Plumb records every AI agent at your edge, proves which ones are real, joins the crawl to Search Console and GA4, then uses AI on that record to draft the rules and the briefs your desk would otherwise write by hand.
01A · Agent traffic
Fetching the last 30 days…
| Agent | Platform | Type | Requests |
|---|
Which crawlers ignored our robots.txt this month?
get_robots_compliance · measured
Five training crawlers fetched pages your robots.txt disallows: ClaudeBot (3), Bytespider (2), CCBot, Amazonbot and GPTBot (1 each). Want edge rules drafted for them?
Cloudflare Radar · 28 days to 21 Jul 2026
HUMAN · State of Agentic Traffic, Jul 2026
A Cloudflare Worker logs every agent request on your content routes. IPs are hashed at the edge.
Which AI agents are reading us?Each agent is checked against its vendor’s published IP ranges, its signature and your robots.txt.
Is that really Googlebot?Crawls sit next to Search Console, GA4 and AI referrals, so what AI takes and what it returns share a table.
Which queries are AI Overviews eating?Plumb drafts the edge rules and the commission briefs, then re-measures after you ship them.
What should we block?01Questions Plumb answers
Pick one. The bot answers come from the live edge log on plumbtest.xyz; the search answers come from a sample finance publisher. Every chart is the same data the product shows.
13 AI agents made 103 requests in 30 days, on top of 104 from search engines and SEO tools. Amzn-SearchBot alone made 49%.
Hover a bar for platform and type
Top 10 of 13 agents; the rest made 3 requests between them.
02 AI inside
The same models that read your pages read your data. Plumb turns the joined edge, search and citation record into recommendations, and labels every inferred number so an editor knows what is fact and what is a model.
Five training agents ignored robots.txt on 12 requests; one Googlebot copy was spoofed; block-robots-ignorers is the top rule (12 requests).
get_robots_complianceget_agent_identity“how to get approved for Capital One Venture X”
Write a new piece targeting "how to get approved for Capital One Venture X." This query drew 7,288 impressions in 28 days but only 95 clicks, a 29.6% click collapse against AI Overview cannibalization. It is the largest single lost-click source in inventory.
Why write new rather than refresh: nothing in the current library ranks for this. The nearest existing piece (/personal-loans/2026/06/29/story-0055) is off-topic and off-vertical. There is no rescuable position to fix.
Block the agents that ignored robots.txt
(http.user_agent contains "ClaudeBot") or (http.user_agent contains "Bytespider") or (http.user_agent contains "CCBot") or (http.user_agent contains "Amazonbot") or (http.user_agent contains "GPTBot")Evidence
A cached three-line recap on each dashboard page, regenerated daily, every line naming the tool it called.
Accept a query and Plumb writes the brief: why refresh over write-new, the intent, the structure that earns citations.
Block, challenge or append, each with the evidence lines behind it. Plumb never applies them; you do.
Ask Plumb rides along in the dashboard and in Claude through MCP, with the same 22 tools.
03The product
Real screens from the public demo. The same pages run on your own logs and accounts once you connect them.
Every AI agent that hit your site, by platform and day, with the readers each one sent back.
Live data · plumbtest.xyzOpen it in the live demoYour biggest leak is “LightStream vs SoFi personal loan.” It drew 31,439 impressions over 28 days, but clicks fell 25.6% while you still rank #11, which is why it’s marked refresh rather than write-new.
I drafted the brief and left it in your commission queue as pending. Every figure above is tagged inferred.
04Ask
Add Plumb as a custom connector and ask in plain English. The public demo server needs no token.
https://plumb-api-767347077340.us-central1.run.app/mcp/public/05Trust
An editor, a lawyer and a CFO will each ask where a number came from. Plumb is built so the answer is one click away.
Every number says whether it was measured, sampled or inferred. Pages that run on sample data carry a banner that says so.
Each classification traces back to the raw request, the rule that fired and the feed version that decided it.
Signature-Agent: chatgpt.comRFC 9421 signatures. Web Bot Auth checked against each vendor’s key directory.
203.0.113.24 → 9f3c…e1Hashed at the edge. The Worker hashes client IPs. The raw address is never stored.
allow · challenge · blockProposal-only rules. Plumb drafts allow, challenge and block rules. You apply them.
plumb_oat_••••OAuth 2.1 for MCP. Dynamic client registration, PKCE and refresh tokens.
raw 14d · rollups ∞14-day raw retention. Raw events roll off after two weeks. Daily rollups stay.
baseline lockedBefore and after any change. Lock a per-agent baseline before a policy change, then measure what it did.
Plumb · Monday digestMonday digest. The recaps that matter, emailed to the desk each week.
brief-04213.pdfExportable evidence. Briefs and reports go to PDF for people who never log in.
06Changelog
The MCP server returns structured data, so the agent-traffic view renders in hosts that support MCP Apps.
Requests checked against 12 published vendor IP-range feeds. Spoofed User-Agents get flagged.
Your live robots.txt set against what each agent actually fetched, with a verdict per agent.
Streamable HTTP, dynamic client registration and a no-token public demo server.
07FAQ
What you need to connect, who Plumb is for, and how it treats the numbers.
A Cloudflare Worker on your content routes for the edge log, plus Search Console and GA4 for the search and referral side. Not on Cloudflare? Server-log upload works too.
Publishers whose articles earn money one at a time: affiliate and commerce sites, B2B trade media, paid newsletters. It is not a brand-sentiment tracker and it does not write articles.
No. Plumb measures what reaches your edge whether or not you have a deal, which is also what you want in hand before you negotiate one.
Measured comes from systems you own. Sampled comes from probes Plumb runs. Inferred comes from modeled relationships, and is labelled that way everywhere it appears.
No. Plumb proposes allow, challenge or block rules per agent with the evidence behind each one. You decide, and you apply them in Cloudflare.
No. It gives editors a better instrument for the calls they already make.
Connect one site and see which agents read it, which ones are real, and what came back, in the first week.