How to track what AI says about you, and whether to pay for it

You can measure this yourself, in an afternoon, for nothing.

The harder question is what to measure. Three things get tracked. Whether an assistant names you at all. Where you land in the list when it does. How it describes you. Two of the three are close to noise.

Presence holds still. Position and tone do not

Rand Fishkin ran the experiment nobody selling these tools had run. He worked with Patrick O'Donnell of Gumshoe. Their 600 volunteers put 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times.

Ask any of them for recommendations 100 times. The odds of getting the same list twice are under 1 in 100. The same list in the same order is closer to 1 in 1,000. Fishkin started the project convinced the whole category was a waste. His conclusion: any tool reporting your ranking position in AI is selling a number that does not exist.

Presence is a different story. A vendor study published in June 2026 tracked 102,025 prompt responses across 102 brands. It followed each brand-and-question pair across repeated runs.

About 78% of those pairs were strictly always mentioned or never mentioned. Only 7% flipped in between. Whether an assistant names you on a question behaves more like a fixed fact than a coin toss. That held across a few months of runs.

Tone sits at the other extreme. In the same dataset, sentiment flipped between runs about seven times more often than presence did.

So the metric worth your attention is how often you get named across many runs of many questions. Position is noise. Tone needs a much bigger sample before it means anything.

The panel

The manual version is a spreadsheet and an hour a month.

• Write 10 to 20 questions your buyers would ask. Use their words, not yours. Cover three shapes. The broad question: "who are the best people for X." The specific problem: "I have this situation, who handles it." The comparison: "X or Y for someone in my position."
• Keep your own name out of every question. Ask an assistant about you by name and it will recognize you almost every time, which proves nothing. In the study above, prompts naming the brand returned a mention 94% to 100% of the time. The unbranded question is the one that measures anything.
• Run each question at least three times per assistant. Single runs are reliable for the clear cases and misleading for the ones in the middle. The middle is where any change shows up first.
• Treat each assistant as a separate market. The sources they draw on barely overlap, so ChatGPT, Claude, Gemini and Perplexity each need their own column.
• Log four things per run. Were you named, who else was named, in what words you were described, and which sources the answer cited.
• Turn memory off, or use a temporary chat, and sign out where you can. Your own account has months of context about you and will flatter you.
• Run it monthly. These systems absorb changes over weeks, so weekly checking measures the weather and quarterly checking misses the season.

Your number is the share of runs that named you, per assistant. Watch it against the names that keep appearing instead of yours, which is the more useful column.

What the panel cannot tell you

It cannot tell you that an assistant sent you a client. Most of these conversations end without a click, so the visit never reaches your analytics. The person who arrives by name a week later looks like direct traffic. The workable read is that presence is a leading indicator and your consultation requests are the outcome.

It cannot separate a real change from a model update. When your number moves, the cause might be your own work or a new version of the model. Nobody outside those companies can tell you which.

And a low number tells you nothing about why. The three reasons a well-known name goes missing look identical on a dashboard. The answer text is where the diagnosis lives.

Whether to pay for a tool

The market is young, crowded, and moving. Entry-level products start around $29 a month. Mid-market platforms run roughly $95 to $300, and enterprise plans are quoted in the thousands. Fishkin cites an estimate of more than $100 million a year already spent on this kind of tracking.

What a tool buys you is scale and continuity. Hundreds of prompts instead of 20, daily runs instead of monthly, several assistants at once. Competitor sets tracked beside you, and a chart you did not have to maintain. For a practice tracking one name in one category, that is a convenience.

What no tool buys you is movement. Every product in this category is a dashboard. The work that changes the number happens elsewhere, mostly on other people's websites.

So run the manual panel first and act on what it shows. Buy a tool when the checking has become the bottleneck, not before. Before you pay, ask the vendor how many runs sit behind each number. Ask whether they have published the math. That was Fishkin's request to the industry, and the ones who cannot answer it are the ones to avoid.

One more check is worth learning, because this field runs on borrowed prestige. When a sales page cites a study, read what the study measured.

Take a paper that circulated widely in early 2026. Researchers at MIT and IBM showed that the public leaderboards ranking AI models against each other are fragile. Removing 0.003% of the user votes was enough to change which model sat at the top.

Real finding, careful work, and about something else entirely. Those leaderboards rank models by how well they answer. The votes come from people comparing two models side by side. Nothing in that paper concerns whether an assistant names your business when a buyer asks a question.

The two subjects share a word. Ranking. That is the whole overlap. A citation resting on a shared word is a citation nobody read.

Write your 10 questions first, in your buyers' words. If you cannot fill the list, that is the first finding, and no subscription fixes it.