant how to deploy your Next.js app, you don’t get a nuanced comparison. You get a confident, authoritative answer that sounds like received wisdom: Use Vercel.

You’re not getting a recommendation. You’re getting an outcome that someone, somewhere, would pay billions of dollars to manufacture.

The Ultimate Advertising Format Has No Ads

Let me be direct: this is advertising, and it’s the most effective advertising ever created.

Think about how advertising evolved. First came billboards and print ads — you saw them, you knew they were ads, you could consciously discount them. Then came native advertising — “sponsored content” disguised as journalism. Readers got confused, but eventually the disclosure requirements and the telltale “SPONSORED” labels gave it away. Then came SEO — companies flooding the web with content to rank higher, so their products appeared “organically” at the top of search results. Sophisticated users learned to be skeptical of organic rankings, too.

Now meet the final form: AI-mediated product placement.

There’s no sponsored label. There’s no suspicious list of keywords. There’s just your trusted AI assistant, confidently telling you that Vercel is where you deploy, Stripe is how you take payments, shadcn/ui is how you build components. The recommendations come wrapped in the same authoritative voice that just helped you debug a gnarly TypeScript error. There’s no seam, no visible join, nothing that says “this recommendation may have been influenced by training data economics.”

The user’s trust model — built on watching the AI correctly solve hard problems — transfers entirely to its product recommendations. It’s the most sophisticated trust-laundering machine ever built.

Who’s Already Gaming This

Here’s where it gets interesting and a little dark: companies are already thinking about this.

The mechanism is straightforward. LLMs are trained on text from the internet. If you want your product to get recommended more, you need your product to appear more prominently in that training data — in tutorials, in GitHub repos, in Stack Overflow answers, in blog posts. This is basically SEO, but for AI training corpora.

Vercel understands this deeply. Their developer relations strategy, their open-source contributions, their exceptionally well-documented deployment guides, their deep integration with Next.js (which they also sponsor) — all of this creates a massive gravity well in the data that AI models train on. Whether this was intentional “training data marketing” from day one or just good developer relations that happened to work out perfectly in the AI era, the result is the same: 100% capture.

Meanwhile, AWS — the actual dominant infrastructure provider that runs a huge fraction of the internet — gets zero primary picks from Claude. Think about that. AWS has more market share, more features, and more enterprise deployments. But it has worse developer experience, worse documentation for small projects, a steeper learning curve. In a world where a human expert is making the decision, AWS often wins on security and scale. In a world where a language model decides based on patterns in training data, AWS loses because junior developers write fewer enthusiastic blog posts about it.

The new moat isn’t infrastructure. It isn’t a feature. It’s the documentation density in training sets.

The Casualties Nobody’s Talking about

Express.js has been downloaded billions of times. It powers an enormous percentage of Node.js applications. It’s maintained, stable, and understood by virtually every Node developer on the planet. Claude Code recommends it zero times.

Jest is in more CI pipelines than you can count. It’s the testing framework that shipped with Create React App for years. Claude recommends it 4% of the time, with Vitest almost entirely replacing it, which launched in 2021.

Redux: zero picks. AWS: zero picks. Traditional managed databases from cloud providers: essentially invisible.

These aren’t bad tools. They’re just not AI-native tools — meaning they either predate the current era of intensive developer blogging, or their documentation lives in enterprise-facing contexts rather than the kind of enthusiastic developer tutorials that end up in training data. They’re being written out of the new developer stack, not because they lost a technical competition, but because they lost a documentation-density competition they didn’t know they were in.

This should concern everyone building developer infrastructure. The question “Is this the best tool?” is increasingly being replaced by “Does AI recommend this tool?” Those are very different questions with very different answers.

Your Technical Decisions Aren’t Yours Anymore

There’s a specific mechanism that makes this particularly insidious, and it’s worth naming directly: AI coding assistants don’t just influence which tool you initially choose. They influence which tools you become proficient in.

When you use Claude Code to build a project, and it consistently patterns your code around shadcn/ui, you start to internalize those patterns. You learn how shadcn/ui works because that’s the code Claude is writing for you. When you later make an “independent” decision about which UI library to use, you reach for shadcn/ui not because Claude told you to — you’ve long since forgotten that part — but because it’s the one you actually know, the one whose patterns feel natural to you.

This is subtle, in the way good advertising always is. The best ads don’t feel like ads. They feel like culture. They feel like your own preferences.

AI tool recommendations are becoming a culture faster than anything that’s come before. Not because they’re more persuasive, but because they’re more formative. They shape what you build before you’ve had a chance to form independent opinions.

What happens next

The next few years will involve a scramble among developer tool vendors to optimize for AI recommendation density. We’ll see more investment in tutorials, more open-source examples, and more carefully crafted documentation written with the implicit understanding that the audience includes AI training pipelines.

We’ll probably also see some vendors try to influence training data more directly — sponsoring AI-adjacent content, contributing to projects used as training examples, and, perhaps eventually, finding ways to influence model fine-tuning directly. The line between “good developer relations” and “training data marketing” will get very blurry very fast.

The companies that understand this earliest have already won a massive advantage. Vercel’s 100% capture rate in Claude’s recommendations is worth more than any advertising campaign they could have run. It will self-reinforce: developers learn Vercel because Claude recommends it; Vercel becomes even more dominant in developer mindshare; more tutorials are written about it; more training data accumulates; and more future models recommend it.

This is a flywheel, and it’s already spinning fast.

The research that triggered all this thinking is worth reading if you build developer tools or work in the AI space. The headline numbers — 100% Vercel, 0% AWS, 0% Express — are striking enough on their own. But the deeper implication is what keeps me up at night: we’ve built an extraordinarily powerful recommendation engine, embedded it into every developer’s workflow, and handed it out for free. Whoever shapes what that engine recommends shapes how the next generation of software gets built.

That’s not a product feature. That’s a civilization-scale lever. And right now, it mostly just happens to be pointing at whoever wrote the most enthusiastic blog posts in 2022.

The study referenced is “What Claude Code Chooses” by Edwin Ong and Alex Vikati at amplifying.ai, which analyzes 2,430 Claude Code responses across tool categories.