An AI visibility audit is a structured check of whether AI assistants can find, understand, and recommend your brand — and who they recommend instead of you. Run through the ten steps below and you'll leave with a concrete list of what to fix, ordered by impact.

Do the steps in order: the first five check whether AI systems can read you at all, the last five check what they actually say.

Part 1 — Can AI systems read you?

1. Fetch your homepage without JavaScript

Most AI crawlers don't execute JavaScript. Fetch your homepage the way they see it:

curl -s https://yourdomain.com | sed 's/<script[^>]*>.*<\/script>//g'

If your value proposition, product names, and key copy aren't in that output, they don't exist for AI systems. This single check catches the most damaging and least visible failure mode — a beautiful site that is literally blank to machines. (We've been bitten by this ourselves; it's why server-side rendering is step one, not a nice-to-have.)

2. Read your homepage as a stranger

Can someone who has never heard of you answer three questions from the first screen of text: What do you do? Who is it for? What makes you different? If a human can't, a model can't. Vague taglines are the enemy — models recommend brands they can describe.

3. Check your structured data

Look for JSON-LD on your key pages (view source, search for application/ld+json). Minimum bar: Organization on the homepage, Product or Service where relevant, FAQPage if you have an FAQ. Validate with Google's Rich Results Test. Structured data turns inference into fact.

4. Verify crawl plumbing

robots.txt — are you accidentally blocking AI crawlers (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot)?
sitemap.xml — current, complete, and free of noindex pages?
Canonical tags — one canonical hostname (www vs apex resolved with a 301)?

5. Confirm your key pages are indexed

site:yourdomain.com on Google, and check Search Console coverage. Retrieval-based AI engines lean on search indexes; a page that isn't indexed can't be retrieved into an answer.

Part 2 — What do AI systems actually say?

6. Ask the four assistants your money questions

Take the 5–10 questions a buyer would ask ("best X for Y", "alternatives to Z", "is [your brand] good for…") and put them to ChatGPT, Gemini, Claude, and Perplexity. Record: Are you named? How early? What exactly is said? One round is a snapshot, not a measurement — but it will already show your biggest gaps.

7. Check for hallucinations

While you're there, look for wrong facts: outdated pricing, discontinued products, confusion with similarly named companies. A brand hallucination repeated to thousands of users is a real business problem, and you only find it by looking.

8. Map who appears instead of you

For every question where you're absent, note which brands ARE named. That's your evoked set gap — and the citation sources behind those answers are a literal to-do list of the directories, reviews, and comparison pages you need to be in.

9. Audit your third-party footprint

Search for your brand plus your category terms. Are you in the roundups, directories, and review sites the AI engines cited in step 8? Models trust consensus across sources far more than your own site's claims.

10. Establish a repeatable baseline

A one-off audit decays fast — AI answers shift by platform, phrasing, and week. Decide the prompt set, run it across every platform on a schedule, and track visibility, rank, and share of voice over time. This is the step where a spreadsheet stops scaling — and exactly what geopsy automates: the prompt set, the four platforms, the scoring, and the trend history.

What to fix first

Triage in this order: readability failures (steps 1–5 — cheap to fix, they gate everything), then absence on high-intent questions (steps 6, 8 — usually a third-party-footprint problem), then hallucinations (step 7 — targeted corrections at the sources models cite).

New to the discipline? Start with what GEO is, then the tactical follow-up: how to rank in ChatGPT and AI search.