GEO Audit: How to Check What AI Actually Sees

What this helps you with

Most teams "optimize for AI" without checking what AI actually receives. A GEO audit answers three questions:

1. Can AI read your brand and products? 2. Is the data consistent (price, specs, policies)? 3. Are there evidence-backed fixes you can apply safely?

Three steps

1. Pick pages: start with your top 10 product pages plus the homepage, FAQ and one policy page. 2. Collect evidence: JSON-LD, FAQ anchors, raw-md and visible text — from the rendered page, not the theme source. 3. Score and fix: list gaps by impact (blocking, warning, nice-to-have), then fix and re-collect.

What a good score looks like

A useful GEO audit is not a single number. It has four parts:

• Readability: crawlers are allowed; pages load; raw-md exists.
• Structure: Product/Offer or page-type schema is valid; FAQ is consistent.
• Consistency: the same SKU has the same price, weight and return window everywhere.
• Freshness: evidence has a timestamp; fixes are re-checked.

If your audit cannot show the source and time of each finding, it is a guess, not an audit.

Want the full method? Read the expert version: geo-audit.expert

Chapter 1: Evidence model

1.1 First-party, not simulated

AI crawlers do not execute JavaScript. An audit that only reads theme source or uses a generic crawler will miss what a browser-rendered page exposes. Collect from a real browser:

• rendered JSON-LD;
• FAQ anchors and counts;
• raw-md view (if deployed);
• canonical and hreflang.

Store the raw payload with a hash and timestamp. Re-validate on the server; never trust a client-provided passed flag.

1.2 Four scoring dimensions

1.3 Gates

A GEO audit should gate changes, not just report:

• P0 inconsistency or invalid schema → block publication;
• P1 missing raw-md / partial FAQ → warn, allow with approval;
• P2 style/coverage suggestions → backlog.

Chapter 2: Reproducibility and operations

2.1 Re-run discipline

• fixed URL list and fixed template versions;
• same collection method (browser + Accept headers);
• record engine/model versions if AI answers are involved;
• keep raw evidence for at least 90 days.

2.2 What makes an audit trustworthy

1. every finding links to evidence; 2. every number has a timestamp; 3. “unmeasured” is reported, never converted to pass; 4. fixes have snapshots and rollback.

2.3 Common mistakes

• auditing only the homepage;
• treating llms.txt as a silver bullet;
• measuring rankings without checking citations;
• changing themes without re-collecting evidence.

Back to beginner version:geo-audit.novice