Declared crawlers, undeclared cadence.
Bots that obey User-Agent norms but ignore robots.txt at scale, sweeping documentation and content surfaces for fine-tune corpora.
Operators · ClaudeBot · GPTBot · Bytespider · PerplexityBot
Your docs train someone else's model. Your /pricing page feeds a competitor's
comparison agent. Your /api reference sits inside RAG indexes you've never heard of.
WhoWatches is the access point to the BotConduct Observatory — receiver-side
intelligence on who is reading you, what they are after, and whether their declared identity
matches their behavior.
If you run an AI-native product, you've seen at least three of these already. The BotConduct Observatory names them, characterizes them, and correlates them across the cohort.
Bots that obey User-Agent norms but ignore robots.txt at scale, sweeping documentation and content surfaces for fine-tune corpora.
Operators · ClaudeBot · GPTBot · Bytespider · PerplexityBotPeriodic, surface-scoped retrieval consistent with vector-store refresh cycles. Same paths each week. Different identities each week.
Operators · Anonymous · cohort-correlatedBrowser-like fingerprint, agentic request pattern: form discovery, link traversal, structured-data extraction. Not bot — not human.
Operators · Browser agents · MCP clients · operatorsHeadless, rotating IPs, narrow path scope: /pricing, /models, /api. Coincident with public announcements.
No agent. No DNS change. No proxy in front of your traffic. Receiver-side instrumentation drops in downstream of your WAF and feeds either the Observatory or your own stack.
Pick the integration that matches your shape. Observatory membership includes webhook delivery — pipe attribution events straight into your warehouse, your SIEM, or a Slack channel.
// POST https://yours.io/whowatches/event
// X-Whowatches-Signature: sha256=…
{
"event": "actor.observed",
"property": "yourcompany.com",
"actor": {
"id": "anthropic.claudebot",
"declared": "ClaudeBot/1.0",
"category": "declared_ai_crawler",
"confidence": 0.94
},
"observed": {
"window": "7d",
"requests": 47,
"sessions": 12,
"surfaces": ["/pricing", "/docs", "/api"]
},
"classification": "extraction",
"behavior_consistent": true,
"cohort_correlated": false,
"first_seen": "2026-05-10T08:14:22Z",
"last_seen": "2026-05-17T13:51:09Z"
}
Cloudflare, your CDN, your bot manager — they answer "should I block this?". The Observatory answers "who is showing up, on purpose, day after day?". We sit downstream and capture signals those systems do not collect and cannot correlate.
/pricing firstThe Observatory accepts a limited cohort each quarter so that observations remain interpretive, not industrial. Bulletins are written, not generated. Membership is reviewable; the Desk reserves the right to decline continuation at any cycle.
Receiver-side intelligence on automated activity directed at your public assets. Weekly bulletins surface new actors and behavioral changes. Monthly deep dives interpret what the cycle means, strategically — written by analysts.
Those answer "should I block this request?". The Observatory answers "who is showing up, on purpose, day after day?". You keep your WAF; we live downstream. Cloudflare tells you bots hit you; we tell you which ones, with what intent, and how their behavior compares across the cohort.
We don't enforce. We attribute. You take the attribution event and decide — block, rate-limit, serve a watermarked variant, or just log it. Most teams start by logging, run a cycle, and then act with intent instead of reflex.
That's the part conventional bot tools miss entirely. Agents fingerprint as browsers but behave nothing like users — no scroll, no idle, deterministic path traversal, form discovery without form submission. We classify these as agentic and tag them with their declared identity when present (e.g. anthropic.computer-use).
Yes. MCP clients have a distinct request shape we can characterize even without declared identity. API extraction at training-corpus volume produces a behavioral signature that is hard to spoof at scale.
Because the value of receiver-side observation depends on cohort-wide correlation — the same operator hitting your property and ours is a signal that no isolated stack can read. The cohort is what produces interpretation; the subscription is what gets you inside it. The Desk reserves the right to decline continuation at any cycle if cohort fit lapses.
The Observatory operates as a closed oracle. We disclose findings, not method. This protects the integrity of observations and the long-term value of the cohort. Approved members receive interpretation; they do not receive the playbook.
BotConduct is a receiver-side investigation outfit. WhoWatches is the public access point to its observatory. The same desk that signs the bulletins runs the cohort. botconduct.org →
Read the public Observatory bulletin or apply for receiver-side intelligence on your own property.
