We Analyzed 37,000 Websites: How Site Speed Impacts Google Rankings

The study starts from search queries rather than from a list of popular websites, because a rankings study needs ranking positions, and a top-sites list tells you nothing about what those sites rank for.

Keywords. We assembled 9,553 keywords stratified across ten commercial verticals (ecommerce, software, finance, legal, health, real estate, travel, education, home services, and general information) and three search-volume bands.

Search results. For each keyword we pulled the top 20 organic results, producing 170,017 ranking positions across 122,650 unique URLs and 39,689 unique domains. We used a third-party SERP API on Google US, mobile, to avoid scraping Google directly.

Speed. For every ranking URL we queried the Chrome User Experience Report (CrUX), the same real-user field data Google's page-experience systems draw on. We captured the 75th-percentile values for Largest Contentful Paint, Interaction to Next Paint, Cumulative Layout Shift, Time to First Byte, and First Contentful Paint, along with a pass or fail flag for all three Core Web Vitals. Field data existed for 64,864 ranking positions, covering 37,036 unique websites. That is the analysis set behind every speed number in this report.

Controls. We attached each domain's Tranco popularity rank so we could test whether any speed advantage was really just popular sites being fast.

Coverage note: 38% of ranking URLs had CrUX field data, higher near the top of the results. Why that works in a competitor's favor rather than against it is covered in the limitations section.

The share of ranking pages that pass all three Core Web Vitals climbs as you move up the results.

Page-one results (positions 1 through 10) pass Core Web Vitals 47.6% of the time. Page-two results (11 through 20) pass 44.2%. The top three positions pass 50.8%, and position #1 on its own passes 49.2%. It is a clean, monotonic slope.

Raw load time tells the same story. The median real-user LCP improves from 2,171 milliseconds on page two to 2,099 milliseconds on page one. First Contentful Paint follows the same curve (1,708 ms on page two versus 1,628 ms on page one), as does Time to First Byte (827 versus 796 milliseconds).

The differences in the middle of the pack are small. But the direction never wavers, and it gets stronger as you approach the positions that actually drive clicks.

When we split results by how competitive the site is, using popularity as the proxy, the speed advantage is far from uniform. It is largest among the most popular, highest-traffic sites.

Among those top-tier sites (roughly the most popular bracket by Tranco rank), page-one results pass Core Web Vitals 46.7% of the time versus 39.1% on page two. That is a gap of nearly eight points, with a 95% confidence interval of 6.5 to 8.8 points on tens of thousands of observations. As you move into less competitive results, the gap narrows toward zero.

This is also the subgroup where the data is hardest to fully untangle, because the highest-traffic sites are exactly the ones most affected by the measurement quirks described later. Google has described its page-experience signal as a tiebreaker between otherwise-comparable results, and the place that tiebreaker should matter most is the place we see the widest gap.

More than half of the pages currently outranking any given site fail Core Web Vitals.

Page-one results pass just 47.6% of the time, which means 52% fail. Even at position #1, only 49.2% pass all three metrics. Passing Core Web Vitals is not a requirement to rank, and failing them is not a wall that keeps a page off page one.

The right way to read that is as an opening: most competitors have left speed untouched. This is one of the few ranking inputs where the field is wide open.

The obvious objection to any speed-and-rankings finding is that fast pages might rank well only because big, authoritative sites happen to be fast. We tested that by adjusting for each site's popularity. When we do, the speed-to-rank relationship does not shrink. It moves from a raw correlation of +0.087 to +0.116 after the adjustment.

Popularity is an imperfect stand-in for the link, relevance, and brand signals Google actually uses, and it is partly downstream of ranking itself, so we treat the result as supporting evidence rather than proof. But the direction is consistent: comparing sites of similar standing, the faster ones rank a little higher.

Every Core Web Vital correlates with rank in the same direction, but they are not equal. We calculated the correlation between position and each metric inside every keyword's own results, then averaged across roughly 3,900 keywords.

Largest Contentful Paint led at +0.090, followed by Time to First Byte at +0.080, then Cumulative Layout Shift at +0.033 and Interaction to Next Paint at +0.023. These are small correlations. An r near 0.09 means speed explains under 1% of the variation in rankings on the open web. This is a faint but consistent signal, not a dominant one.

Loading is also where pages break. Most ranking pages clear visual stability (77% pass CLS), but loading and interactivity are tighter, with 63% passing LCP and 63% passing INP. Among the pages that fail Core Web Vitals outright, 67% fail on LCP and 60% on INP, versus 42% on CLS.

Speed targets should be set against your sector, not against the whole web. The pass rate of pages that actually rank ranges from about 67% in legal down to 33% in travel.

Legal (67%), health (55%), education (49%), home services (49%), and finance (49%) sit at the faster end. Real estate (42%), ecommerce (42%), general information (41%), software (40%), and travel (33%) sit at the slower end. Script-heavy industries like travel and software are home to the slowest pages that still manage to rank, which suggests other signals carry more weight there.

The strength of the speed-to-rank link also shifts by sector. It is a little stronger in real estate (+0.134), health (+0.125), and general-interest queries (+0.115), and weakest in finance (+0.044), though all of these are small and carry the same caveats as the rest of the study.

A profile of what a typical page in Google's top 20 looks like, so there are concrete numbers to measure against.

The median ranking page loads its main content (LCP) in about 2.1 seconds, responds to interaction (INP) in 152 milliseconds, holds layout stable at a 0.02 CLS, and returns its first byte (TTFB) in 813 milliseconds, with First Contentful Paint around 1.66 seconds.

Beat those figures and a page is already ahead of the typical result it competes with. Pull comfortably under them, toward Google's "good" thresholds of 2.5-second LCP, 200-millisecond INP, and 0.1 CLS, and it is in the top tier of the field. For the full playbook that gets a page from average to top-tier, read our guide to page speed and SEO.

The ranking case is the smaller half of the argument. Google has treated Core Web Vitals as confirmed ranking signals since 2021, and continues to weight page experience as part of its core systems (Google Search Central). The larger and better-established effect is on what visitors do once they arrive.

Research from Portent found that a site loading in one second converts at 3.05%, dropping to 0.41% at five seconds, and that each additional second of load time can cut conversions by up to 7%. Think with Google reports that 53% of mobile visitors abandon a site that takes longer than three seconds to load. Deloitte's Milliseconds Make Millions study found that improving mobile load time by a tenth of a second lifted retail conversions by around eight percent.

The ranking effect is small but real. The conversion and revenue effect is large and well-documented.

This is observational research. It shows association, not proof of causation. Anyone who tells you a correlation study proves causation is selling something.

The limitation worth understanding in detail is coverage. Google's field data only exists for pages that get enough real traffic, so 38% of the ranking URLs in the sample had data, and that coverage is higher near the top of the results: 46% of position-#1 URLs have field data, falling to 35% at position 20.

For a competitive site the bias runs in the reader's favor. The pages the study leaves out are low-traffic also-rans, not the rivals a competitive site measures itself against. Blending millions of low-traffic pages into a web-wide average dilutes the effect, so the gaps in real competitive results are likely larger than the numbers reported here.

Two smaller notes: URL-level and origin-level field data were combined, and the signal is clearer in URL-level data (+0.086 LCP-to-position correlation, versus +0.031 for origin-level). The SERP is a single snapshot in time, while the field data is a trailing 28-day aggregate.

The pattern here lines up with what we see in adjacent field data. Just under half of the mobile web passes Core Web Vitals in our Core Web Vitals pass rates study, so a page that passes is already ahead of the typical result Google returns. And because Google grades page experience on mobile field data, the mobile vs desktop speed gap study is a useful reminder that the ranking assessment applies to the slower of the two devices, not the faster one.

Sample. 9,553 keywords stratified across ten verticals and three search-volume bands. Top 20 organic results per keyword, for 170,017 ranking positions across 122,650 unique URLs and 39,689 domains.

Ranking data. Third-party SERP API, Google US, mobile, collected within a tight window.

Speed data. Chrome UX Report API, 75th-percentile real-user field data for LCP, INP, CLS, TTFB, and FCP, plus a Core Web Vitals pass or fail flag. Field data existed for 64,864 ranking positions covering 37,036 unique websites.

Controls. Tranco popularity rank per domain, used to adjust the pooled speed-to-rank correlation.

Limitations. Observational data (association, not causation). Field-data coverage is 38% of URLs and weighted toward higher positions. URL-level and origin-level CrUX were combined. The SERP is a single snapshot; field data is a trailing 28-day aggregate. Effect sizes are small in absolute terms and largest within the most competitive subgroup.

How to cite this study. PageSpeed Matters, "We Analyzed 37,000 Websites: Site Speed vs Google Rankings," June 2026.

Third-party figures referenced for business-impact context are drawn from Google Search Central, Portent, Think with Google, and Deloitte Digital, each linked inline. All ranking and speed figures are from our own dataset.