Does website speed matter for AI search? We tested it against 10,000 sites

Every SEO checklist has a section on page speed. Core Web Vitals, load time, mobile performance, all treated as table stakes for ranking. It's a reasonable assumption that the same logic carries over to AI: a fast, technically sound site should be easier for ChatGPT and Perplexity to find and recommend too.

We tested it directly. For our AI Visibility Report, we audited 10,000 US local businesses through the Insites 360-degree platform, capturing full performance data (Core Web Vitals, server response time, Largest Contentful Paint, First Input Delay) and correlated it against real ChatGPT and Perplexity outcomes. The result surprised even us: website performance showed no meaningful relationship with AI visibility.

The numbers, flat across the board

Businesses found by both ChatGPT and Perplexity averaged a performance score of 59. Businesses not found by either averaged 62.4, slightly higher, if anything, than the businesses AI actually surfaced.

The rest of the performance metrics told the same story:

We saw the identical pattern when we looked specifically at recommendation rather than just awareness. Businesses in the "highly recommended" bucket averaged a performance score of 58.2 and a server response time of 214.8ms. Businesses that were never recommended averaged 59.1 and 245ms respectively, essentially indistinguishable, and if anything marginally better for the highly recommended group only on response time.

Core Web Vitals pass rates were low across every group we measured (1.1% to 3.5%), and didn't separate the surfaced businesses from the invisible ones in any consistent direction.

Why this makes sense once you think about how AI actually works

Traditional search engines prioritise page speed because it's a genuine proxy for user experience, a slow page loses visitors, so Google has strong incentive to reward faster ones. AI models aren't browsing the web the same way. They're pulling and processing raw content directly, often without fully rendering a page or waiting for it to load at all. A slow first paint or poor Largest Contentful Paint simply isn't something the model experiences the way a human visitor does.

That also explains a pattern we saw elsewhere in the report: AI visibility correlated far more with total word count and page depth than with "visible words per page", another hint that these systems are reading the underlying content of a site rather than what actually renders in a browser.

What this doesn't mean

This isn't a licence to ignore site performance. Speed still matters enormously for conversion, for user experience, and for traditional SEO, none of that has changed. What it does mean is recalibrating expectations: a Core Web Vitals fix is not an AI visibility fix, and agencies selling performance work under an "AI optimisation" banner are overselling what that work will actually do for ChatGPT or Perplexity visibility. It's the same pattern as llms.txt and other "AI SEO" tactics: plausible-sounding technical fixes that don't actually move the needle.

Where to actually spend the effort

If a client asks whether their site needs a performance overhaul for AI visibility specifically, the honest answer is no. Point the effort instead at what the data shows actually correlates: Google Business Profile strength, review volume, local pack presence, directory consistency, and genuine content depth. Keep the performance work on the books for the reasons it's always mattered, just don't sell it as the AI visibility lever.

Insites' AI visibility reports already separate these two scores, so a client never mistakes a performance grade for an AI visibility one.

This analysis is drawn from Insites' AI Visibility Report, based on a clean audit of 10,000 US local businesses across ChatGPT and Perplexity. Run a free AI visibility check on any business in seconds.