November 21, 2026
Heard in AI

Supabase says its lead in AI model recommendations was an accident

On a Latent Space episode, an AEO tracker cited in the conversation was described as showing every frontier model naming Supabase first for an integrated backend. Supabase says the lead came from lucky timing, and co-founder Ant Wilson credits community content that he believes models absorbed.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on 2 episodes of Latent Space, published October 22–28, 2026

If you ask an AI model which service should provide the backend for a new app, the answer is Supabase, according to an AEO tracker cited on the Latent Space podcast. In an episode published October 28, 2026, host swyx spoke with Supabase co-founders Paul Copplestone and Ant Wilson. During the conversation, the tracker was described as showing "every single frontier model" recommending Supabase "as your primary choice for an integrated backend."

AEO stands for answer engine optimization. It is the work of shaping what AI chatbots and agents say when someone asks them which product to use. A backend is the part of an app that users never see. In the conversation, "integrated backend" was defined as Supabase's original bundle: the database, auth (the login system), server functions and the APIs that let other software talk to the app. AI agents now write more and more software. When an agent builds the app, the service its model suggests by default can decide what the app runs on.

A head start before the boom

The conversation offered a theory for how Supabase got there. Under that theory, a company needed to be on the market for three to four years before large language models took off. During that time it also had to present itself as "the leading competitor to the leading thing at the time." Supabase fit that window. In the conversation, that position was what let a challenger take over once the models arrived.

Supabase's answer was that it got lucky. The company could not have timed it better, came the reply, and the timing was a complete accident. Supabase had held that position from the start without planning it. Now it invests heavily in AEO, and the job is to keep the lead.

Resend did it on purpose

The discussion then compared Supabase with companies founded after the language-model boom. It mentioned interviews with some of those founders and a conversation with the team at Resend, an email service for developers, about why an LLM kept recommending them. The explanation given was that Resend follows a lot of good AEO practices.

One example of that kind of practice is Resend's llms.txt file, a plain-text index of its documentation written for agents and automation. It lists the email jobs the service handles, such as transactional messages, batches, marketing, inbound mail, templates and sending domains. It also links to setup guides, official code libraries, a command-line tool, an MCP integration and agent skills.

Other infrastructure companies measure the effect directly. In a Latent Space episode published October 22, 2026, Erik Bernhardsson, CEO of the cloud computing company Modal, described internal benchmarks that check whether agents recommend Modal. When they don't, he said, Modal responds "primarily through content." By its founders' account, Supabase had the lead first and started working to keep it later.

What the models absorbed

Wilson pointed to a second cause: the community. One of Supabase's early growth strategies, he said, was "spending so much time with the community." Developers then wrote blog posts, made YouTube videos and recorded podcasts about how to use Supabase and when, along the lines of "here's how I built this thing."

"I'm sure that all just got slurped up by the model," Wilson said. Language models learn from huge amounts of public text. Years of tutorials and build stories about one product become part of what a model knows about which tools solve which problems. The conversation also noted that being open source helps.

Every channel, one public archive

The community content had a practical problem behind it. The discussion turned to Discord, described as the best chat tool but not public, so help given there does not scale, and to whether Supabase had found a fix.

Wilson said the company looked at some tools early on but decided against them, because someone would still have had to decide what should be public and what should not. The bigger debate, he recalled, was where to host the community at all: Discord, a Slack group or Reddit. After trying a few options, Supabase concluded that "Discord users are going to use Discord. Reddit users are going to use Reddit," and that it had to be in all of those places. "It was a lot of work," he said. He named Orb and Common Room among the tools that appeared later. The discussion agreed that services copying Discord chats to public question-and-answer sites exist, though none has clearly stood the test of time.

Supabase built its own system. Whenever someone asked a legitimate question in any of those places, the company copied the question and its answer into GitHub Discussions and mirrored them onto Supabase's website. In practice, Supabase used GitHub as a content management system, the kind of software that stores and publishes a website's pages. GitHub Discussions is a forum attached to a code repository or organization. Maintainers can sort posts into categories, pin important threads and mark a reply as the accepted answer to a question.

The conversation linked that archive to the content models later absorbed. Supabase had a simpler reason for keeping it up at the time: people asked the same questions over and over, so the company decided to surface the answer each time.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02

Connected ideas and articles

From the conversation

Podcast episodes

Latent Space

Supabase’s Plan for a Billion AI-Built Apps

Episode published This article draws on 28:19–29:59 and 32:13–34:01 (approximate times)

Article history

Updates to this article

Tags