How to Track ChatGPT Mentions Without a Paid Tool

The short version

A manual process can produce a defensible baseline if you fix the prompt set, run each prompt at least five times, and log the full answer text rather than a yes or no.

Doing those three things is the entire difference between a measurement and an anecdote, and none of them requires a tool.

Last updated: 30 August 2026. Rechecked quarterly.

Most manual tracking fails for the same reason: somebody asks ChatGPT about their category a few times, writes down whether they appeared, and treats the result as data. That produces a number with no interval and no way to compare it to next month’s.

Why does one check tell you nothing?

Because the same question does not reliably produce the same answer. OpenAI’s advanced usage guide states that “Chat Completions are non-deterministic by default (which means model outputs may differ from request to request),” and even with a fixed seed it offers only mostly deterministic output.

So a single run is one draw. If you appear, you have proved the engine can name you. If you do not, you have proved nothing at all, because the next run might.

One run supports an existence claim. Only repeated runs support a percentage, and a percentage is what everyone wants to report.

What does the manual method actually require?

Five things, and the discipline costs more than the effort.

A manual baseline that survives scrutiny

Fix the prompt set and write it downBetween 10 and 20 unbranded questions, phrased the way a buyer would ask them. Save the file. Do not edit it mid-period, because an edited set makes every comparison meaningless.

Use a clean session every timeLog out, or use a fresh private window. A session that remembers your earlier questions is answering a conversation rather than a query.

Run each prompt five timesFive is the practical minimum that produces a fraction rather than a coin flip. Ten is better if the prompts matter.

Paste the whole answer, not a verdictStore the text. A yes or no cannot be recounted later under a different rule, and you will change the rule.

Record the date and the model versionThe model answering next quarter may not be the one answering today. Without the version, a change is unattributable.

Step four is the one people skip and the one that pays. Stored answers let you re-code the whole baseline when you decide that being listed as a runner-up should not have counted as a mention.

Ten prompts at five runs is fifty answers. On one engine that is an afternoon, and it is a genuine measurement rather than a look.

What do you record for each answer?

Six fields, and a spreadsheet is entirely adequate.

Field	Why it earns its column
Prompt text, verbatim	A reworded prompt is a different prompt. Punctuation included.
Run number and date	Lets you compute variation within a day as well as across weeks.
Full answer text	The only primary evidence you will have. Everything else is derived from it.
Named or not	Your mention count. Simple string matching on your brand name.
Named as the recommendation, or as context	The judgement call. Write the rule down before you start applying it.
Competitors named	Turns your sheet into a share-of-voice denominator instead of just a presence count.

The fifth field is where manual tracking earns its keep over a cheap tool. You are reading the sentence, so you can tell the difference between being recommended and being the thing a competitor is described as better than. Most tools cannot.

What can you get for free that is not manual?

One thing, and it is the strongest data in this whole exercise. Microsoft’s AI Performance report in Bing Webmaster Tools shows “the total number of citations that are displayed as sources in AI-generated answers,” page-level citation counts, and “the key phrases the AI used when retrieving content that was referenced in AI-generated answers.”

That is a census rather than a sample, it costs nothing, and the grounding queries are the only place you will see the actual words that pulled your pages into an answer. It covers Microsoft surfaces only, which is a real limit and does not make it less worth having.

Verify the site first. It takes about ten minutes and a surprising number of teams have Search Console configured and nothing on the Microsoft side.

Where does the manual method break down?

Three places, and all of them are about scale rather than rigour.

Engines. Doing this properly on one engine is an afternoon. Doing it on four is most of a week, every period, forever, and the tedium is what actually kills it rather than the hours.

Frequency. A baseline you refresh quarterly cannot detect a change that happened six weeks ago and reverted. Anything you want to catch inside a month needs checking inside a month.

Google’s two AI surfaces. Its documentation on AI features notes that “AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary,” so they are two separate manual jobs, and neither can be automated without browser work, which is the boundary drawn in build your own AI visibility tracker.

When is paying actually cheaper?

Count your own hours honestly. Ten prompts, five runs, four engines is 200 answers to collect and code, and that is a day of somebody’s month, every month, indefinitely.

Against that, the cheapest published entry in the category is Otterly at $29 a month for 15 prompts on four core engines, read on 30 August 2026. Whether that is better value than a day of your own time is arithmetic you can do faster than we can.

The honest recommendation is to do it manually once, properly, before buying anything. A single rigorous baseline tells you whether the number moves enough to be worth watching, and it makes you a much harder customer to sell to, because you will know to ask how many times each prompt gets queried. The wider field is in best AI visibility tracking tools, and where free options stop is set out in cheap AI visibility tracking tool.