How to Actually Measure Whether AI Recommends You

The shrug is the opportunity

Ask a business owner whether they show up in Google, and they can usually tell you. Ask whether they show up when someone asks ChatGPT, Gemini, or Perplexity for a recommendation in their category, and you'll get a shrug.

That shrug is the entire opportunity in GEO right now — and it's also the thing almost nobody is measuring, including the agencies charging for it. Here's the measurement framework we use, in enough detail that you could run a rough version yourself this afternoon. If you can measure it, you can't be sold a screenshot.

Measure recommendation, not recall

The first mistake is measuring the wrong thing. People check whether an AI knows they exist — "tell me about [my company]" — the model dutifully summarizes their website, and they feel reassured. That's not the metric that matters.

Customers don't ask AI to describe a company they've already chosen. They ask AI to choose for them: "who's a good X near Y," "what's the best tool for Z," "recommend a firm that handles W." The metric that matters is whether you appear in the answer to a neutral, customer-shaped question you did not name yourself in. Recommendation, not recall.

Build a prompt set that mirrors real demand

Step one is building a prompt set that mirrors real demand. Sit down and write the twenty to forty questions your actual customers would type — in their words, with their constraints.

Not "AIRAX GEO agency," but "how do I get my business to show up in AI search," "affordable SEO agency in Tokyo that speaks English," "is GEO worth it for a small firm." Include the qualifiers real people use: budget, location, language, industry, company size.

Test every engine, then repeat for a rate

Step two is testing across engines, because they don't agree. ChatGPT, Gemini, Perplexity, Claude, and the AI Overviews sitting on top of Google search each have different training, retrieval, and freshness. You can be invisible in one and cited in another. Where a tool exposes its sources — Perplexity is the clearest — record not just whether you appeared but what it linked to.

Step three, the one people skip: account for non-determinism by repeating. The same prompt can yield different answers on different runs, so a single check is noise, not data. Run each prompt several times, across fresh sessions, and record a rate — "we appeared in 3 of 5 runs for this query" — rather than a yes/no. That's a number you can watch move, and one a guarantee-seller can't fake, because it's an average over repetition, not a lucky screenshot.

Bucket: present, volatile, absent

What you do with the results turns measurement into strategy. Sort your prompt set into three buckets.

• Present — you show up consistently. Protect it, and understand why.
• Volatile — you show up sometimes. This is where small improvements to your facts, content, and third-party mentions pay off fastest, because the model is already on the fence.
• Absent — you never appear. Decide whether that query is worth fighting for, or simply isn't your customer.

Re-run monthly and watch the delta

Then re-run it monthly. The point of a baseline is the delta. If your citation frequency for the volatile and absent buckets isn't climbing over a quarter, whatever you (or your agency) are doing isn't working — and you now have the evidence to say so.

This is exactly the accountability the current market lacks: not "trust us, AI loves you now," but "here are the same forty queries, across the same five engines, run five times each, last month versus this month." That's the whole framework — neutral prompts; every engine; repeat for a rate; bucket present, volatile, absent; re-run monthly. Measure first. It's the cheapest thing you can do, and the one thing the hype can't survive.