- Oct 30, 2026
- 3 min read
A Monthly Website Check in One Command: Web Analytics + Search Console
I’d already written about each API on its own: Search Console with a service account and Cloudflare Web Analytics via GraphQL. Knowing how to query them isn’t the same as looking. So both now sit behind one command I run once a month, during the same Sunday hour as the outbound review:
make analytics # both
make analytics-cf WEEKS=12
make analytics-gsc DAYS=28
What it pulls
| Source | Call | Credential |
|---|---|---|
| Cloudflare Web Analytics | GraphQL rumPageloadEventsAdaptiveGroups, one query per 28 days (longer ranges are refused), by date, host, path, referrer, country, device, browser | The wrangler login (the make target runs wrangler whoami first to refresh it) |
| Search Console | searchAnalytics/query for totals, queries, pages and dates; sitemaps for when Google last read each one | Service account JSON, read-only scope |
Raw responses are saved as JSON next to the script (gitignored), so a closer look never needs a second API call.
The report
- Visits per site per week, 12 weeks, every host in the account.
- Per site, last 28 days: top referrers, countries, pages, devices, browsers.
- Search, last 28 days vs the 28 before: clicks, impressions, average position.
- Top queries and pages by impressions.
- “Page 1, no clicks”: pages at position ≤ 10 with impressions and zero clicks.
- Impressions per week and when each sitemap was last read.
Bots or people?
My busiest site shows about 3,000 visits a week. The breakdown:
| Signal | Share |
|---|---|
| No referrer | 96% |
| Desktop | 86% |
| Browser “Unknown” | 64% |
| Countries | US, Singapore, China |
People arrive from somewhere (a search engine, a link) and mostly on phones. This is crawlers running JavaScript. The real referrals in four weeks: Bing ~250, DuckDuckGo ~50, Reddit ~30. Read the shares, not the total.
Sampled numbers move between runs
Web Analytics data is sampled, and for a small site the sample changes from query to query: the same week came back as 200 visits on one run and 30 on the next. For small sites, trust the trend over several weeks and the proportions, never one week’s count.
Two Search Console gotchas
- Rows are ordered by clicks. With zero clicks everywhere, the “top 10” is alphabetical. Sort by impressions yourself.
- Most queries are anonymised. A page with 50 impressions may show two visible queries. Write for the page’s topic, not for the few queries you can see.
What one month showed
- Submitting sitemaps worked. One site went from ~8 impressions a week to 481 two weeks after its sitemaps were submitted. Google had simply never been told about most pages.
- Fewer impressions isn’t always worse. Another site lost most of its impressions while its average position improved from 53 to 31 and clicks went up. What vanished were impressions on results pages 6–8, which nobody scrolls to anyway.
- Organic search brings about one click a week across all my sites. It’s compounding slowly. New business, for now, comes from outbound.
The list that turned into work was number 5, and that’s the next post.
Takeaways
- Put the APIs behind one command, then actually run it.
- Judge traffic by its shares (referrer, device, browser) before its size.
- Sampled data: trends, not single weeks.
- “Page 1, no clicks” is the cheapest SEO work there is.