How to check what AI says about your business (5-minute test)

Before investing anything in AI visibility, run this test. The core takes five minutes, requires no tools beyond a browser, and shows you exactly where you stand. Below is the full protocol we use as the first step of every audit — including the scoring sheet and the mistakes that make results lie to you.

Step 1 — Ask the assistants what your customers ask

Open ChatGPT (with web search enabled), Perplexity and Gemini, and give each of them two prompts:

1. The buying question: "best [your service] in [your city]" — phrased the way a customer would, not the way your industry talks. A dental clinic tests "best dentist for implants in Austin", not "premier dental implantology provider".
2. The brand question: "What do you know about [your business name]?" — and if your name is shared with other companies, add your city or domain to see whether the model can even tell you apart.

Record three things per answer: are you mentioned, are you cited (linked as a source), and is the information correct? Save screenshots — you'll want the before/after later.

Step 2 — Check whether AI crawlers can reach you

Open yoursite.com/robots.txt and look for lines blocking GPTBot, ClaudeBot, PerplexityBot or Google-Extended (the full agent lists are documented by OpenAI and Google). A surprising number of sites block AI crawlers by accident — a security plugin, a CDN's "block bots" toggle, a robots.txt template copied years ago. If a crawler is blocked, that system cannot read your site. No content strategy compensates for a locked door.

While you're there, confirm sitemap.xml is referenced and loads, and that your key pages return HTTP 200 rather than redirect chains.

Step 3 — Check your structured data

Run your homepage through validator.schema.org. You're looking for three levels, in order: any structured data at all; an Organization or LocalBusiness with your real name, contacts and services; and entity links (sameAs) pointing to your profiles elsewhere. If the validator finds nothing, AI systems must guess who you are — and they guess conservatively: they leave you out rather than risk being wrong.

Step 4 — Check what the model believes without the web

Ask ChatGPT the brand question again with web browsing off (or ask "what do you know about X from your training data"). This reveals whether your business exists in the model's learned knowledge — the layer that answers when browsing is unavailable, which is still most consumer usage. A business absent from training data but visible via search is winnable this quarter; absent from both means you're starting from zero and every week of delay compounds.

Step 5 — Check your facts are consistent across the web

Compare your site, Google Business Profile, top directories and review platforms: same name, same services, same address and phone, same pricing signals. Models cross-reference sources before stating facts; contradictions read as low trust, and a low-trust entity gets skipped or — worse — described with the wrong data. This is the step almost everyone fails without knowing it.

Scoring your results

Check	Pass	Points
Mentioned in ≥1 assistant's buying answer	Yes	25
Cited as a source (linked)	Yes	20
Brand info correct in all answers	Yes	20
No AI crawlers blocked	Yes	15
Valid Organization/LocalBusiness schema	Yes	10
Facts consistent across ≥3 external profiles	Yes	10

0–40: you're invisible and probably leaking customers to whoever the assistants name instead. 40–75: fixable gaps — the technical ones close in days. 75+: defend the position and start measuring monthly.

Four mistakes that make the test lie

1. Testing from a logged-in, history-rich browser. Assistants personalize. Use a clean session or incognito, or your history will show you a rosier picture than your customers see.
2. Testing once. Answers vary between runs. Ask each question twice; treat a mention that appears in one of two runs as "unstable", not "present".
3. Testing only your brand name. Customers who already know your name were never the problem. The buying-question results are the ones that pay rent.
4. Fixing content before access. If Step 2 fails, nothing else you do is visible. Order matters: access → facts → content → authority.

How often to re-test

Monthly, same questions, same clean-session protocol, results in the same sheet. AI platforms ship changes constantly; a stable monthly delta is the only honest KPI of AI visibility. That cadence — baseline, fix, re-measure — is what we publish about our own site in our self-audit case study.

FAQ

Which assistant matters most for my business?

Where your customers already are: consumer services see the most ChatGPT and Google AI Overviews exposure; research-heavy B2B sees more Perplexity. Test all three — the fixes overlap almost entirely.

My site is new and appears nowhere. Is that fixable?

Yes — that's the normal starting state. Technical access and structured data are fixable in days; mentions and citations build over months. See the five reasons AI recommends competitors for the usual culprits in priority order.

What if AI says something wrong about us?

Find the source it's echoing — usually an outdated directory listing or old page — fix it there, and strengthen correct facts with structured data. Models repeat the most consistent version of you they can find.