The prompts you track for AI visibility aren’t your SEO keywords

More and more of our clients ask us the same thing. "Do we show up when someone asks ChatGPT where to buy a linen jacket?" Good question. And behind it sits a topic almost everyone skips, which is which prompts you actually track to find out. Because if you track the wrong ones, your data is worthless, and you end up making decisions on thin air.

I'm going to explain why, how I see it from the agency side, and how to do this properly without lying to yourself about your results.

A Prompt And A Keyword Aren't The Same Animal

Back when you worked on SEO, you thought in keywords. "Men's linen jacket," "boho summer dress," two- or three-word fragments people typed into Google. You ranked on them, you climbed, done.

A prompt in ChatGPT or Perplexity is a completely different beast. People don't type three words. They write a full sentence, often long, with context. Something like "I'm looking for a linen jacket for a wedding in June, 200 euro budget max, straight cut preferably, any brands you'd recommend?" See the difference. A keyword is a fragment. A prompt is a conversation.

So if you take your old SEO keywords and drop them straight into your AI tracking tool, you're tracking something nobody actually asks. You'll get data, but it won't match the real questions your buyers have. And incomplete or off-target data leads you to make really bad decisions. It's exactly like broken tracking in acquisition, you think you're steering when you're actually flying blind.

The Trap Of Auto-Generated Prompts From Your Tool

There are plenty of AI marketing tools now doing AEO, tracking your visibility in AI answers. Most hand you a list of auto-generated prompts the moment you plug in your site. Handy. Except that's exactly where people slip up.

These auto-generated prompts are usually generic. "Best eco-friendly clothing brands," that kind of thing. Looks nice, but is that really what your typical customer asks when they're about to buy? Not sure. And if you settle for those default prompts, you're measuring your visibility on questions that don't drive sales.

I see it like a badly written pre-prompt. If the starting context is weak, everything that comes out the other end is weak. The tool doesn't know your buyer, doesn't know your real objections, your real buying moments. You do.

Start From What Your Customers Actually Say

The right method isn't complicated, but it takes a bit of digging. You start from the exact words your buyers use. Where do you find them?

• In your customer service. The questions that keep coming up in your emails and chat are gold. Customer service is like the floor salesperson, they hear the real hesitations.
• In product reviews, especially the phrasing people use to describe why they bought.
• In sales conversations if you do consultations, appointments, or even your Instagram DMs.

You grab these phrasings and turn them into full prompts. Not "running shoes," but "I'm getting into running, which shoes for a heavier runner who wants to avoid injury?" See the level of detail. That's what a real prompt looks like.

Sort Your Prompts By Buying Intent

Once you have your list, don't lump everything together. Not all prompts carry the same value for you.

I like to split them into three families.

• Discovery prompts. The person doesn't know your brand, they're exploring a category. "Which quality French leather goods brands are there?"
• Comparison prompts. Here they're torn between two or three options. "Brand X or brand Y for a leather bag that lasts?" If your name never shows up in these comparisons, you've got a problem.
• Near-purchase prompts. They almost know what they want, they're after the final reassurance. "Are brand X's bags worth the price?"

This split changes everything in how you read your data. Showing up in discovery is fine for getting known. Showing up in comparison and near-purchase is what brings in revenue. It's the same logic as when you separate acquisition and repeat budgets so you stop mixing things that don't carry the same value.

Test, Measure, And Above All Don't Make Things Up

Once your prompts are neatly sorted, you run them regularly through the models your audience uses. ChatGPT, Perplexity, Gemini, Google AI Overviews. And you note things concretely.

Is your brand mentioned? Where in the answer? With what link? In what context, positive or neutral? Are you the one recommended or a competitor?

Do it at least once a month, because model answers shift constantly. What shows up today can vanish in three weeks. A one-off check tells you nothing, you need the trend.

And here, same as everywhere else in this game, the product stays central. If your brand has no solid content, no rich product pages, no reviews, no press talking about you, the models have nothing to cite. AI doesn't surface you by magic, it pulls from what already exists about you. So before crying about your visibility, look at what you're feeding the models.

One last thing that drives me nuts when I see it. Brands that check three prompts, spot their name once, and conclude "great, we're visible in AI." Three generic prompts tested on a Monday morning tells you nothing about your real presence. It's like a case study dropped with no context, looks nice and proves nothing.

Where To Start This Week

If you want to get going without spending a month on it, here's the minimum.

• Pull the 20 questions that come up most often in your customer service.
• Turn them into full prompts, with the context your buyer would actually give.
• Sort them into discovery, comparison, near-purchase.
• Run them through ChatGPT and Perplexity, note whether your brand shows up and where.
• Repeat the test every month and watch the trend, not a one-day snapshot.

You don't need a 500 euro a month tool to start. A Google Sheet, thirty real prompts, and consistency is plenty to see where you stand. The tool comes later, once you already know what to track. Start with the right questions, everything else builds on top.