How to Check If ChatGPT Mentions Your Brand (Free AI Visibility Audit)

A growing share of buying research now starts as a question to an AI assistant: “best email platform for a Shopify store,” “alternatives to [the incumbent] for a small firm,” “is [your company] legit?” Those conversations produce no referral traffic, no search-console row, no ad impression — the answer either includes you or it doesn’t, and you never find out which. So the question behind this post is a fair one: does AI actually know my brand exists, and what does it say?

You can answer it yourself, by hand, today. The method below works; it also has real limits worth understanding before you trust one good (or bad) answer. At the end there’s a free audit that does the first pass in one step — but the manual version teaches you what the results mean, so start here.

The manual check: five steps

1. Ask buyer questions, not vanity questions

“What do you know about [my company]?” is the vanity prompt, and it’s the least informative one — a model can retrieve something about almost any company with a website once named. The prompts that decide revenue are the ones where you are not in the question: category prompts (“best [what you sell] for [who you serve]”), problem prompts (“how do I [job your product does]”), and comparison prompts (“[rival A] vs [rival B] — what else should I look at?”). Whether you appear unprompted in those answers is your actual AI visibility.

2. Use a clean session

Assistants personalize. If you’ve discussed your own company in past chats, memory and conversation context can surface it in answers a stranger would never see. Turn memory off or use a temporary chat, and phrase prompts as a buyer would — no insider vocabulary, no product names only your team uses.

3. Repeat across engines

ChatGPT, Perplexity, Claude, Gemini, and Copilot draw on different training data, different browsing behavior, and different citation habits. Being invisible in one and prominent in another is common — Perplexity in particular browses and cites by default, so it tends to reflect the current public record, while a non-browsing answer reflects training-data memory that can run months behind. Check at least three engines before concluding anything.

4. Ask the same question more than once

Answers are sampled, not looked up. The same prompt to the same engine can name you on the first run and omit you on the second. One appearance proves you’re possible; it doesn’t prove you’re reliable. Three to five runs per prompt gives you a rough frequency, which is the number that actually matters.

5. Record what it says, not just whether you appear

When you do appear, read closely: does the description match what you sell today, or an old positioning? Does it name the right audience? Does it attach you to the right competitors? A stale or wrong description in a buyer’s chat can cost you the deal you never knew you were in. Save the answers with dates — this is your baseline for next quarter.

Why one chat session proves less than it seems to

The manual check’s weakness isn’t effort, it’s interpretation. A single session confounds four variables — engine, sampling, personalization, and time — and shows you the output without the inputs. If you’re absent, it can’t tell you why; if you’re present, it can’t tell you how durable that is. To act on the result you need to know what the answer was built from.

What actually determines whether AI knows your brand

AI answers about companies are assembled from the public record — training data plus, for browsing engines, whatever public sources a live search surfaces. Three inputs dominate:

• Third-party corroboration.Engines lean on independent pages — review sites, comparison articles, directories, press — more than on your own site, because that’s how they hedge against self-description. A brand only its own website describes is a brand AI describes tentatively or not at all.
• Consistency of the record.If your site, your directory listings, and the articles about you describe three different companies, the model’s summary inherits the confusion. Old positioning that was never cleaned up keeps getting served to buyers indefinitely.
• Crawl access. Browsing engines can only read what your robots.txt lets their crawlers read. Plenty of sites blocked GPTBot and friends in 2023 as a precaution and never revisited the decision — which for a discovery-stage brand mostly means invisibility, not protection.

The one-step version

The audit half of this — what the public record says about you, and how an AI engine reading it is likely to describe you — is mechanical enough to automate. Dozier’s free AI visibility checkruns it on any company domain: enter the domain, wait a couple of minutes while it reads live public sources, and get a headline read plus the top findings, each with the number of sources behind it. No account, no email, no API key; it says plainly that it’s a point-in-time snapshot, and it’s capped per visitor per day so it stays available. You can also run the same audit from inside your assistant via the free MCP server.

Run it on your own domain first, then on a rival you consider well-known — the gap between the two reads is the clearest picture of where you stand.

If the answer is “AI doesn’t know you”

The fix depends on which input failed. A thin record needs more independent pages describing what you do — the unglamorous work of listings, comparisons, and press. An inconsistent record needs the cleanup pass nobody scheduled. Blocked crawlers need a one-line robots.txt decision made deliberately instead of by default. And in every case the result is worth re-checking on a cadence, because models update, browsing surfaces new sources, and your competitors keep publishing. That last part — noticing when the answer about you changes — is what Dozier does continuously, with every finding cited to the source that produced it. The same discipline applies to researching rivals with an assistant, which is its own minefield — ChatGPT for competitor research covers those failure modes.