There Is No Rank in ChatGPT, Only a Mention Rate

‘What is our rank in ChatGPT?’ is the wrong question, and it is worth correcting before anyone answers it. A rank in ChatGPT is not a property your brand has. It is a property of one answer, and the next answer to the same question will usually have a different one. The question that can be answered is another: across many answers, how often does the assistant name you at all?

That sounds like a quibble about wording. It is not. A great deal of reporting on AI visibility still arrives as a position, ‘third in ChatGPT for industrial pumps’, and the number is read the way a Google ranking used to be read: as a place you hold until someone takes it from you. This piece takes that claim apart, one assumption at a time, and then sets out what to measure instead and how many runs a reading needs before it means anything.

What ‘third in ChatGPT’ quietly assumes

A sentence like ‘we are third in ChatGPT’ rests on three assumptions, and it needs all of them to be true.

• That there is a list, the same one each time, in which a third place exists.
• That the order of that list is a judgement: being third means the assistant rates you below two others and above the rest.
• That the single answer someone looked at stands for the answers everyone else gets.

Each one can be checked. None of them survives the check.

Rank in ChatGPT: there is no fixed list

A search engine ranking is a stored ordering. The index exists before your query, the ordering is computed against it, and two people asking the same thing a minute apart see broadly the same page. A generated answer has no such object behind it. The model writes the reply word by word, choosing at each step among plausible continuations, and a list of brands is just more words. Why that produces different replies to identical prompts is covered in our piece on LLM answer volatility. What matters here is the consequence for position.

On 28 January 2026, SparkToro published a test of exactly this. About 600 volunteers ran 12 brand recommendation prompts through ChatGPT, Claude and Google’s AI Overviews or AI Mode, 2,961 runs in total. The authors put the chance of ChatGPT or Google’s AI returning the same list of brands in any two of 100 responses at less than 1 in 100. For the same list in the same order, they estimated something closer to 1 in 1,000 runs. Claude repeated itself slightly more often, not enough to change the picture.

So the first assumption fails on its own terms. If the list is different nearly every time, ‘third’ describes one draw from a very large set of possible lists. It is a fact about a reply, not about you.

The order is not a verdict

Suppose you set the instability aside and look only at the answer in front of you. Does third place at least tell you that the assistant prefers two brands over yours?

Nothing in the answer says so. An assistant listing suppliers is not sorting a table by a score. It might open with the brand it has seen described most often, group the rest by type, or put a local option last because the question mentioned a country. The order is produced by the same process as the sentences around it, and it is subject to the same sampling. In the SparkToro data, order was the least stable thing measured: the membership of the list moved, and the sequence moved far more.

There is a second reason position misleads. Being named first in a passing list and being named third with a paragraph explaining why you suit the buyer’s case are not ranked the way the numbers suggest. A position collapses that difference into a digit and keeps the wrong half.

One answer does not stand for the others

The third assumption is the one that does the most damage in practice, because it is how a screenshot ends up in a board deck.

On 14 July 2026, Dmitrij Żatuchin posted a preprint on arXiv (2607.13304) that asks where the noise in brand answers comes from. It covers 12,933 responses about 20 Central and Eastern European brands, in 8 query languages, from three models: GPT-5.2, Gemini 3 Flash and Perplexity. What it measures is the tone of each answer about a brand, not whether the brand is named, so it is not a direct test of mention counts. Its structure of noise is still the relevant part.

Two findings matter for anyone reading a position. First, on a scale where 1 means perfectly reliable, a single answer scored about 0.01 at ranking brands against one another. The full design, crossing prompts, languages and models, reached about 0.36. One answer, in other words, carries almost no information about how brands compare. Second, the language of the query was the largest systematic source of variation, larger than the identity of the brand itself. The same brand asked about in two languages can look like two different brands.

A position read off one answer, in one language, from one assistant, is therefore the least reliable number in the whole exercise, presented as the most precise.

What to measure instead: mention rate

Mention rate is the share of answers, out of a known number of runs, in which the assistant names your brand. Named in 7 answers out of 10 is a mention rate. Third place is not.

The difference is that a rate is built to absorb variation rather than be fooled by it. Each run is a sample. If the assistant names you on some runs and skips you on others, the rate records exactly that, and it stays comparable from one measurement to the next as long as the prompts, the language and the assistant are held still. SparkToro reached the same conclusion from its own data: how often a brand appears, counted over many runs, was the signal worth keeping, and position was not.

A mention rate should always travel with three things: the number of runs behind it, the language of the prompts, and the assistant. A rate with no run count attached is a position in disguise.

How many runs a mention rate needs

This is where most readings fall short, and the arithmetic is simple enough to check with any binomial calculator.

Run a prompt 10 times and find your brand in 5 answers. The honest reading of that result, at the confidence level statisticians normally use, is that your true rate sits somewhere between roughly 2 in 10 and 8 in 10. Ten runs cannot distinguish a brand the assistant almost always names from one it rarely does. Moving from 4 out of 10 last month to 6 out of 10 this month is not a change. It is the same uncertainty seen twice.

To pin a rate near the middle down to about 10 points either way, you need around 100 independent answers. To get to about 5 points either way, around 385. Those numbers are why one person checking on a Tuesday cannot produce a trend, and why measuring more often is not the same as measuring better.

The Żatuchin preprint adds a useful twist about where those runs should come from. The common practice is to repeat each prompt about five times, and past the fifth repeat each extra run barely reduced the error. Per query spent, adding languages and models reduced it far more. The practical translation:

• A hundred answers from one prompt repeated a hundred times measure that prompt very precisely and your market barely at all.
• A hundred answers from twenty prompts run five times each, across the assistants and languages your buyers use, measure something closer to how the market sees you.
• Pooled rates hide detail, so keep the per prompt and per language figures underneath the headline number.

Reading a report that hands you a position

None of this means a report that shows a rank in ChatGPT is dishonest. It means the number needs context it rarely ships with. Four questions sort it out:

• How many runs is this position based on? If the answer is one, it is an anecdote.
• Is it an average position across runs, and what happens in the runs where you were not named at all? An average rank that silently drops your absences flatters you.
• Which language and which assistant? A position in English on ChatGPT says nothing about German on Gemini.
• Would the same reading, repeated tomorrow with the same prompt, land on the same number? If nobody has tried, nobody knows.

This is also how we built our own reporting. PSentry puts the prompts of your market to ChatGPT, Claude, Gemini and Perplexity, in each language you sell in, on two scans a month, and builds a visibility score per language from those answers. That does not make any single scan a final verdict: it is a sample too, and the trend across scans is what carries the information. We do not change what the assistants say. We count it.

The short version for the next meeting where someone announces a rank: ask how many answers it came from. If the reply is a screenshot, the position is real for that screenshot and for nothing else.

What buyers and marketers ask next

Can I find out where my brand ranks in ChatGPT?

Not in any lasting sense. ChatGPT writes each answer afresh, so the brands and their order change from one reply to the next. What can be measured is how often it names you across a known number of runs.

What is a mention rate in AI search?

A mention rate is the share of answers, out of a stated number of runs, in which an AI assistant names your brand. It should always be reported with the run count, the language and the assistant.

How many runs does a reliable AI visibility reading need?

Ten runs cannot tell a brand that is almost always named from one that rarely is. Pinning a rate near the middle down to about 10 points either way takes around 100 independent answers, and those answers say more about your market when they come from many prompts, languages and assistants than from one prompt repeated.

Does our visibility in English on ChatGPT show how German buyers see us?

No. The language of the query can shift the answers more than the brand itself, so each language you sell in needs its own measurement.