How to audit whether AI engines mention your brand

Someone asks ChatGPT for "the best option in your category," and it names three businesses. If you're one of them, you just won a customer you'll never see in your analytics. If you're not, you lost one — silently, with no log, no referral, no trace.

That's the problem with AI visibility: it's invisible by default. You can't audit it the way you audit traffic, because there's no traffic to audit. You have to go ask the engines directly, on purpose, and write down what they say.

Here's a repeatable method to do exactly that — and to turn a fuzzy "are we showing up?" into a number you can move.

Why you can't see this in your analytics

Your analytics show what happened after a click. AI answers happen before one — often instead of one. When ChatGPT recommends you in a conversation, there's no visit to attribute, no UTM, no referrer. The buyer forms an opinion, and you never appear in any report.

It gets harder. AI answers are non-deterministic: ask the same question twice and the shortlist can shift. Phrasing, timing, the user's region, and the model's own randomness all move the result. So a single check tells you almost nothing — a brand can be named in three runs out of five and absent in the other two.

That has one clear consequence: auditing AI visibility has to be deliberate and repeated. Not a one-time spot check, but the same prompts, across the same engines, on a schedule — so you're reading a trend, not a coin flip.

The audit, step by step

1. List the prompts your buyers actually use

Start from real intent, not vanity queries. Nobody types your brand name into ChatGPT to discover you — they describe a need. Write down the prompts a buyer uses before they know you exist:

• Category + location: "best [category] in [city]," "[service] near me that's open late"
• Comparison: "[you] vs [competitor]," "is [competitor] worth it"
• Alternative: "alternatives to [big competitor]," "cheaper option than [X]"
• Job-to-be-done: "how do I [the problem you solve]," "who can help me [outcome]"

Aim for 15–30 prompts that cover your real buying situations. These are your test set, and you'll reuse them every cycle — consistency is what makes the trend readable.

2. Run them across the engines that matter

Run every prompt through each engine your buyers use: ChatGPT, Perplexity, Gemini, and Google AI Overviews. They retrieve and ground differently, so the same prompt yields different shortlists and different sources on each — being named in one is no guarantee of the others.

Because answers vary, run each prompt more than once per engine (three is a sane minimum) and treat "named 2 of 3 times" as a real, recordable signal. One run is noise; a few runs is a measurement. Use fresh sessions so prior chat history doesn't bias the result.

3. Record the same four things every time

For each prompt-and-engine, capture four facts — nothing more, nothing fancy:

What to record	The question it answers
Mentioned?	Were you named at all in the answer?
Cited / linked?	Did the engine link to you, or just say your name?
Competitors	Which rivals showed up — and how consistently?
Sources	Which pages did the engine lean on to build the answer?

That last column is the gold. The sources an engine cites are the map to your fix list — they're the pages you need to be mentioned on. Log them in a simple sheet: one row per prompt-engine-run, those four columns, and a date.

4. Turn it into a visibility score

Roll the raw log into one number so you can compare cycles and engines at a glance. A simple, honest version:

• Mention rate — share of runs where you're named (a 0–100% per engine).
• Citation rate — share of runs where you're actually linked, not just mentioned.
• Share of voice — your mentions versus the competitors that keep appearing.

Average the mention rates across engines into a single visibility score, and keep the per-engine breakdown beside it. The exact formula matters less than using the same one every time. The score's only job is to move — up when your work lands, flat when it doesn't.

5. Build the gap list and prioritize fixes

The audit isn't the goal; the gap list is. For every prompt where you're absent or out-cited, ask why, then act on the highest-leverage gap:

• You're not on the cited sources. The single most common gap. The engine leaned on roundups, directories, or community threads that never mention you. Fix: earn mentions on those specific pages — pitch the roundup, claim the directory, show up in the thread.
• Your content isn't answer-shaped. You're on the right topic but buried in narrative. Fix: lead with the answer, use the buyer's question as a heading, keep facts clean and quotable.
• Your facts aren't machine-extractable. Inconsistent name, hours, or claims; no structured data. Fix: make schema.org and your core facts consistent everywhere so the engine can quote you confidently.

Work the gaps that block the most buyer prompts first. One missing roundup mention can cost you a dozen answers at once.

6. Re-measure on a schedule

Run the same prompts, same engines, same way — monthly is a sensible cadence — and watch the score move. GEO changes don't show up overnight; corroboration takes time to propagate across the web and into the models. A scheduled re-run is the only way to separate "the fix worked" from "the shortlist reshuffled this week."

What good and bad look like

Bad is invisible and unmeasured: you've never run the prompts, you assume good Google rankings carry over, and the engines name competitors you've barely heard of. Good is being named consistently — not once, but across most runs and most engines — cited with a real link, and trending upward cycle over cycle.

The most common gaps, in order: not being mentioned on the sources the engines cite; content that's relevant but not structured as a clean answer; and inconsistent, unverifiable facts the model can't safely repeat. Notice the pattern — almost every gap traces back to that fourth column from step 3.

The honest truth is that doing this by hand across four engines, dozens of prompts, and several runs each gets heavy fast — which is exactly the loop PromptHawk automates. But the method works whether a tool runs it or you do, so the real first move is simple: write down ten prompts your buyers actually use, and go see who the engines name.