Search Engine Answer research artifact

Cloudflare Content Format: 25-Request Observation Sheet

Five request variants were sent to five Cloudflare-owned pages. This sheet separates the requested Accept value from the served Content-Type and preserves the sample boundary.

Run: 2026-09-01 00:28:29–00:28:38 UTCSample: 25 GET requestsPages: 5Client: declared SEA lab user agent

Download raw CSVCloudflare field definitions

Result summary

25/25responses returned HTTP 200
5/5text/markdown requests returned Markdown
20/20other request variants returned HTML
92.67%weighted body-byte reduction in the five HTML/Markdown pairs

Boundary: the test client was Search Engine Answer, not an AI crawler. The result verifies negotiation behavior on this sample; it does not measure crawler preference, indexing, training, citation, referral, or conversion.

Request-to-response matrix

Five Cloudflare pages tested for each request variant
Request variantRequestsHTTP 200HTMLMarkdown
No Accept header5550
*/*5550
text/html5550
text/markdown5505
application/json5550
Total2525205

Observed HTML and Markdown body sizes

Body bytes from explicit text/html and text/markdown requests
PageHTML bytesMarkdown bytesReductionMarkdown / original tokens
Markdown for Agents reference186,38517,19090.78%4,289 / 46,485
Analyze AI traffic131,85815,18488.48%3,791 / 32,863
AI Crawl Control changelog151,90423,84984.30%5,951 / 37,870
Docs for Agents121,14210,76391.12%2,685 / 30,185
Markdown for Agents blog post534,78515,56197.09%3,856 / 133,525
Five-page total1,126,07482,54792.67%20,572 / 280,928

Reductions are Search Engine Answer calculations. Token values are the x-markdown-tokens and x-original-tokens response headers returned with the Markdown representations.

Raw CSV fields

Method and repeatability

  1. Send one GET per page and request variant with a declared, non-impersonating client user agent.
  2. Preserve headers and bodies privately; record response metadata and body hashes in the public ledger.
  3. Compare request Accept and response Content-Type before calculating size differences.
  4. Keep real crawler demand, representation parity, citations, referrals, and conversions as separate studies.

The capture script and private raw responses are retained in the Batch 24 editorial package. Re-run the test before treating the numbers as current because page content and edge behavior can change.