How to Check If ChatGPT and Perplexity Actually Know Your Brand

Ask ChatGPT about your own company and you'll get one of three answers: an accurate description, a confident-sounding paragraph that's subtly wrong, or nothing at all. Most founders never run this test in a structured way — they ask once, feel either relieved or annoyed, and move on. That's not enough to know if you check if ChatGPT knows your brand in any way you can act on.

Here's a method you can actually run yourself in about 20 minutes, across the four systems that matter, with a way to tell a real mention from a hallucination.

Why "I asked ChatGPT once" isn't a real test

A single prompt tells you almost nothing. ChatGPT's answers depend on whether it's using parametric knowledge (facts baked into the model from training) or live web search, and those two modes can produce completely different answers to the same question about your brand.

If your brand is well established in the training data, you get a solid, specific description. If it isn't, you get hedging, vague generalities, or the model quietly answers about a competitor with a similar name instead. That's not a bug you can fix with better copy — it's a structural gap between what the model "remembers" and what it can look up.

So the test has to isolate which mode produced the answer, and it has to run more than once.

How to check if ChatGPT knows your brand: five prompt types, run per platform

Don't just ask "what is [brand]?" Run these five prompt shapes, each of which catches a different failure mode:

1. Direct identity: "What is [brand]? What does it do?"
2. Category recommendation: "What's the best [your category] tool for [your ICP]?" — this is the one that actually matters for revenue, because it tests whether you get recommended without being named first.
3. Feature-specific: "Does [brand] have [specific real feature]?" — catches feature-conflation, where the model attributes a competitor's feature to you or vice versa.
4. Comparison: "[Brand] vs [named competitor] — which is better for X?" — catches comparison-page averaging, where the model blends facts from a stale roundup article.
5. Adversarial: embed a false premise and see if the model corrects it. "Given that [brand] doesn't support [thing you actually support], what should I use instead?" A trustworthy answer pushes back on the false premise; a hallucinating one just runs with it.

Run each prompt at least twice, on different days if you can. Answers vary run to run — a single response is a sample, not a verdict.

Real mention vs. hallucination: how to tell the difference

This is the part most guides skip. A confident-sounding wrong answer looks identical to a correct one on the surface. Two checks separate them:

Check the citation, not just the claim. If the platform shows sources (ChatGPT Search and Perplexity both do this — hover or click the citation), open them. If the cited page doesn't actually say what the model claimed, that's a hallucination wearing a citation as a costume, not a verified fact.

Check for internal consistency across runs. Ask the same question three times. A real, retrievable fact stays roughly the same across runs. A hallucinated detail — a wrong founding year, a feature you don't have, a price that's off — tends to drift or contradict itself between attempts, because the model is generating a plausible-sounding answer each time rather than recalling one fixed fact.

If you catch a hallucination, the fix isn't "ask it to be corrected" (session-based, doesn't persist). The fix is making the correct fact easy to retrieve from a source the model actually trusts — your own site, a Wikipedia-style reference, or a page that ranks well enough to get pulled into search results.

Test the platforms separately — they don't share an answer

This is the step people skip because it's more work, and it's the one that matters most. ChatGPT, Perplexity, Gemini, and Google's AI Overviews pull from different source pools and use different retrieval mechanisms, so a brand can be well-represented on one and invisible on another:

| Platform | Primary mechanism | What it tends to favor | |---|---|---| | ChatGPT | Training data (parametric) + optional web search via partner search providers | Long-established, widely-referenced entities; Wikipedia and major outlets | | Perplexity | Live retrieval, source-heavy by default | Fresh content, forums like Reddit, recent pages | | Gemini / Google AI Overviews | Google's index, via query fan-out (your one question becomes several related searches, synthesized together) | Pages already ranking reasonably in classic Search, plus broader long-tail coverage | | Google AI Overviews specifically | Same index as organic Search — no separate schema or special markup required | Content that would already be eligible for rich results |

According to OpenAI's own documentation on ChatGPT search, when search is used, ChatGPT rewrites your query into one or more targeted queries sent to partner search providers, and shows inline citations you can click through — which is exactly the mechanism you're checking in the "real mention vs. hallucination" test above.

Perplexity documents its own retrieval behavior similarly — it treats structured, well-sourced pages as higher-confidence candidates for direct answer extraction, according to Perplexity's developer documentation. And Google's own AI features documentation confirms both AI Overviews and AI Mode may issue multiple related searches across subtopics before composing an answer — which is why a single-prompt test on Google misses most of what actually happens behind the answer.

Testing only one platform and assuming the others match is the single most common mistake in this exercise. Search Engine Land's guide to fixing brand hallucinations makes the same point from the correction side: the fix always has to be platform-specific, because the underlying retrieval is platform-specific. Run the same five prompts on all four.

Turning this into a repeatable check

Doing this by hand once is useful. Doing it every time you want to know if a launch or a new piece of content moved the needle gets tedious fast — which is exactly the workflow AI visibility tracking is built for: running your actual buyer-relevant prompts across assistants on a schedule and reporting where you're mentioned, where you're not, and who's getting recommended instead.

If you're new to the underlying concept, our guide to what AI search visibility actually means covers the fundamentals this method builds on.

Common mistakes when running this test

• Asking only brand-name prompts. "What is [brand]?" tells you if you exist. It tells you nothing about whether you get recommended when someone doesn't already know your name — the category-recommendation prompt is the one that actually predicts new customers.
• Testing once and calling it done. AI answers are generated per query and vary between runs. One good answer isn't a trend.
• Confusing a citation with accuracy. A model can cite a real page and still misstate what it says. Always open the source.
• Ignoring robots.txt. None of this works if GPTBot, ClaudeBot, or PerplexityBot are blocked from your site. Check with a free robots.txt tester before assuming your content is the problem — sometimes it's crawler access.

FAQ

How do I check if ChatGPT knows my brand? Run five prompt types — direct identity, category recommendation, feature-specific, comparison, and an adversarial false-premise prompt — at least twice each, and check whether cited sources actually support the claims made. A single prompt isn't a reliable test.

Why does ChatGPT give different answers about my brand on different days? Answers depend on whether the model is drawing from training data or live web search, and both can vary run to run. That's why the method above calls for repeating each prompt rather than trusting one response.

Do I need to test ChatGPT, Perplexity, and Gemini separately? Yes. They use different retrieval mechanisms and source pools — strong presence on one doesn't predict presence on another.

What's the difference between a hallucination and a real mention? A real mention traces back to a citation that actually supports the claim and stays consistent across repeated runs. A hallucination is a confident-sounding claim that either has no real source behind it or drifts when you ask again.

Run your first check today — pick five real prompts your buyers would type, and go through all four platforms before you decide whether your AI visibility is a content problem or a crawler-access problem.