Answer: At least seven times on each AI tool, every day. Ask fewer times and the answer still jumps around from run to run. That floor comes from a study of repeat asking (Source: Schulte, Bleeker, and Kaufmann, 2026). Two answers to the same question named the same brands less than 1 time in 100 (Source: SparkToro, 2026).
The number is higher than almost anybody asks. Most tools in this field ask each question once and print the result as a score. The rest of this page shows what the research found, and what a small number of asks can and cannot prove.
Why does one answer tell you almost nothing?
Because the answer changes each time. About 600 people ran 12 questions through the most used AI apps 2,961 times. Two answers named the same set of brands less than 1 time in 100. The same names in the same order came up closer to 1 time in 1,000 (Source: SparkToro, 2026).
A local audit found the same thing another way. It started from a complete list of every cafe, restaurant, and bar in two markets, 4,776 of them, then read 2,208 answers from ChatGPT, Claude, Gemini, and Perplexity. The four systems agreed on their top twenty only 33% to 54% of the time (Source: Pitenin, 2026).
So one answer is a single draw from a shuffled deck. It is not a reading of where a business stands.
How many asks does the research actually ask for?
Seven, as a floor. One team ran the same questions over and over on four AI search tools. It then worked out how many runs it takes before the hit rate stops swinging. Seven runs per question per day made the error small enough to use. Eight if the websites behind the answers matter too (Source: Schulte, Bleeker, and Kaufmann, 2026).
The author of the 2,961 run test sets the bar higher still. To really know what an AI recommends, he says to ask over and over, usually at least 60 to 100 times, then average the answers (Source: SparkToro, 2026).
The two numbers answer different questions. Seven asks pin down how often a name shows up. Seeing the very same list twice is a much harder test, and that is where 1 in 1,000 comes from.
What does a zero out of three result really prove?
Less than it looks. Say an AI really does name a business 1 time in 3. Ask three times and you miss it about 3 times in 10. Zero out of three is not zero.
The same sum keeps going. Ask five times and the miss falls to about 13 times in 100. Ask seven times and it is about 6 times in 100. More asks shrink the odds of a false zero. They never erase it.
That is plain arithmetic, not a study, and anybody can check it on a calculator. It is also the reason an honest result is a range rather than one number.
Do the answers settle down if you wait?
No. One audit asked again two weeks later. The answers differed about as much as two tries on the same day. The churn is the AI picking at random, not the market moving (Source: Pitenin, 2026).
Day to day, the drift is wide. Across four AI search tools, the brands named one day overlapped the next day's list only 45% to 59% of the time. The websites behind those answers overlapped only 34% to 42%. The same team suggests adding up a rolling two to four weeks before trusting a per business number (Source: Schulte, Bleeker, and Kaufmann, 2026).
Is it better to ask one question many times, or many questions a few times each?
Both matter, and most checks do neither. The same team warns that one or two questions only show how those questions behave. Its advice is a long and varied question list on top of the repeat asks (Source: Schulte, Bleeker, and Kaufmann, 2026).
The wording changes the answer as much as the repeating does. One study ran six AI models across five kinds of product. Plain questions left well known brands out. Questions that spelled out a need brought back a different set, and brands missing from the plain answers could still be reached (Source: Malthouse, Lee, Yang, Pal, and Feng, 2026).
So spread the asks across many real buyer questions as well as repeating each one. A deep count on two questions pins down those two and leaves the rest dark.
What can no amount of asking tell you?
Three things, and more asking does not touch any of them.
Nobody can tell you why one name came back instead of another. These systems do not report the reason. A confident story about the cause is still a story.
Nobody can turn a count into a promise. A review of 45 studies found no method with a steady, lasting effect across AI tools on being found at all (Source: Martinez, 2026).
And nobody can hand you one true number. The honest output is how often a business was named, over a stated question list, with the date and the ask count printed beside it.
Where does that leave an owner who wants a straight answer?
With a dull rule that works: ask many times, ask on every AI, and write down the count and the date. This is the thing we do. We ask ChatGPT, Perplexity, Gemini, and Claude the questions your customers really ask, many times each, and we publish the question list, the number of asks, and the date next to every result. The first check is free.
Free. No card, no call.