- In openllmrank's test of 75 answers from five AI engines, 51 named Next.js, Supabase, Vercel and Stripe together.
- Agreement fades below the core stack, and 7 of 8 named AI models were 2024-era versions. The sample is small and the publisher sells AI-visibility tracking.
- Check how AI engines describe your product and category, engine by engine, and fix the sources they read.
A founder starting a new app used to ask colleagues, read comparison sites or call a vendor. Now many start with a question to a chatbot. The answer shapes a purchase before any salesperson hears about it.
That is the idea worth taking from a new test by openllmrank. A shortlist of technology is increasingly written by a machine that reads other people's blog posts. The shortlist is consistent, which makes it persuasive. It is not necessarily current.
What the test found
openllmrank asked ChatGPT, Claude, Gemini, Perplexity and Grok five versions of the question "what should I build an app with?" Each version was asked three times. That gave 75 answers.
The engines did not offer a menu. The database pick was close to unanimous: 89% of answers pointed to Supabase. Hosting on Vercel came up in 84%, the Next.js framework in 83% and Stripe payments in 71%. All four appeared together in 51 answers. When the question asked directly for a stack, 42 of 45 answers named all four.
Where the agreement thins out
Below the core, the answers diverge. Resend was the most common email product, but it appeared in only 43% of answers. PostHog for analytics appeared in 39%, and Sentry for error monitoring in 25%. Gemini did not name Sentry once.
The engines also split on how to build. Gemini named an AI app builder or coding tool, such as Lovable, Bolt, v0 or Cursor, in 11 of its 15 answers. ChatGPT did so in 1 of 15.
Payments showed one oddity. Gemini named Lemon Squeezy, a merchant of record that handles sales tax, in 12 of its 15 answers. Every answer from ChatGPT, Claude and Perplexity skipped it.
The practical point: a blended average hides these gaps. Each engine tells a different story about the same market.
Why the answers look the way they do
An AI engine answers from the pages it has read, not from a live survey of what works. The report shows how much that matters. The sites the engines cited most were vendors' own blogs, not documentation. Makerkit's site, makerkit.dev, turned up as a cited source in 11 answers. The company sells starter code for Next.js and Supabase apps, and one cited page was its own guide to the 2026 SaaS stack.
Age shows up too. Only 8 answers went as far as naming a particular AI model version. In 7 of them, the version dated from 2024, such as GPT-4o-mini or Claude 3.5 Sonnet. The researchers suggest this is probably because many of the pages the engines read date from then. That is their inference, not a tested cause.
Most engines did not favor their own makers. On the question about building an AI app, Gemini recommended OpenAI and Anthropic models in all 3 of its answers and never its own Gemini. Grok's 15 answers never mentioned Grok or xAI. ChatGPT was the exception: it leaned toward OpenAI in 3 of 3 answers on that question, and named Anthropic in 1. The samples are small, so this is a direction, not a rate.
Read the limits before the headline
This is one snapshot from one publisher, with 15 answers per engine. Answers change as new pages appear.
Perplexity supplied 396 of the 807 cited links, so the source table mostly reflects its reading list. ChatGPT cited only 28 links across its 15 answers.
openllmrank also sells AI-visibility tracking, and it discloses that it runs on the same default stack the engines recommend. Both facts argue for reading the findings as a sample, not a market ruling.
Adoption data tells a separate story
Does the recommendation match what is deployed? Not necessarily. In WebPulse's September 2026 scan of the Tranco top 10,000 domains, a platform was detected on 2,491 sites. Next.js accounted for 791 of them.
That measures established, high-traffic sites. It says nothing about new apps or about what the engines caused. It is a reminder that AI recommendations and real adoption are different measurements.
What leaders should ask
First, ask your marketing or product team what each major AI engine says about your product and your category. Ask engine by engine, not as an average.
Second, ask which pages the engines appear to read. In this test, vendor blogs and starter-kit posts carried weight. If a competitor's post defines your category, that is a content problem you can fix.
Third, ask whether your technical choices rest on an AI-written shortlist. If an engineer cites an AI recommendation, check the version numbers. In this test, nearly every model version the engines named was two years old.
The report's authors see open ground in the lower layers. In their view, no vendor has yet claimed analytics, monitoring or email in AI answers. That is their judgment, and they sell tracking. For vendors in those categories, the more modest reading is that the engines often stop before reaching them, rather than recommending against anyone.
A recommendation can be consistent and still be out of date. Treat the AI shortlist as a first draft, and check it before you sign anything.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: openllmrank.io.
