Server-side citation data from a live, instrumented three-site portfolio — including a working local-service business — measured from the only vantage point that can't be gamed: the server's own access logs.
Every number on this page was measured, not estimated. It comes from reading raw server access logs on three live websites — a Sarasota pressure-washing company, a consumer photo app, and this methodology site — and identifying, verifying, and counting each time an AI answer engine pulled a page to build a response for a real user. Nothing here is a survey, a projection, or a vendor's dashboard. It is a ledger.
The most important comparison in this report is two instruments describing the same sites on the same day.
How long does it take between an AI crawler ingesting a page and an answer engine actually using it? Measured across every crawl-to-citation round trip in the archive, the platforms behave like different species:
| Platform | Round trips observed | Median lag |
|---|---|---|
| Perplexity | 3 | 11 minutes |
| OpenAI / ChatGPT | 10 | 9.2 days |
| Google (AI answers) | 16 | 20.3 days |
| All platforms | 30 | 6.9 days |
The single most consequential finding for anyone still optimizing purely for rank. Comparing Google Search Console position data with measured citations, page by page, across the portfolio:
| Query (local service site) | Google position | Times cited by AI |
|---|---|---|
| "roof cleaning lido key" | 3.0 | 0 — never |
| "roof cleaning bayshore gardens" | 5.3 | 0 — never |
| "hoa paver sealing siesta key" | 43 | 6 |
Classifying every measured retrieval on the local-service site by page type produces an unambiguous winner:
| Page class | AI retrievals |
|---|---|
| Cost & pricing guides | 205 |
| Homepage | 202 |
| Long-form articles | 96 |
| Service & location pages | 70 |
The most sobering number in the ledger. AI crawlers are voracious — but ingestion is a warehouse, not a shelf:
On July 11, Common Crawl — the nonprofit dataset that major LLM training corpora are built from — performed a verified crawl of the local-service site: 181 pages in 104 minutes, one polite request every 35 seconds, robots.txt honored. Cross-checking against Common Crawl's public index confirmed 23 pages already in the May 2026 corpus, including two PDFs.
Two lessons worth publishing. First, this is a different door than citation: content taken by Common Crawl doesn't get quoted — it gets learned, becoming part of what future models simply know. Second, Common Crawl discovers by following links, and it missed four of the site's six most-retrieved pricing pages — the same under-linked pages the answer engines have to work hardest to find. Internal linking, dismissed in some AI-optimization circles, still decides what enters the training corpus.
During this measurement window, the instrument caught live impersonation: a Google Cloud VM announcing itself as an Anthropic answer agent, rented cloud machines claiming to be Google's and Anthropic's crawlers, and a probe session cycling through user-agent variants to test the site's defenses.
Measurements are produced by a private server-side instrument reading raw access logs across the portfolio: citation events are detected by proprietary log forensics, de-duplicated so multi-page fetches within one answer count once, filtered of all operator traffic and tooling, tagged and excluded during any operator self-testing, and identity-verified as described in Finding 07.
Detection mechanics, scoring weights, and instrumentation fingerprints are deliberately not documented here: publishing how the measurement works would let subjects game the measurement. The live totals above update in delayed batches for the same reason — a real-time counter is an oracle that can be probed. Findings are published; the instrument is not.
This page is updated monthly with new measured findings. The measurement continues. A citable, self-contained PDF edition of this report is maintained at a stable URL alongside this page.