If you run one website, checking Google Search Console, analytics, and your payment dashboard takes two minutes.
If you run 12 small sites solo (a mix of AI tools, craft pattern makers, calculators, and micro-SaaS), two minutes turns into a ritual you quietly skip—and skipping it is how silent failures eat a week of traffic before you notice.
Over the past few months, three silent bugs bit my own portfolio:
- One site's Google impressions dropped by ~90% after a metadata tweak, and I didn't notice for 7 days.
- Two sites had their main landing pages stuck in "Crawled – currently not indexed" for weeks while I assumed they were live in search.
- A third-party API started returning 403s and broke a core user flow without throwing a 500 on the homepage.
So I built AllSites—a single-screen portfolio console running on Cloudflare Workers + D1 that pulls uptime, first-party visitors, Google Search Console, Bing Webmaster Tools, and read-only revenue (Stripe, Dodo Payments, PayPal, Lemon Squeezy) into one "Needs attention today" queue.
Along the way, building a real multi-site monitoring product surfaced four sneaky data-engineering bugs that almost every DIY dashboard gets wrong. Sharing them here in case you're building your own telemetry stack.
1. Cloudflare Web Analytics RUM Uses Adaptive Sampling (Why Small Sites Show Multiples of 10)
When I first wired Cloudflare's GraphQL Analytics API (rumPageloadEventsAdaptiveGroups) into the dashboard, my smaller sites looked bizarre:
- Day A:
100visits - Day B:
0visits - Day C:
100visits
Every single number was an exact multiple of 10 (10, 20, 50, 100, 120), even when server logs showed 12–14 real people per day.
Why: Cloudflare Web Analytics RUM uses adaptive sampling on rumPageloadEventsAdaptiveGroups. When the sampler picks sampleInterval = 10 (or 100), a single sampled visit gets multiplied by 10 or 100—and if no visit happens to be sampled on a quiet day, that day reads 0. On top of that, ChatGPT strips referrers, so Cloudflare buckets AI search traffic under direct.
The fix: Keep Cloudflare RUM as an optional historical fallback (clearly labeled as sampled), and ship a <2 KB unsampled first-party script (t.js) that filters obvious bots, respects privacy (no raw IP storage, no cross-site tracking), excludes your own browser tests, and preserves utm_source / ChatGPT referrals. Once a site has real first-party visits, prefer those exact counts immediately.
2. Never Permanently Invalidate an Integration on a Single Timeout
A classic cron trap: your scheduled Worker loops through 11 sites sequentially, hitting an external GraphQL or Webmaster API with a 20-second AbortSignal.timeout. One site hangs for 21 seconds, throws AbortError, and your catch block dutifully runs:
UPDATE cloudflare_connections SET status = 'invalid', invalid_reason = ? Congratulations—a single transient network hiccup just permanently killed hourly syncing for the entire workspace until the user notices and reconnects.
The fix:
- Isolate every site in its own
try / catchinside the connection loop so one slow property never blocks the other ten. - Separate auth failures from transient failures. Only mark a connection
invalidon genuine credential rejection (HTTP 401/403, revoked OAuthinvalid_grant, or explicit invalid-token codes). On timeouts,5xx, or rate limits (ThrottleIP), keepstatus = 'active'and back off the next retry by 10–60 minutes.
3. Never Compare a Half-Finished "Today" Against a Full Week
If you alert on "visitors dropped >40% week-over-week" using rolling timestamps that include today at 8:00 AM, your dashboard will cry wolf every morning.
Even trickier: if you switch a site to a newly installed tracker on Day 14, the previous 7-day window only has 6 days of data—creating a fake +16.7% spike (or a fake drop if the first day was partial).
The fix:
- Always evaluate weekly deltas and drop alerts strictly on complete calendar days ending yesterday (
yesterday - 6throughyesterdayvs.yesterday - 13throughyesterday - 7) in the workspace's own timezone. - Suppress percentage-drop alerts on tiny baselines (e.g.,
< 50visits/week or< 100GSC impressions/week), where 3 fewer people looks like a 50% crash.
4. Indexation Checks Shouldn't Require Logging Into 12 GSC Tabs
Before you even look at Search Console, most indexation failures are caused by basic technical mismatches: a 301 chain, a stray x-robots-tag: noindex header, a canonical pointing somewhere else, a robots.txt rule, or a page missing from sitemap.xml.
I split that diagnostic engine out into a standalone, no-signup free tool: AllSites Google Index Checker. Paste any public URL and it checks:
- Final HTTP status & redirect chain
-
robots.txtGooglebot-style matching - HTML meta + HTTP
x-robots-tagdirectives - Canonical alignment
- Sitemap presence
- Visible text & JavaScript rendering risk
- Simple homepage discovery link check
Try It or Poke Holes in the Stack
If you're juggling 3–40 small sites and want uptime, title-change monitoring, GSC/Bing, exact human visitors, and multi-provider revenue in one place, AllSites is live with a 7-day free trial (no card required)—or just bookmark the free Google Index Checker.
Happy to share more about running scheduled multi-tenant sync jobs on Cloudflare Workers + D1, handling GSC/Bing API quirks, or reconciling Dodo/Stripe/PayPal webhooks. What's the worst silent failure you've caught too late on your own sites?