We analyzed 572 real contractor websites.

Collection. Site measurements and the ranking pull were run in June 2026. The Core Web Vitals reproducibility snapshot is pinned to Google’s public CrUX BigQuery release 202605 (May 2026).

Frame. Independent US home-services contractors in four verticals (HVAC, roofing, kitchen & bath remodeling, and pest control) operating with their own website. The sampling frame was built from commercial business databases cross-checked against state license registries, across 13 metros stratified by four Census regions × three population tiers (seeded random draw, seed 20260611; Texas held under a 40% home-bias cap). No Google Maps, Places, or Business Profile data is used anywhere in the pipeline: that data is prohibited from export by the Google Maps Platform Terms.

Margin of error

At 572 sites the pooled margin of error is roughly ±4% (95% CI, worst case); per vertical at ~150 sites it is roughly ±8%. Each statistic on this page carries its own denominator and ±. Margins of error (±) are computed under a simple-random-sampling assumption. This is a stratified, non-probability frame drawn from commercial business databases, so every ± is indicative precision, not formal statistical inference. Between-vertical differences are only called out where the gap clears ~17 points, the level a 150-per-group sample can detect.

Field vs lab: never blended

Field data is real Chrome users (CrUX, 28-day p75, opted-in Chrome only). Lab data is a single simulated load on Google’s infrastructure (Lighthouse via PageSpeed Insights, emulated mid-tier mobile, throttled 4G). They are reported in separate sections and never combined in one number or chart. CrUX-covered sites skew toward higher-traffic, likely-healthier sites, so field-only figures overstate the health of the whole population. That is exactly why the two cohorts are split.

Named checks

Fetches were polite and robots-respecting: an identifying user-agent (okaybabe-research/1.0), no more than one request per host every few seconds, at most two pages per site (homepage + one service/contact page), each fetched twice (raw HTML + rendered DOM). Only aggregate statistics are published: no per-site results, no screenshots, no personal data.

Reproducibility: pinned CrUX snapshot

SELECT yyyymm, origin, rank,
       p75_lcp, p75_inp, p75_cls,
       fast_lcp, avg_lcp, slow_lcp,
       phoneDensity, desktopDensity
FROM `chrome-ux-report.materialized.metrics_summary`
WHERE yyyymm = 202605
  AND origin IN UNNEST(@origins)

Limitations

• This is a stratified, non-probability frame; margins of error are indicative, not formal inference.
• CrUX covers only sufficiently-trafficked origins, so the ~88% with no field data is a coverage fact, not a quality verdict; field pass-rates describe only the covered minority.
• Commercial business databases under-count website-less and subdomain-only firms, so true no-website prevalence is higher than any database-field figure; it is not published here as ground truth.
• Lab scores are a single simulated load, not real-user experience.
• Ranking and authority figures are third-party model estimates (DataForSEO Labs, indexed June 2026), labeled as such and never blended with the field/lab data.
• Schema and on-page figures are prevalence, not proven ranking drivers. No causal claim is made from this data.