The method: Agent Entry Chain (AEC)
A source is only used by an AI agent if it survives a chain of stages. The Agent Entry Chain gives that chain a name so it can be cited, measured and audited.
Definition
The probability that an AI agent discovers, opens, understands, cites and acts on a source is the product of the per-stage probabilities. Because it is multiplicative, a single near-zero stage collapses the result regardless of the others.
P(use by agent) =
P(discovery)
× P(crawlability)
× P(indexing)
× P(query matching)
× P(source selection)
× P(extraction)
× P(action)The seven stages
| # | Stage | Question it answers | Signals |
|---|---|---|---|
| 1 | Discovery | Can the agent find the URL? | Sitemap, backlinks, index presence, llms.txt, mentions. |
| 2 | Crawlability | Is the agent allowed and able to fetch it? | robots.txt, HTTP status, no anti-bot walls. |
| 3 | Indexing | Is the source cached/indexed when the query arises? | Search and AI index coverage, freshness, canonical. |
| 4 | Matching | Does content match the query intent? | Intent pages, clear titles, user vocabulary. |
| 5 | Selection | Is this source chosen among rivals? | Authorship, dates, evidence, disambiguation, coherence. |
| 6 | Extraction | Can the agent read it uncorrupted? | Correct Content-Type, semantic HTML, Markdown, faithful JSON-LD. |
| 7 | Action | Can the agent do more than read? | API, MCP, captcha-free forms, machine-readable contact. |
The weakest link governs
The model is diagnostic: find the lowest-probability stage and fix it first. The 2026-06-12 audit of this site is the example — stage 6 (extraction) was near zero because /docs/*.md was served as application/octet-stream, and stage 1 (discovery) was crippled by relative links the agent's fetcher would not follow. No amount of good content elsewhere compensated those two near-zeros.
How to measure each stage
See the measurement page for the operational metric behind each stage.