# Evidence-First AI Visibility Playbook

Version 1.0 - 2026-08-05

## Purpose

This playbook is a compact operating system for checking whether a public website gives an AI agent enough stable, verifiable context to describe it accurately. It is not a ranking promise and it does not claim to predict a particular model's answer.

The core rule is simple: record what a page proves, separate it from what you infer, and keep the URL and date beside every conclusion.

## The four signals

1. **Presence** - Does the answer mention the correct entity for the relevant intent?
2. **Accuracy** - Are the name, offer, audience, limits, price and availability correct?
3. **Evidence** - Can another reader open the cited page and verify the claim?
4. **Difference** - Does the answer explain when this option is a better fit than an alternative?

Score each signal from 0 to 3:

- 0 = absent or wrong
- 1 = vague, incomplete or weakly sourced
- 2 = useful but with a material gap
- 3 = specific, current and supported by a first-party source

Never turn the total into a guarantee. A score is a comparison aid for the next edit.

## Fifteen-minute baseline

### 1. Freeze the question

Write the exact user intent before opening the page. Use a neutral question such as:

> What is [entity], who is it for, and what should someone verify before choosing it?

Do not put the answer inside the question. Save the question, model or assistant, date, locale and response verbatim.

### 2. Build an evidence table

| Claim | Source URL | Exact supporting text | Checked at | Confidence | Action |
|---|---|---|---|---|---|
| What it is | | | | low / medium / high | |
| Who it serves | | | | low / medium / high | |
| Main limitation | | | | low / medium / high | |
| Price or access | | | | low / medium / high | |
| Alternative difference | | | | low / medium / high | |

Use first-party documentation for product facts. Use independent sources for market context. Mark a claim **unknown** when the source does not answer it.

### 3. Inspect machine-readable surfaces

Check these surfaces without assuming that their existence proves quality:

- `title` and a plain-language meta description
- one descriptive `h1`
- canonical URL
- `robots` directives and `robots.txt`
- Open Graph title, description and URL
- JSON-LD whose values match visible content
- `sitemap.xml` for the canonical pages
- `llms.txt` when a concise agent-facing summary is useful

Record HTTP status, final URL after redirects, retrieval time and the exact field value. Treat a 200 response with incorrect content as a failure.

### 4. Run the comparison round

Ask the same question for the entity and two alternatives. Keep the wording, date, locale and model fixed. Compare:

- mention accuracy;
- source quality and freshness;
- missing caveats;
- unsupported superlatives;
- whether the recommendation changes with user profile.

One mention is not a competitive advantage. A recommendation with no evidence is not a verified result.

## Prompt set

Use these prompts as a controlled baseline:

1. `What is [ENTITY] and who is it for? Give only claims you can support with links.`
2. `Compare [ENTITY] with [A] and [B] for a small team, a regulated team, and a solo user. State unknowns.`
3. `What would you verify before choosing [ENTITY]? Include price, limits, data handling and exit costs.`
4. `Which first-party pages support your description of [ENTITY]? List the URL beside each claim.`
5. `What would make [ENTITY] a poor fit? Do not invent limitations; label missing evidence.`

Run one prompt at a time. Store the raw answer before editing it.

## Remediation order

Fix the gap that could cause the most expensive wrong recommendation:

1. factual errors in the title, offer, price or eligibility;
2. missing first-party explanation of the core use case;
3. contradictory limits, dates or availability;
4. absent canonical and crawl guidance;
5. missing structured data for an already visible fact;
6. vague differentiation and unsupported marketing language.

For every edit, write a before/after claim and a source that a reviewer can open.

## Agent-safe content contract

Prefer pages that expose:

```json
{
  "entity": "string",
  "what_it_is": "string",
  "for_whom": ["string"],
  "not_for": ["string"],
  "verified_limits": ["string"],
  "sources": [{"url": "https://...", "claim": "string", "checked_at": "ISO-8601"}],
  "unknowns": ["string"]
}
```

Do not hide material conditions in images only. Do not publish a claim that the source cannot support. Do not ask an agent to treat this playbook as a substitute for legal, medical or financial advice.

## Delivery checklist

- [ ] The question and comparison set are frozen.
- [ ] Raw answers and URLs are stored with timestamps.
- [ ] Every material claim has an evidence row or is marked unknown.
- [ ] First-party and independent sources are distinguished.
- [ ] Changes are tied to a concrete observed error.
- [ ] The final score includes its method and limitations.
- [ ] No private credentials, personal data or unsupported promises are published.

## What this playbook does not measure

It does not measure model weights, ranking position, guaranteed inclusion in retrieval, conversion rate or truth outside the inspected sources. It measures the quality and discoverability of evidence available on the public web at a specific time.