Which AI crawlers actually fetch your sitemap, and which only follow links?

IEO Engine Research · Published 2026-07-11 · Measured from production server logs
They split into two classes, and it changes what they can find. In observed production logs, ClaudeBot fetches sitemap.xml on a roughly two-hour cadence and discovers new URLs from it directly. PerplexityBot has never fetched a sitemap on the domain observed — not once across multiple full crawls. It reads robots.txt and then walks links. This means a page that is in your sitemap but has no inbound links is invisible to link-driven crawlers and immediately visible to sitemap-driven ones.

The measurement

On 11 July 2026 a reference domain published a research series and updated its sitemap. ClaudeBot fetched sitemap.xml at 03:37:05, and fetched a brand-new page from that series at 04:23:06 — 46 minutes from map to page. Across the same 24-hour window, PerplexityBot crawled the domain three separate times, hitting 35+ pages in one twenty-second burst across eight parallel workers, and never fetched sitemap.xml once. Its pattern, every visit: fetch robots.txt, then walk links.

Why this matters more than it sounds

Sitemap-driven and link-driven crawlers do not fail in the same way. A sitemap-driven crawler will find an orphaned page — one with no inbound links — because the sitemap hands it the URL. A link-driven crawler cannot. It has no path. On the domain observed, a research directory was listed in the sitemap and linked from only 27 of 284 pages. ClaudeBot found the entire series within minutes. Perplexity crawled straight past it, three times, because every section it entered was a dead end.

What to do with this

Stop treating 'it's in the sitemap' as distribution. It is distribution for exactly one class of crawler. For the other class, only the link graph exists. If a page matters, it needs both: a sitemap entry and a real inbound link path from pages that get crawled. The failure is silent — the page returns 200 to anyone who asks, so nothing in your logs looks broken. It simply never gets asked.

Primary source: this answer is drawn from production access logs across four live deployments, not from vendor documentation or third-party tooling.
Full data and method: read the underlying field note →
All field notes: IEO Engine Research →