Original research · Wave 0 · August 2026

Physician, Heal Thyself

Name: Physician, Heal Thyself — engine properties vs their own rules (Aug 2026)
Creator: SynapseIN
License: https://creativecommons.org/licenses/by/4.0/

The engines publish the rules the rest of the web must live by. We ran the same binary checks we run on client sites against eight of their own front doors — robots.txt, llms.txt, no-JS text, structured data, meta completeness. What the machines found at the measurers’ doors.

By Manas Panda & the SynapseIN teamPublished: August 30, 2026 · Cross-checked with second methods before publication

Key findings: Bing’s own webmaster page is invisible without JavaScript (12 words). OpenAI and Perplexity 403 plain readers while their robots.txt sits open — doors not closed, fingerprint-calibrated. Zero of eight engine properties serve llms.txt. One engine fully practices what it preaches: Google.

Five findings

P1 — The visibility teacher is invisible without JavaScript. Bing Webmaster Tools’ own pages serve exactly 12 words of no-JS content: “Bing Webmaster Tools — You need to enable JavaScript to run this app.” Confirmed on two URLs. The page that teaches webmasters how to be visible fails the textual-content rule Google publishes for everyone.

P2 — The companies that read the world calibrate their doors by reader type. openai.com and perplexity.ai returned 403 — homepage and /llms.txt — to plain non-JavaScript readers (browser-costumed curl, default curl, and a headless browser engine), while their robots.txt sat open. Yet an independent managed-rendering crawler on a different network was served Perplexity’s page in full. Not closed — fingerprint-calibrated.

P3 — Nobody serves llms.txt. Not one of eight engine properties: six 404s, two 403-guarded (absence unconfirmable from outside). Google’s non-use is published policy — consistent. We report the observation without inferring intent for the rest.

P4 — One engine fully practices what it preaches. Google’s AI-features guidance page: 2,096 no-JS words, valid JSON-LD, title and meta present. OpenAI’s own crawler-rules page: readable but zero structured data, no meta description. Anthropic’s memorable /claude-bot docs URL 404s (verified twice, two days apart). The bar the engines set for everyone is met, at their own doors, by exactly one of them.

P5 — Two of our own claims died in verification. We kept the bodies. An initial pass flagged DuckDuckGo for serving HTML at /robots.txt and for invalid JSON-LD. Both were our measurement artifacts; both retracted after re-fetch and a fixed extraction. A census that can kill its own findings is the only kind worth publishing.

The eight doors, property by property

Plain non-JS fetch, browser user-agent, redirects followed. *DuckDuckGo JSON-LD: one valid block after the P5 extraction fix (raw keeps as-collected values + correction note).

Engine	Property	robots	llms.txt	Page	No-JS words	JSON-LD	Notable
Google	developers.google.com/search/docs/appearance	200	❌ 404	200	2,096	1 valid	The only full pass: textual, structured, described — the standard it asks of everyone.
OpenAI	platform.openai.com/docs/bots	200	❌ 404	200	2,052	0 valid	Readable crawler rules (2,052 words) but zero structured data, no meta description.
OpenAI-www	openai.com/	200	🚫 403	403	0	0 valid	Door calibrated: robots.txt open, but homepage and /llms.txt 403 plain readers.
Anthropic	www.anthropic.com/	200	❌ 404	200	697	0 valid	Readable, schema-less; the memorable /claude-bot docs URL 404s (verified twice).
Perplexity	www.perplexity.ai/	200	🚫 403	403	0	0 valid	Door calibrated: 403 to plain clients; a managed rendering crawler on another network was served in full.
DuckDuckGo	duckduckgo.com/duckduckbot	200	⚠ HTML	200	2,320	1 valid*	Rich textual page (2,320 words) with valid WebSite schema*.
Bing	www.bing.com/webmasters/about	200	❌ 404	200	12	0 valid	The webmaster page's entire no-JS content is 12 words: “You need to enable JavaScript to run this app.”
CommonCrawl	commoncrawl.org/	200	❌ 404	200	459	0 valid	Plain, readable, honest — like the crawl itself.

Cross-check log — published in full

Every headline claim re-attacked with a different method than produced it, before publication.

Claim	Second method	Verdict
Bing 12-word page	Second URL + re-fetch	confirmed, verbatim, both URLs
OpenAI / Perplexity 403s	Default-UA curl + headless Chromium	confirmed across reader types
“Doors are closed”	Independent managed crawler, different network	corrected → calibrated by reader type
No llms.txt anywhere	Re-fetch all eight	confirmed
Google-only full pass	Word recount + JSON-LD re-parse	confirmed
Wave-1 spine (UBS 403, Mayo 403, Cleveland llms.txt)	Independent re-fetch	all hold
Two DuckDuckGo claims	Re-fetch + fixed extraction	killed — retracted (P5)

Named in this study? If you represent a property above and believe a finding is outdated or wrong, write to hello@synapsein.com — we re-run the same checks within 48 hours and publish the dated delta. Corrections are logged and dated, never silently edited.

What it means

The machine-readable web is a standard enforced downward, not upward. That is clarifying, not damning — it tells every business owner to measure their own door instead of assuming the doors of the engines have been held to the same standard.

Methodology & limitations

2026-08-30, 00:30–01:15 IST, single vantage point (India), plain fetches with a browser user-agent, 12s timeouts, 0.5–0.7s politeness. JSON-LD extracted only from served <script type="application/ld+json"> blocks (payload-embedded strings excluded after P5). Limitations: n=8, one evening; a 403 to our reader is a fact about our reader, not a universal claim; SPA-by-design is an engineering choice, not a violation — we measure against Google’s published guidance as the industry’s explicit standard. Re-run monthly alongside the wealth study; deltas will show who fixes their door.

Data & license

Raw data: physician-raw.json — CC BY 4.0: reuse freely, attribute SynapseIN, link the study. Corrections are logged and dated, never silently edited.