How to Audit Your Brand in ChatGPT, Gemini, and Perplexity

An AI visibility audit shows how your brand appears when people ask AI tools about your category, competitors, products, and trust signals. The useful version is simple: build the right prompts, run them across engines, record mentions and sources, then diagnose whether the problem is visibility, credibility, clarity, or crawlability.

You can run the first version of this audit in an afternoon, with nothing more than a spreadsheet and free accounts on ChatGPT, Gemini, and Perplexity. This guide gives you the prompt set, the tracking template, and a way to read what you find. Nothing to buy, nothing to install.

Why audit all three engines, not just one

ChatGPT, Gemini, and Perplexity often answer the same question differently, because each one retrieves from a different place. ChatGPT's search retrieval leans on Bing's index. Gemini draws on Google's index and benefits from Googlebot's rendering. Perplexity runs its own crawler and its own ranking on top. Add different training data and the fact that these systems phrase answers probabilistically, and you get three engines that can disagree about your category on any given day.

How different are the answers, exactly? Honestly: it varies by query, by language, and by week, and any fixed percentage would be stale by the time you read it. Run the same question through all three engines and you will usually see the divergence yourself within the first two or three prompts. That is the point of the audit: a brand that looks strong in one engine can be invisible in another, and checking only one gives you false confidence. If you want the adoption data behind why this now matters for Indonesian brands, our Indonesia AI statistics reference keeps it sourced and current.

The seven prompt categories

A complete audit asks seven kinds of questions: discovery, trust, comparison, problem, local, expert, and purchase. Write two or three prompts per category for your own brand, then run every prompt in both Bahasa Indonesia and English, because engines answer each language from partly different sources. Phrase prompts the way a real customer types, not the way a marketer talks. The examples below imagine an Indonesian skincare brand; swap in your own category.

1. Discovery

Discovery prompts test whether engines surface you when nobody names you. This is the hardest category to win and the most valuable.

• "What are the best local skincare brands in Indonesia?"
• "Rekomendasi skincare lokal yang bagus untuk kulit berminyak?"

2. Trust

Trust prompts test what engines say when someone checks you out before committing.

• "Is [your brand] legit? What do reviews say about it?"
• "Apakah produk [your brand] sudah terdaftar BPOM?"

3. Comparison

Comparison prompts test how engines frame you against named competitors, and whose sources they trust for the verdict.

• "Which is better for sensitive skin, [your brand] or [competitor]?"

4. Problem

Problem prompts describe the customer's need without naming any brand, to see whether you appear as a solution.

• "My skin reacts to strong actives. What ingredients and brands should I look for?"
• "Skincare apa yang aman untuk ibu hamil?"

5. Local

Local prompts add a location qualifier, because answers change when geography enters the question.

• "Where can I buy good local skincare in Jakarta?"

6. Expert

Expert prompts borrow an authority's voice to test whether credible third parties are associated with your brand.

• "What do dermatologists in Indonesia recommend for beginner routines?"
• "Brand skincare lokal apa yang sering direkomendasikan dermatolog?"

7. Purchase

Purchase prompts sit at the bottom of the funnel, where a wrong or missing answer costs the most.

• "Is [your brand]'s vitamin C serum worth buying?"
• "Review [your brand] sebelum beli, apakah recommended?"

What to record: the tracking template

Record ten fields for every prompt you run, one row per prompt per engine. These ten columns are the whole template; copy them into any spreadsheet and you are set.

The template is deliberately boring. The value is in the discipline of filling it for every prompt and every engine, not in the design of the sheet.

How to read the results

Every row lands in one of four states: mentioned, cited, misrepresented, or absent, and each state means something different.

• Mentioned. Your brand is named in the answer. The engine knows you exist, usually from third-party coverage or training data. Good, but fragile if no sources back it.
• Cited. One of the linked sources is your site or a page about you. This is the strongest state: the engine is not just recalling you, it is actively reading and referencing you.
• Misrepresented. You are named, but the facts are wrong: an old address, a discontinued product, a mix-up with a similarly named company. Treat this as the most urgent state, because it usually traces to inconsistent information across the sources engines read.
• Absent. You do not appear where competitors do. Use the gap-hypothesis column to diagnose which problem binds: visibility (too little third-party coverage), credibility (nobody independent vouches for you), clarity (engines cannot tell who you are), or crawlability (your site is unreadable to AI crawlers).

Read the sources-cited column across all your rows before deciding anything. It shows you where each engine actually gets its category answers: which directories, which listicles, which publications. If those pages do not include you, that is the gap, and no amount of publishing on your own site fixes it alone. If most of your rows read absent, the usual causes and their fixes are in Why Your Brand Is Missing From AI Answers.

What a manual audit can't see

A manual audit is a snapshot taken from one chair, and three limits come with that. First, location: engines adapt answers to locale and language settings, so prompts run from outside Indonesia (or through a foreign VPN) may not match what your customers in Jakarta actually see. Second, consensus: the same prompt can return different answers minutes apart, so one run is one sample, and a fair read needs repeated runs with majority patterns. Third, trend: a single audit tells you where you stand, not which direction you are moving, and the monthly re-runs that build a trend line are where good intentions usually go to die.

None of this makes the manual version worthless. It makes it a starting point. If you would rather skip the spreadsheet, our free GEO Audit runs the automated version of this method and scores your brand's AI visibility in about 30 seconds. And if the results show gaps you want help closing, that diagnosis is exactly where our GEO practice starts.