Most founders I speak to have never once asked an AI assistant about their own brand. They have rank trackers, alert tools and analytics dashboards for Google, but the conversation where a growing share of buying decisions now start, a buyer typing "what's the best tool for X" into ChatGPT or Gemini, goes completely unwatched.
You don't need a budget to fix that. You need an afternoon, a spreadsheet and a bit of discipline. This is the audit I run, and it works whether you sell software, consumer products or services.
## Step 1: Write down the questions buyers actually ask (30 minutes)
Open a spreadsheet and list 15 to 25 questions a real buyer would ask on the way to purchasing something in your category. Not keyword strings. Full questions, in the buyer's words.
A few patterns that matter:
- Category questions: "What's the best electric bike for city commuting under 80,000 rupees?"
- Comparison questions: "Is X better than Y for a small team?"
- Problem questions: "How do I stop my invoices from being chased manually?"
- Brand questions: "Is [your brand] reliable?" and "Who owns [your brand]?"
- Switching questions: "What do people use instead of [the category leader]?"
Pull phrasing from your sales calls, support tickets and reviews, not from your keyword tool. AI questions are longer and more conversational than search queries, and the wording changes the answer.
## A few traps to avoid
Three mistakes skew most first audits. The first is testing while logged in: assistants personalise, and a model that knows you founded the company will flatter you. The second is asking leading questions. "Why is [your brand] the best?" tells you nothing; the buyer's question is "what's the best option for my situation", and the honest answer to that is the one you need. The third is arguing with the model in the same chat. Correcting it mid-conversation changes that conversation, not the underlying knowledge, and it makes your notes useless. Record, don't debate.
## Step 2: Ask every assistant, logged out (60 minutes)
Take your list to ChatGPT, Gemini, Claude and Perplexity. Ask each question in a fresh chat, logged out or in a clean session, because personalisation and history can skew what you see.
For each question and each assistant, record three things in your spreadsheet:
1. Are you mentioned at all? Yes or no.
2. If yes, what is said? Copy the exact sentence. Note whether you are recommended, listed neutrally, or mentioned with a caveat.
3. Who else is named, and in what order? Order is a rough proxy for confidence.
Expect this to be sobering. Most brands discover they appear in fewer than a third of the category questions they assumed they owned, and the comparison questions are often answered with two-year-old information.
## Step 3: Trace the sources (45 minutes)
This is the step that turns the audit from interesting to useful. For the answers that mention you wrongly, or don't mention you at all, ask a follow-up: "Where did you get that?" or "What sources describe [brand]?"
Perplexity will show citations directly. The others will usually name the publications, review sites, forums or directories they drew from if you ask. Write those down. You will start to see the same handful of domains repeating: a couple of review platforms, a Reddit thread, a listicle from 2023, your own comparison page, a directory you forgot you were listed on.
That list is your actual AI marketing surface. Those are the pages teaching the models what to say about you.
## Step 4: Mark the gaps and the errors (30 minutes)
Go through the spreadsheet and tag every row:
- Absent: you should be in this answer and you are not.
- Wrong: the answer contains a factual error (old pricing, dead feature, wrong positioning).
- Stale: the answer is directionally right but describes a version of you from a year or more ago.
- Good: accurate, current, and roughly how you would describe yourself.
The absent and wrong rows are your work queue. The good rows tell you which sources are doing their job, so you can support rather than disturb them.
## Step 5: Fix the cheapest things first (the rest of the afternoon)
You will not fix everything today, and you don't need to. The highest-yield fixes are usually boring:
- Update the directory and review profiles the assistants cited. Half of them will have descriptions you wrote in a hurry two years ago.
- Correct factual errors at the source. If a review site has your old pricing, ask them to update it. Most will.
- Add or refresh the pages on your own site that answer the comparison and problem questions directly, in plain language. Assistants lean heavily on clear, well-structured first-party pages.
- If a single outdated listicle is poisoning several answers, it is worth an email to the author with the corrected facts.
## Step 6: Set a reminder to do it again
AI answers drift. Models retrain, new reviews pile up, competitors publish. Put a monthly reminder in your calendar to rerun steps 2 to 4 with the same question list, so you can see movement rather than anecdotes. If you want the monitoring to run continuously instead of monthly, that's exactly the problem I built Vouch to solve; it tracks what AI assistants say about your brand over time. You can find it at vouchgeo.com.
## What a good result looks like
After one afternoon you will not have "fixed AI". You will have something better: a baseline. You will know where you show up, what is being said, which sources are shaping it, and which three or four fixes will move the most answers. Most brands are flying blind on all four. An afternoon is a cheap price for stopping. And unlike most marketing work, every fix you make compounds: the sources you correct today keep teaching the next model, and the one after that.
Disclosure: I am the founder of Vouch (vouchgeo.com), a tool that monitors what AI assistants say about your brand.

