Nobody has shown that a slow site keeps you out of AI answers
No AI company publishes a speed rule for getting named in an answer, and we could not find a study that tested one. Google’s page on AI features, updated December 2025, says: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” Google’s one stated bar: your page is indexed and allowed to show in Search with a snippet.
We read the bot documentation from OpenAI, Anthropic and Perplexity too. None names a load time, a Core Web Vitals score, or a number of seconds before it gives up.
Silence is not proof that speed does nothing. Nobody can prove a negative, so we will not claim one. The rest of this splits what has been measured from what is only claimed.
Speed is proven to keep people from leaving
The settled half is about people, not machines. Vodafone ran a split test, published by Google on web.dev in March 2021. Half the traffic got a page where the largest thing on screen loaded 31% sooner, 5.7 seconds instead of 8.3. The other half was left alone, at roughly 100,000 clicks a day each. After the change, the faster half sold 8% more, and 15% more of its visits turned into leads.
What Vodafone changed: they stopped making the browser build the page and sent finished HTML.
Deloitte’s “Milliseconds Make Millions”, March 2020, found a tenth of a second faster on mobile was associated with 8.4% more retail visitors buying. That study watched brands rather than testing them, and Google paid for it. Associated with, not caused by.
Google’s web.dev gives two targets: largest thing on screen in 2.5 seconds, first byte back in 0.8.
The AI bots do not run your scripts
Vercel measured its own site, nextjs.org, and reported in December 2024 that none of the major AI bots run JavaScript. Its words: “They can’t read client-side rendered content.” That is one company measuring one site at one point in time. Google’s Gemini is the exception, since it rides along with Googlebot, which does build the page.
The operators say so themselves. Common Crawl, whose archive feeds a lot of AI systems, says in its FAQ: “Currently, JavaScript is not executed and Cookies are not used.” Google’s developer guidance, updated March 2026, says to send finished HTML, since “not all bots can run JavaScript”.
The Web Almanac’s 2025 report, on July 2025 data, found the middle mobile home page ships 2,559 KB. Only 22 KB of that is HTML. Under one percent of what you send is readable without running anything.
None of that is a speed claim. A page that shows nothing until scripts finish gives a reader nothing and a machine nothing to quote.
A broken page hurts more than a slow one
In that same December 2024 measurement, about 35% of the pages ChatGPT’s bot asked Vercel for came back missing. Claude’s bot was around 34%. Googlebot was 8%.
What Google says makes it back off is errors. From its docs updated December 2025: “Google’s crawling infrastructure reduces your site’s crawling rate when it encounters a significant number of URLs with 500, 503, or 429 HTTP response status codes.” Those codes mean your server broke or told Google to back off. Slow but working is not on that list.
Google also says it “waits for the content for a limited time, then passes on whatever it received”. It has never published how long. Anyone quoting you a number of seconds is guessing.
What we cannot promise you about speed
We cannot promise that a faster site gets you named in an AI answer. We have not seen it tested, and no AI company says it is so.
The one peer reviewed study we found on what moves AI citations is “GEO: Generative Engine Optimization”, accepted to the KDD 2024 conference. It ran 10,000 queries. Quotations lifted how often a page got cited by about 41%, statistics about 33%, naming your sources about 28%. Stuffing keywords went backwards. Page speed was never tested. What is known to move AI citations is the shape of the writing.
The search agency Onely says to get your first byte under 200 milliseconds, and writes that “slow pages get deprioritized by crawlers with limited budgets”. That is an agency’s claim, not a rule published by anyone who runs these systems.
Apply the same standard to anyone who sells you a score. Google says of its own test that “having good lab data does not necessarily mean real-user experiences will also be good”. It is a simulated load on a simulated phone. Google draws the ceiling too: “Google Search always seeks to show the most relevant content, even if the page experience is sub-par.”
What to check on your own site this week
Turn scripts off in your browser and reload your own site. In Chrome it is under settings, site settings, JavaScript. If your prices, hours and service list disappear, that is close to what most AI bots get.
Then click your main pages and watch for anything that errors or lands on a “not found” page. Broken is the documented problem, so clear that first.
Last, your prices and hours should be text on the page, not a picture of a price list. A machine cannot read a picture.
We build every site as plain HTML files, so the words are in the file before any script runs. No database, no plugins. That alone does not get you named.
If you want the AI side checked for you, our free scan does it. We ask the questions your customers ask, write down who gets named, and reply by email after a person reviews it. It is a check and not a fix. Ask if you want a call.