---
title: Evidence Model
description: What SeeLLM observes directly, what it samples, and what it infers.
audience: [human, agent]
type: concept
version: v1
stability: stable
lastVerified: "2026-08-01"
prerequisites: []
related: [/docs/concepts/ai-traffic-types/, /docs/guides/weekly-trends/, /docs/api/analytics-endpoints/]
artifacts: []
---

## Why the distinction matters

SeeLLM combines several kinds of evidence. They answer different questions and must not be treated as interchangeable. A crawler request proves access to a URL; it does not by itself prove that an AI answer cited, quoted, or sent a person from that page.

## Evidence types

### Crawler visit

An observed server-side request classified as an AI crawler using request metadata such as user agent, network information where available, headers, and behavior. User agents can be spoofed, so high-stakes decisions should use verified vendor IP data when the vendor publishes it.

### AI retrieval

A user-directed or search-oriented fetch when the platform exposes a distinguishable request signal, such as `ChatGPT-User` or `Claude-User`. Retrieval is stronger evidence of live use than a training crawl, but it still is not proof of a visible citation.

### Sampled citation

A monitored URL appeared as a source in a documented sample of generated answers. Citation checks are samples: results can vary by prompt, engine, geography, model, account, and time.

### Quoted text

Text from a monitored page matched text in a sampled answer using SeeLLM's matching method. The result must include the matched passage, sample context, and confidence or matching rule.

### AI referral

A human request reached the site with a detectable referrer from an AI product. Some clients strip referrers, so measured referrals are a lower bound rather than a complete count.

### Crawled, not cited

Crawler evidence exists, but the page was absent from the stated citation sample. This identifies a gap worth investigating; it does not prove that the page is never cited.

### Source gap

Another source appeared in the same sampled answer set where the monitored page did not. A source gap is tied to the recorded query set and observation window.

## Provenance shown with every result

SeeLLM should identify whether evidence is **observed**, **sampled**, or **inferred**, together with its URL, platform, observation window, and collection method.

## Known limits

- Crawler identities can be spoofed unless network ownership is verified.
- AI products can strip referral headers.
- Citation results are non-deterministic and must be sampled repeatedly.
- A crawler visit cannot establish how a model used the fetched content.
- Absence of observed evidence is not proof that no activity occurred.