Docs
Plumb, page by page.
What each page is for, where its numbers come from, how to read it and who uses it. The i beside any page title in the product links here.
01Getting started
- 01Connect your site
Settings generates a Cloudflare Worker for your content routes. Paste it in, and every request from a bot reaches Plumb. Client IPs are hashed at the edge; the raw address is never stored. Not on Cloudflare? Upload server logs instead.
- 02Connect Search Console and GA4
Two OAuth clicks. Search Console gives impressions, clicks and rank per query. GA4 gives the sessions AI assistants send back.
- 03Read the first week
Agent Traffic fills within the hour. Bot Policy needs a day of requests before verdicts are worth reading. Opportunities and Commissions need Search Console history, which arrives on the first sync.
Monitor
Agent Traffic
Every AI agent that requested a page on your site, by platform and by day, with the human readers each platform sent back.
Where it comes from
Your Cloudflare Worker reports each request to Plumb. Search engines and SEO tools are counted separately from AI agents. Measured.
How to read it
- AI-agent requests exclude Googlebot, Bingbot and SEO crawlers; those sit in their own count so the AI number stays honest.
- Retrieval means an assistant fetched the page to answer a question right now. Training means a crawler is collecting pages. Indexer means an AI search index.
- Crawls per referral compares pages a platform took with readers it sent back. A blank ratio means it sent nobody.
- Unknowns are user agents the registry does not recognise yet. Review them before you block them.
What to do with it
- Check who is reading your site before a licensing or blocking decision.
- Find the pages agents want most in the top-pages table.
- Watch the daily chart after a policy change to see what it did.
Who uses it
Audience team, infrastructure, anyone preparing for an AI-company conversation.
Monitor
Overview
One screen for the week: the biggest leak, the evidence behind it, and which page owns the next step.
Where it comes from
The overview endpoint joins edge requests, GA4 sessions and the content catalog. Mostly measured; the revenue-at-risk figure is inferred.
How to read it
- The briefing paragraph is written by Plumb's reader from the week's numbers and names the tool behind each claim.
- Consumption numbers (crawls, top platform, URLs touched) are the story. Session numbers are context.
- Every card links to the page that explains it, with the same filters.
What to do with it
- First stop on Monday.
- Send the page recap to people who will not log in.
Who uses it
Every role.
Monitor
Visibility
How often AI assistants cite you, per platform, against named competitors on the topics you care about.
Where it comes from
A weekly panel of 500 queries run against ChatGPT, Claude, Gemini and Perplexity. Sampled, and noisy.
How to read it
- The same question returns different citations hour to hour. Treat week-to-week moves under 15 points as noise.
- Use it for share of voice against competitors, never for an absolute visibility claim.
- Head-to-head cards show the eight-week win rate against one competitor.
What to do with it
- Weekly review with content strategy.
- Answer why a competitor is winning a topic cluster.
Who uses it
SEO and audience strategy, editorial leadership.
Monitor
Mission Control
A live, full-screen view of crawler hits as they land, one lane per bot.
Where it comes from
A streamed feed from the edge. Measured, with a synthetic fallback when the stream is unreachable, labelled as such.
How to read it
- A lane with a red border is a bot marked block that is still getting 200s. That is a compliance problem.
- The anomaly ticker flags rate spikes and unexpected 404s.
What to do with it
- The ops screen during an incident.
- The clearest demo of first-party measurement.
Who uses it
Infrastructure and on-call.
Diagnose
Bot Policy
What your robots.txt says, who listened, whether each agent is who it claims to be, and the rules Plumb proposes in response.
Where it comes from
Your live robots.txt, the edge log, twelve vendors' published IP ranges and RFC 9421 signatures. Measured. Rules are proposals; Plumb never applies them.
How to read it
- Ignored means the agent fetched a path your robots.txt disallows for it. Honored means it stayed out.
- Spoofed means a request wore a known agent's User-Agent from outside that vendor's published ranges. Unverifiable means the vendor publishes no ranges, so nobody can check.
- Each proposed rule carries the evidence lines that produced it. The block rule targets agents that ignored the file; the challenge rules target look-alikes.
What to do with it
- Weekly review of new and unknown crawlers.
- Copy a proposed WAF expression into Cloudflare.
- Export the policy files (robots.txt, llms.txt) or the litigation record for counsel.
Who uses it
Infrastructure, SEO lead, general counsel.
Diagnose
Inventory Yield
Which sections and stories are read heavily by AI without sending readers back.
Where it comes from
Edge requests joined to the catalog, Search Console positions and an ad-yield estimate. The ratio is measured; the dollar figure is inferred.
How to read it
- Sections below the reference line have the widest gap between what AI takes and what comes back.
- The under-yielder queue lists URLs by worst ratio.
- Quadrants: moat (cited and ranking), opportunity (ranking, not cited), dark (neither).
What to do with it
- Give editors the weakest quadrant as a fix list.
- Export for the ad-yield review.
Who uses it
Managing editor, ad-revenue analyst.
Diagnose
Funnel
Visit, engaged, newsletter, registration, subscription, per channel, so you can see whether AI-referred readers ever convert.
Where it comes from
GA4 events walked per anonymous visitor. Measured.
How to read it
- Engaged means scrolled past half the page or stayed a minute.
- A less-than-or-equal sign means the sample was too small; the figure is an upper bound.
What to do with it
- Find the rare sections where AI readers convert.
- Compare logged-in and anonymous AI readers.
Who uses it
Subscriptions, audience product.
Diagnose
Channel Health
Where sessions come from and how the mix is shifting, including the AI channel.
Where it comes from
GA4 sessions with referrer classification. Measured.
How to read it
- AI share moves in percentage points. Two points is a real change.
- AI-referred cohorts retain less than direct readers do. That is normal.
What to do with it
- Investigate any overview alert about channel mix.
- Export for the monthly finance review.
Who uses it
Audience team, growth.
Diagnose
Content Health
A 0 to 100 readiness score per article, with the fixes most likely to earn a citation.
Where it comes from
Catalog signals scored against citation research. Inferred.
How to read it
- Bands: 80 and up is strong, 60 to 79 is fine, 40 to 59 needs work, under 40 is unlikely to be cited.
- The under-cited queue is articles with heavy crawl and weak citation.
What to do with it
- Monthly portfolio review.
- Check a draft before it publishes.
Who uses it
Editorial ops, SEO lead.
Diagnose
Raw Events
Read-only SQL over the warehouse when a page does not answer the question.
Where it comes from
SELECT and WITH only, three-second cap, 50,000 rows.
How to read it
- Cmd+Enter runs the query.
- Saved queries stay with the page they were written on.
What to do with it
- Ad-hoc investigation.
- CSV download for spreadsheet work.
Who uses it
Data and analytics.
Act
Opportunities
Queries where you still draw impressions but clicks are falling, ranked by what is still recoverable.
Where it comes from
Search Console impressions, clicks and rank, with a flag for queries where an AI Overview rendered. Measured; the priority score is inferred.
How to read it
- Bleeding means impressions held while clicks fell. Someone else answered the question, usually the results page itself.
- Refresh means you still rank and a structural fix can recover it. Write new means nothing in the library ranks for it.
- Priority weighs impressions, the size of the click drop and how close you are to the top.
What to do with it
- Accept, dismiss or snooze each query. Decisions are kept.
- Accepting a query drafts a commission brief.
Who uses it
SEO lead, desk editors.
Act
Commissions
The week's editorial work order: write new, refresh or retire, each with a brief and an owner.
Where it comes from
Built from Opportunities, Emerging Demand and Content Health. Inferred, and labelled as directional.
How to read it
- Each brief says why this call over the alternative and which structural elements earn citations.
- Outcome tracking starts when the piece publishes: citations earned and AI-referred revenue attributed to it.
- Retire candidates are past the 13-week citation half-life with falling crawl volume.
What to do with it
- The artifact for the planning meeting. PDF export is one click.
- Share the URL; the filter state travels with it.
Who uses it
Managing editor, desk editors.
Act
Emerging Demand
Questions trending up in AI answers where competitors are cited and you are not.
Where it comes from
The same weekly 500-query panel as Visibility. Sampled.
How to read it
- Opportunity: demand rising, you are absent. Defense: you slipped. Moat: you lead and it is stable.
- Check a query to test an idea before committing a writer.
What to do with it
- Feeds Commissions.
- Validate editorial hunches cheaply.
Who uses it
SEO lead, desk editors.
Act
Editorial
Which stories broke out in AI answers, by byline and by desk, and how fast.
Where it comes from
Edge requests and citations windowed by story. Resonance is a z-score against the section median. Sampled and measured.
How to read it
- Time to first citation is in hours.
- Ignored stories did well with human readers but were never picked up by AI. The diagnosis label is a hypothesis.
What to do with it
- The Tuesday editorial meeting.
- Click a story for the platforms that cited it.
Who uses it
Managing editor, desk leads, reporters.
Prove
Reports & Alerts
Scheduled digests to inboxes and rule-fired alerts, with a log of what fired.
Where it comes from
Reports run on a schedule; alerts watch a metric and fire to Slack, email or a webhook.
How to read it
- Preset templates cover the common cadences, including the Monday digest.
- The firing log is the history.
What to do with it
- Subscribe the people who will not open the dashboard.
- Alert on AI referrals dropping or a competitor passing you.
Who uses it
Anyone who wants the numbers pushed to them.
Prove
Reconciliation
Your edge log set against each marketplace vendor's reported totals.
Where it comes from
First-party counts are measured. Vendor columns are synthetic until live vendor connectors ship, and say so.
How to read it
- Variance bands: within 2%, review under 8%, material at 8% or more.
- Annual dollar impact is variance times the average rate card.
What to do with it
- Monthly check of every marketplace statement.
- Evidence for a billing dispute.
Who uses it
Licensing lead, finance, counsel.
Prove
Per-Platform View
A quarterly consumption-and-return report for one AI platform, built for a licensing conversation.
Where it comes from
Edge requests and sessions filtered to one model family. Measured.
How to read it
- Top-20 crawled and top-20 session-producing pages side by side.
- High overlap between the two lists suggests retrieval, not training.
What to do with it
- The pre-meeting packet for any AI-company conversation.
- PDF export is formatted as a deal pack.
Who uses it
Licensing lead, counsel.
Prove
Pricing Intelligence
A rate-card suggestion per section and per bot, for publishers deciding what to charge AI crawlers.
Where it comes from
Crawl volume, citation density, ad-yield parity and a willingness tier per bot. Inferred.
How to read it
- Base price times section factor times ad-yield parity times bot tier.
- It is a starting point for a negotiation, not a market price.
What to do with it
- Read before any marketplace conversation.
- Export keyed for Cloudflare's pay-per-crawl rule builder.
Who uses it
Licensing lead, finance.
Prove
Scenario Explorer
An estimate, with bounds, of how much of your direct traffic started in an AI conversation.
Where it comes from
A Bayesian estimate from the conversion lift of direct over organic. Inferred.
How to read it
- The output is a range, not a number. The sweep chart shows how it moves with the prior.
What to do with it
- A defensible hidden-AI-share figure for the board.
- Evidence that AI influence extends past attributed sessions.
Who uses it
Data lead, finance.
System
Agents & MCP
Use Plumb from Claude, Cursor or your own agent. Every table and workflow in the product is also a tool call.
Where it comes from
A hosted MCP server with 22 tools, OAuth 2.1 for your tenant, and a public no-token server that reads the demo tenant.
How to read it
- Tools mirror the pages: get_agent_traffic, get_robots_compliance, get_agent_identity, propose_edge_rules, list_bleeding_queries, create_commission_draft and so on.
- Every tool result carries the same measured, sampled or inferred label the dashboard shows.
- Ask Plumb inside the dashboard calls the same tools, so an answer in Claude and an answer on the page agree.
What to do with it
- In Claude: Settings, Connectors, add a custom connector with the server URL from this page.
- Ask what to commission this week, or which crawlers ignored robots.txt.
Who uses it
Anyone who works in Claude or another MCP client.
System
Settings
Connect your site and accounts, manage members, mint MCP tokens, read the audit log.
Where it comes from
Connectors, members, tenant defaults, audit log.
How to read it
- Roles: admin, analyst, editor, viewer.
- The Worker script for your site is generated here, with your site key shown once.
What to do with it
- First session: connect the Worker, Search Console and GA4.
- Before a board review: read the last 30 days of the audit log.
Who uses it
Admins.