Measurement

How to measure AI visibility

By Arnav Mukherjee, founder of TofuBofu · July 6, 2026

The first time a founder really gets it, the reaction is almost always the same: "Okay, so how do I even see this?" They have accepted that buyers are asking ChatGPT and Perplexity for recommendations, and that they might be invisible in those answers. But visibility feels like a fog. You cannot open a dashboard the way you open Google Search Console. So most people either guess, or they type their own company name into ChatGPT once, see something reassuring, and move on.

That one-off check is worse than useless, because it gives false confidence. Measuring AI visibility properly is not hard, but it is specific. It is a repeatable process with four parts, and once you have it, the fog turns into a number you can actually move.

First, know what you are measuring

An AI answer can put your brand in one of four states, and they are not the same thing. You can be named (you appear in the text), cited (you are linked or footnoted as a source), recommended (the engine actively suggests you), or ignored (you are absent while competitors are there). Being recommended for "best MSP for law firms" is worth far more than being cited once in a footnote. So the goal is not a yes-or-no "am I in AI," it is knowing which state you are in, for which questions, on which engine.

With that in mind, here is the method. Four steps.

Step 1: Build a query set of real buyer questions

This is where most people go wrong. They test their own company name. "Tell me about Acme IT." Of course the engine says something, because you handed it the answer. That measures nothing useful. What you want is the questions a buyer asks before they know you exist: "best managed IT provider for a law firm in Chicago," "alternatives to [named competitor]," "how do I stop my accounting firm's email from getting phished." Fifteen to thirty of those, drawn from how your actual buyers talk, is a solid set.

Include a mix: category questions ("best X for Y"), comparison questions ("X vs Y," "alternatives to X"), and problem questions ("how do I fix Z"). Those three cover the shortlist-building a buyer actually does.

Step 2: Ask each engine, separately

Run every query through ChatGPT, Claude, Gemini, and Perplexity, and keep the results apart. This separation is not optional busywork. The engines pull from different sources and barely agree: only about 11 percent of the domains ChatGPT cites also show up in Perplexity. You can be the top recommendation in one and completely absent in another. Averaging them into a single "AI score" hides exactly the gap you need to see. Score each engine on its own.

Step 3: Sample, because the answer moves

Ask ChatGPT the same question twice and you can get two different answers, with no change on your side. Large language models are non-deterministic, returning different responses to identical prompts even at their most deterministic setting. So a single run is a coin flip: you might be named by luck or missed by luck. Run each query three or more times and average whether you appear. That turns noise into a number you can trust, and it is the step hand-checking most often skips.

Step 4: Record the right things, and track over time

For each query and engine, record three things: whether you were named at all (your mention rate), how prominently (first and recommended, or buried at the bottom), and who showed up instead (your competitor set). Those roll up into a share of voice: your presence against the competitors answering the same questions. Then do the whole thing again next month. One snapshot tells you where you stand today. The trend line tells you whether your work is doing anything, which is the point.

One honest note. When I first ran this by hand on my own company, it was genuinely useful and genuinely tedious: four engines, thirty questions, several runs each, recorded in a spreadsheet, repeated monthly. Doing it once by hand is the best way to understand what you are measuring. Doing it every month, consistently, across six AI engines, is the part worth automating, and it is exactly what a visibility tool exists to do.

What to do