How to track brand mentions in AI

Why a one-time check doesn’t hold

Two forces keep moving the answer. Models get retrained and swapped out, and each version has read a slightly different slice of the web. Meanwhile the web itself changes underneath them. A new review lands. A Reddit thread gets buried. A competitor ships a comparison page that reframes the whole category. Take a snapshot today and it starts going stale by the weekend.

The outputs wobble on their own, too. Same prompt, same afternoon, and you might see different wording or a different brand entirely. That wobble is built into how these systems work, and it’s the reason you watch a brand mention over time rather than trusting whatever one run happened to say.

What to record for each prompt

Pick the questions your buyers actually type. “Best [category] for [use case].” “Alternatives to [competitor].” For every one of them, write down four things.

• Whether you’re named at all. The blunt yes or no. This is the number that goes up or down, and everything else explains why.
• How you’re described. Being named as the budget option reads very differently from being named as the one buyers regret. Note the sentiment, not just the appearance.
• Which sources the answer leaned on. When an engine shows its citations, log them. Those are the pages doing the work, so they tell you what to reinforce and what to go earn.
• How your competitors did on the same prompt.Your slice of the answer is a competitive number, so track the brands beside you. That’s the raw material for share of voice.

Why doing it by hand stops working

Run one prompt once and you’ve learned almost nothing, because of the wobble. So you run each prompt a few times to see what’s typical. Then you repeat it across ChatGPT, Gemini, Perplexity and whatever else your buyers actually open. Then you do the whole thing again next week, because a single week is a dot, and you need a line.

Multiply your prompts by the engines, by the repeat runs, by the weeks, and the spreadsheet quietly turns into a part-time job. That’s the wall most teams hit. The work is simple; it just doesn’t fit in a human’s afternoon.

Pick a cadence, then freeze the prompts

Choose weekly or monthly and hold it. The other half of a clean comparison is a stable prompt set: if you reword the questions between runs, you can’t tell whether the answer changed or your test did. Lock the wording, change it only on purpose, and note the date when you do.

When a number moves, chase the reason, not just the direction. A drop from a fresh competitor page is a content problem. A drop after a model update is a patience problem. They look identical in the chart and call for opposite responses, which is why the “why” column earns its keep.

Tracking only counts if it feeds the work

The reason to do any of this is to catch a drop early, while it’s one prompt and not your whole category, and to learn which sources actually move your mentions so you spend your time on the pages that pay. A dashboard nobody acts on is just a slower way of guessing.

Start with today’s reading. A free AI visibility audit runs your real buyer prompts across the major engines and shows where you land right now. That first pass is the baseline every later run measures against.