Core Web Vitals: real for Google, measured at zero for AI citations

How many sites actually pass, in 2025

The 2025 Web Almanac, compiled by Himanshu Jariyal, Prathamesh Rasam, Humaira and Aaron T. Grogg for HTTP Archive and published 15 January 2026, measured Core Web Vitals across millions of origins using July 2025 Chrome UX Report data. Mobile passed at 62% for LCP, 77% for INP and 81% for CLS, with only 48% of origins clearing all three at once. Desktop did better on LCP (74%) and INP (97%) but slightly worse on CLS (72%), for 56% passing all three. INP is the newer, stricter metric, having replaced First Input Delay in March 2024, and it shows: LCP remains the single biggest failure point on mobile.

The more surprising finding is the shape of who passes. Pass rates follow a U, not a slope: the top 1,000 sites on the web hit 51% mobile pass rates, dropping to 37% for the top 100,000, before recovering to 48% across the wider population. Scale and resources do not translate cleanly into speed; the biggest, best-funded mid-tier sites carry the most legacy code and third-party scripts weighing them down. The one piece of real progress: the mobile-to-desktop gap on INP narrowed from 23 percentage points in 2024 to 20 in 2025, the first measurable step toward closing it.

Is it actually a ranking factor, and how much

Google has never denied that Core Web Vitals affects rankings, but its own framing of how much has shifted depending on who is asked. Addy Osmani, a Chrome engineer, laid out the current position in a retrospective published 2 October 2025: "a fast site won't outrank a more relevant site just because of speed," but "in cases of ties or close calls, having good Web Vitals could be the tiebreaker." Pages that pass all three thresholds see visitors 24% less likely to abandon the page mid-load, a stronger case than the ranking one, and Google's Page Experience update rolled out fully between May 2021 and March 2022.

Google's own John Mueller complicated the tiebreaker framing years earlier. Responding to a thread questioning whether Core Web Vitals mattered at all, he wrote, as reported by Roger Montti in Search Engine Journal on 6 August 2021: "It is a ranking factor, and it's more than a tie-breaker, but it also doesn't replace relevance." Four years apart, two people at Google described the same signal with two different weights. Neither statement is wrong; both point at the same conclusion for a site owner. It moves rankings some, it never overrides a page that better matches the query, and nobody at Google will give you a number to plan against.

What actually predicts an AI citation, and page speed isn't on the list

Every claim above is about Google's ten blue links. Whether Core Web Vitals affects who gets quoted by ChatGPT, Claude, Gemini or Google's own AI answers is a separate, newer, and far less settled question, and until 2026 almost nobody had tested it directly rather than assumed it.

Liam Dunne and Ben Moore at Discovered Labs did the test. Their research, published in August 2026, joined 2 million AI citation observations collected over a 6-month window against 10,000 distinct cited URLs, crawled and scored on more than 60 structural, alignment, recency and infrastructure attributes, across four engines: ChatGPT, Claude, Google AI and Gemini. The performance findings are stated plainly: real-user Core Web Vitals, meaning LCP, INP and CLS, "showed no significant independent effect once domain, content depth, page type and page age were controlled for." Synthetic Lighthouse performance scores showed no significant effect either, after the same controls. The authors' own reading is that performance "likely contributes indirectly through content structure rather than as a standalone citation signal," meaning a slow page might correlate with other problems that do get penalized, without speed itself being the thing measured against.

What did move the needle, by a wide margin, was how closely a page's content matched the actual prompt: a standardized coefficient of +0.37, described as roughly three times larger than the next-strongest page-level signal, title-to-prompt similarity at +0.09. That is the same conclusion our own piece on where AI engines actually quote from on a page reaches from a different dataset: citation is won on content structure and directness, in the first third of the page, not on infrastructure polish. Two independent studies pointing at the same lever is a stronger basis for a content strategy than either one alone.

The AI-crawler-timeout claim that keeps getting repeated without a source

A specific number circulates through 2026 AEO content: that AI crawlers abandon a page after a strict timeout of "1 to 5 seconds," implying page speed decides whether GPTBot or ClaudeBot see your content at all. We traced this claim as far as it goes and could not find a primary, tested source behind the specific figure. It appears repeated across marketing blog posts citing one another rather than any published measurement of crawler timeout behavior. One of the few companies making the claim in print, Discovered Labs itself, states elsewhere that Core Web Vitals function as "a constraint" on AI crawlability rather than "a growth lever," without publishing timeout data to support the constraint framing either.

The confirmed, tested reason AI crawlers miss content is different, and it has nothing to do with how fast a page loads. Vercel's analysis of roughly a billion crawler requests, and a separate 41-day controlled navigation experiment, both found that GPTBot, ClaudeBot and the other major AI crawlers simply do not render JavaScript at all, regardless of speed. We cover the full mechanics, including the exact study data, in do AI crawlers render JavaScript. A slow page that is server-rendered still gets read in full by every AI crawler that requests it. A fast page that requires a script to reveal its content returns an empty response no matter how quickly that script would have run.

What this means for where you actually spend the audit budget

Keep Core Web Vitals work for what it is proven to do: a genuine but limited Google ranking tiebreaker, and a real driver of whether human visitors stay on the page at all. It is not, on the only controlled evidence that currently exists, an AI citation lever, so do not expect a Lighthouse score improvement to move your citation rate, and do not deprioritize Core Web Vitals work either, since the Google ranking and conversion case for it stands on its own.

If AI citation is the specific goal, the evidence says to spend the next dollar on content structure and prompt alignment rather than a faster server: put the direct answer in the first third of the page, use concrete entities and definitive language, and confirm the page is actually reaching the crawlers at all before assuming a structural fix will help. That last check is exactly what turns up the difference between a technical health problem, which our own guide to a competitor outranking you treats as the first thing to rule out, and a content problem that no amount of server tuning will fix. A Technical SEO Audit checks Core Web Vitals, rendering and crawlability together, so the report tells you which of the two you actually have before you spend on the wrong one.

What we could not verify

Osmani's retrospective cites a 2023 pass rate of roughly 68% desktop and 64% mobile, well above the Web Almanac's 2025 figures of 56% desktop and 48% mobile. The two could reflect a genuinely tougher standard, since INP replaced the more forgiving First Input Delay metric in March 2024 and is not directly comparable to it, but we could not independently confirm the exact metric mix Osmani's 2023 figure was built from, so treat the two numbers as measuring related but not identical things rather than a real decline. Discovered Labs' citation research is a single vendor's proprietary panel, and like most AEO research shops, the company also sells citation-tracking tools, a reason to want independent replication rather than a reason to discount a specific, controlled result. No AI company publishes how page load time factors into what gets retrieved before the citation step Discovered Labs measured, so "no effect on citation" describes the step tested, not a guarantee about earlier stages of the pipeline. And the "1 to 5 second AI crawler timeout" claim circulating in 2026 content could not be traced to any primary, tested source; it should be treated as unverified until someone publishes the underlying measurement.

Sources

Core Web Vitals pass rates and the U-shaped popularity pattern come from Himanshu Jariyal, Prathamesh Rasam, Humaira and Aaron T. Grogg, Performance, 2025 Web Almanac, HTTP Archive, published 15 January 2026 from July 2025 CrUX data. The ranking-factor framing and the 24% abandonment figure come from Addy Osmani, The History of Core Web Vitals, 2 October 2025. John Mueller's quote is via Roger Montti, Core Web Vitals Ranking Factor is More Than a Tie Breaker, Search Engine Journal, 6 August 2021. The AI citation findings come from Liam Dunne and Ben Moore, What actually drives AI citations, Discovered Labs Research, published August 2026, a study of 2 million citation observations across 10,000 URLs and four AI engines.

Common questions