If a site's main content only appears after JavaScript runs in the browser, AI search crawlers have historically never seen it.

That's not a guess. Vercel and MERJ tested it directly in December 2024 by monitoring more than 500 million real crawler fetches, and found that GPTBot, ClaudeBot, PerplexityBot and several other major AI bots fetch a page's raw HTML and leave, without ever executing the JavaScript that builds the rest of it.

Only Google, through Gemini's use of Googlebot's existing rendering infrastructure, and Apple's browser-based crawler were doing anything different. That finding held up when checked again against each vendor's own documentation as recently as June 2026.

The rule that's been safe to repeat for two years

The scale of the gap is why it has mattered so much to anyone building a site in React, Vue or a similar framework. One widely cited estimate from February 2026 put the cost at 42% of JavaScript-rendered content never getting indexed by AI systems at all. A 2025 survey of 295 SEO professionals found only a third knew the rule; the rest either guessed wrong or had no idea.

For two years, "serve your main content as plain HTML, because the AI crawlers won't run your JavaScript" has been close to the single most reliable technical fact in AI search optimisation.

What changed on one developer's sites in late September

A developer running several Next.js sites, one of them a Japanese freelance job marketplace, published his server logs this week showing that pattern breaking. On 25 September 2026, two of OpenAI's crawlers, OAI-SearchBot and GPTBot, started firing a specific type of request that Next.js only sends when a page has genuinely been rendered client-side: a prefetch request carrying a ?_rsc= tag in the URL, invisible in the raw HTML and only generated once JavaScript executes.

Before that date, GPTBot was visiting from a single IP address a day. From 25 September, it was visiting from 36 to 131 IP addresses a day, and daily GPTBot requests to his main site rose from 83 to over 150,000 within a few days.

Raw request counts alone can mislead, since a single rendered page fires off many prefetches at once. He corrected for that by counting distinct referring pages, not total requests, and still found a clear jump: OAI-SearchBot's count of genuinely rendered pages rose roughly tenfold over a week, and GPTBot's rose more than a hundredfold.

He checked every request against OpenAI's own published IP ranges to confirm the traffic was genuinely OpenAI's and not a spoofed user agent, and saw the same pattern begin on the same two days across four separate Next.js sites on his server, none of which he had changed.

Why this isn't being reported as settled fact

Three things keep this a signal, not a confirmed change, and they're worth stating plainly, not smoothed over. First, this is one person's sites. Second, he says plainly that he could find no announcement from OpenAI describing any change to its crawlers' behaviour. Third, and most telling, Cloudflare's aggregate data on AI bot traffic share showed OpenAI's overall volume holding flat through September.

That points toward a rollout limited to some sites, not a switch flipped for every GPTBot and OAI-SearchBot request everywhere.

There's also a wrinkle in how it appears to work: most of the pages GPTBot rendered on his sites hadn't been freshly fetched from his server that same day. That suggests OpenAI may be re-rendering HTML it already stored, in a separate process, well after the original crawl, not live at the moment of each crawl.

What this does and doesn't mean for a site right now

Nothing here is a reason to change a site's rendering approach yet, and several reasons not to. Anthropic's ClaudeBot, Perplexity's PerplexityBot, and every other major AI crawler besides Google's and Apple's are still confirmed, as of today, not executing JavaScript, by their own current documentation.

If a site's main content, headline, and structured data only exist in the browser-rendered version of the page, that content is still invisible to most of the AI crawler traffic that matters, this single early signal aside. And even where OpenAI's rendering does appear to be switching on, the evidence suggests it's working from previously stored HTML, not rendering fresh at crawl time. A page that was never crawlable before may not benefit from it yet, regardless.

What's worth doing is watching, not rebuilding. Making sure a site isn't accidentally blocking the AI bots it wants to be cited by in the first place is still the more urgent, better-understood technical question for most businesses. Whether OpenAI's rendering change turns out to be a genuine platform-wide shift, a narrow test, or something else entirely should become clear within the next few weeks, and it's the kind of thing worth checking again once it does.

A few common follow-up questions

Does this mean I can stop worrying about server-side rendering for AI search?

No. Even if OpenAI's crawlers are starting to render JavaScript on some sites, every other major AI-native crawler, Anthropic's and Perplexity's included, is still confirmed not to. Server-rendering critical content remains the safer default.

Which AI crawlers currently render JavaScript?

By vendor documentation through 2026, only Google (via Gemini's use of Googlebot's infrastructure) and Apple's crawler reliably render full pages. OpenAI's GPTBot and OAI-SearchBot may now be rendering on at least some sites, based on one developer's server logs from late September 2026, but this isn't yet confirmed by OpenAI itself or seen at scale.

How would I check whether this is happening on my own site?

Look at server logs for requests from GPTBot or OAI-SearchBot that include query strings JavaScript execution would generate (on a Next.js site, a ?_rsc= parameter), or fetches of image and CSS assets from a bot that previously only ever requested HTML. A plain-HTML fetch with no such pattern suggests no rendering is happening yet.