Disclosure: I'm Lindenbyte, the small studio that built the Apify Actor used below, and this post was
drafted with AI assistance. pagespeed.web.dev is free and fine for one URL at a time, and the PageSpeed
Insights API is free with your own key if you want to write the script yourself. The Actor adds sitemap
input, mobile + desktop in one run, flat CSV/JSON output, pass/fail thresholds and scheduling.
Checking one page in PageSpeed Insights is easy. Checking the 15 templates that actually make your money,
on mobile and desktop, every week, is where people give up. Here is how to do it from a sitemap in one
run, and why you should always read lab and field numbers side by side.
1. Point it at a sitemap
Open the Bulk PageSpeed & Core Web Vitals Checker
and give it a sitemap (a sitemap index works too). For this post I audited the first 15 URLs of
BBC Good Food's public sitemap index on mobile:
{
"sitemapUrl": "https://www.bbcgoodfood.com/sitemap.xml",
"maxUrls": 15,
"strategy": "mobile",
"minPerformanceScore": 70,
"maxLcpMs": 2500,
"requireFieldCwvPass": true
}
No Google API key needed (the Actor has one built in; you can add your own apiKey). The run took about
2 minutes for 15 audits. Each row has the Lighthouse score, lab metrics (LCP, CLS, TBT, FCP, Speed Index),
real-user Chrome UX Report (CrUX) p75 values, Google's Core Web Vitals assessment, the top fix and a
passed flag. A SUMMARY record in the run's key-value store gives the totals.
2. Read lab and field side by side
Result (run 2026-11-13): median performance score 40, median lab LCP 10.4 s, and all 15 pages failed
the thresholds. Sounds like a disaster. Now look at the real-user columns:
| Page (mobile) | Perf | Lab LCP | Field LCP p75 | Lab CLS | Field CLS p75 | Field CWV |
|---|---|---|---|---|---|---|
| /review/best-air-fryers | 40 | 10,876 ms | 1,077 ms | 0.248 | 0.27 | FAILED |
| /review/best-toasters | 39 | 10,426 ms | 876 ms | 0.333 | 0.23 | FAILED |
| /review/best-cast-iron-skillet-pans | 38 | 10,752 ms | 1,061 ms | 0.317 | 0.09 | PASSED |
| /howto/guide/how-to-cook-turkey-crown | 48 | 8,117 ms | 1,568 ms | 0.019 | 0.23 | FAILED |
Lab LCP said 7.8-13.7 seconds; real Chrome users saw a p75 LCP of 1.6 s or better on every page. Lighthouse
simulates one cold load on a throttled mid-range phone over slow 4G; CrUX is what real visitors experienced
over the last 28 days, with warm caches and faster devices. Neither is "wrong", but Google's ranking signal
uses field data.
The real problem shows up in the field columns: 12 of 15 pages failed Core Web Vitals, every one on CLS,
with CLS p75 of 0.21-0.27 on the failing pages that had their own data (good is 0.1 or less; one page also missed on INP). Lab CLS can miss this
entirely: the turkey guide had a lab CLS of 0.019 but a field CLS of 0.23, because real users scroll into
late-loading ads and embeds, and lab only measures the initial load. The cast-iron page is the reverse:
bad in the lab (0.317), fine for real users (0.09).
Two more columns matter here:
-
fieldDataScope: 8 pages had their own CrUX data (url); 7 only had site-wide data (origin, withfieldOriginFallback: true). Origin rows share one number, so don't credit or blame a page for them. -
topOpportunity: on 11 of 15 pages it was "Reduce unused JavaScript" (est. 960-2,400 ms), the first thing to hand to a developer.
For contrast I ran the first 20 URLs of Smashing Magazine's sitemap with "strategy": "both"
(40 audits): median mobile score 96, median lab LCP 1.65 s mobile / 0.33 s desktop, field CWV passing
everywhere it had data. A fast site looks fast in both views. The interesting cases are the gaps.
3. Pull the failures with a few lines of Python
from apify_client import ApifyClient
client = ApifyClient("YOUR_APIFY_TOKEN")
run = client.actor("lindenbyte/bulk-pagespeed-core-web-vitals").call(run_input={
"sitemapUrl": "https://www.example.com/sitemap.xml",
"maxUrls": 50,
"strategy": "both",
"minPerformanceScore": 70,
"requireFieldCwvPass": True,
})
for row in client.dataset(run.default_dataset_id).iterate_items():
if row["error"]:
print("ERROR ", row["url"], row["strategy"], row["error"][:80])
elif row["passed"] is False:
print("FAIL ", row["url"], row["strategy"], "; ".join(row["failedChecks"]),
"| fix:", row["topOpportunity"])
Each failing row carries readable reasons such as "Lab LCP ms 10876 > 2500" or
"Field Core Web Vitals FAILED", so the output can go straight into a ticket or a Slack message.
4. Schedule it weekly and alert on failures
- In Apify Console, save your input as a task (e.g. "weekly CWV - main templates").
- Add a schedule (e.g. Mondays 06:00) for that task.
- Add an integration or webhook on the task (Slack, email, or your own endpoint), and in it read the
SUMMARYrecord:thresholdFailed > 0means at least one page crossed your limits.
Pick thresholds you can actually meet today (a score of 70 would fail every BBC Good Food page above), then
tighten them. The alert then fires when a page gets worse, not every week. Note that it compares against your
thresholds, not against last week's numbers: for trend charts, export each run's dataset and keep them.
Limits (honest ones)
- Lighthouse lab scores move a few points between runs; don't alert on a 2-point drop.
- New or low-traffic pages have no URL-level CrUX data; you get origin data or nothing.
- Google sometimes can't load a page (
FAILED_DOCUMENT_REQUEST): 2 of the 40 Smashing audits came back that way, each only on one device. Those rows haveerrorset and are not charged; just re-run them. - Pages behind logins can't be audited.
Cost: $0.003 per successful audit plus $0.003 per run. The 15-page BBC Good Food run cost about
$0.05; the 40-audit Smashing run about $0.12. A weekly 50-URL, both-device check is roughly $0.30 a run.
Not affiliated with Google, BBC Good Food or Smashing Magazine. PageSpeed Insights, Lighthouse and the
Chrome UX Report are Google products.
Top comments (0)