How to track brand mentions in ChatGPT

Why one check tells you almost nothing

ChatGPT is not deterministic. Ask it "best AI visibility tools" twice and you can get two different lists. We measured this on our own panel: the same ten questions run twice on the same day moved individual answers by 1.86 points on average and the overall score by 0.59 (PROXY, July 31, 2026; the record is in the methodology changelog). That is small at the run level and large enough at the answer level that a single screenshot proves nothing. Whether ChatGPT has web search on also changes the answer: with search off it answers from training memory, with search on it answers from pages it just read. A tracking method has to hold both of those constant.

What to record for every answer

Four things, and they are the same four things a scoring rubric needs. Was the brand named at all (Presence). How early in the answer it appeared, first sentence or item nine of ten (Prominence). How it was described: recommended or merely listed, accurate or stale, in the right category or the wrong one (Context). And whether the answer cited a link to your own domain (Citation Link), because a mention without a link sends the buyer to whoever was linked. A fifth, optional but useful: how many other brands were named alongside you (Competitive Presence), since a one-in-five mention is a different outcome from being the only name.

Build a fixed question panel first

Do not track your brand name. Nobody types your company name into ChatGPT unless they already know you. Track the questions a buyer asks before they know any names. Write ten of them across four shapes: category entry ("best tools for tracking brand mentions in AI"), problem first ("how do I find out if ChatGPT recommends my company"), comparison ("Pondral vs Profound"), and brand adjacent ("is an AI visibility tool worth it for a small agency"). Write them in the buyer's words, not yours. Then freeze the list. Every change to the questions resets your baseline, so treat it like a survey instrument.

The manual method

Open ChatGPT with web search turned on. Ask question one. Paste the whole answer into a spreadsheet row with the date and time. Fill in the four columns above. Repeat for all ten questions. Repeat the whole set on the same weekday every week. After four weeks you have a trend for each question, which is the first point at which a change means anything. Expect this to take about an hour a week for one brand on one engine.

Where the manual method breaks

It breaks at scale and at rigor. Ten questions on one engine is an hour; ten questions on five engines is five hours; three brands is fifteen. It also breaks on consistency: a human grading "how well was it described" drifts from week to week, and a human copying answers by hand cannot force web search on every call or discard an answer that skipped it. And there is no evidence trail a client or a boss can audit.

How Pondral does it

Pondral runs your fixed panel on ChatGPT, Claude, Gemini, Perplexity, and Grok every week, forces web search on every call (an answer that skipped search is discarded and re-run), stores the full answer with the model version and timestamp, and grades each answer on the five factors above: Presence 20%, Prominence 25%, Context 20%, Citation Link 20%, Competitive Presence 15%. The Gaps page lists the questions you are missing from, ranked by impact, and each one can become a writing brief. Free and SMB brands carry 10 questions, Growth 50, Agency 100. The method is published at /methodology and every change to it is logged.

Reading the results

A few points of movement week to week is noise. A drop of 15 or more on a named question is signal; that is the threshold Pondral's change alerts use on Growth and Agency plans. The number to watch first is not the overall score but the count of questions where you are absent, because those are the ones content can fix fastest. Our own brand, on September 21, 2026, scored 79 on ChatGPT and 0 on Perplexity on the same panel. The overall was 38. The gap list, not the headline, is what told us what to do. How the five engines differ is the subject of the companion answer on monitoring across engines; how the tools compare is at best AI visibility platform for tracking brands in ChatGPT.

Start free

The free check at /demo runs 2 engines on a fixed 12-question mini panel in about a minute, with no signup. It is a snapshot, not tracking, but it shows what a grounded answer says about you today. The longer playbook for closing the gaps it finds is the AEO playbook.

Questions people ask next