Answer: Maybe. Ask ChatGPT, Perplexity, Gemini, and Claude a customer's question and count how often they name your business. That count moves on its own, because their makers keep rebuilding them. One study watched 753 questions. Most of the movement left over came from the apps, not the businesses named (Source: Profound, 2026). Ask the same questions every month.
When somebody types a question like who is the best plumber in town, an AI app writes back a short answer that names a few businesses. Being in that answer, or missing from it, is the thing being counted on this page. The honest answer is that one month's rise, on its own, cannot tell a business owner anything. Two counts a month apart look like a before and an after. They are not, unless everything except the business was held still: the same questions, each one put the same number of times, and the date written beside each count. Miss any of that and the second number is measuring something different from the first, so the gap between them is not news about the business. What follows is the evidence for why, and the short list of what to do instead.
Why can an AI app name a business more often when nothing changed?
Because the app changed, not the business. ChatGPT, Perplexity, Gemini, and Claude are pieces of software. The companies that own them keep rewriting it. So the same question, put two months apart, is answered by software that is not quite the same software as last time. Nothing at the business has to change for that.
In July 2026 a company called Profound, which sells software for tracking how often AI tools name a brand, published a test of its own product. It ran two copies of the same tracking setup side by side for two weeks, on the same 753 questions across the same 7 AI platforms. One copy put each question once a day. The other put it ten times a day. That worked out at roughly 129,000 runs in the once-a-day copy and about 860,000 in the other (Source: Profound, 2026).
The extra asking bought much less than expected. Once a day already landed within about 2 percentage points of the ten-times-a-day reading. Ten runs a day cut the day to day wobble by roughly 40%, and the post is blunt about what is left over: "even at ten runs a day, most of the remaining movement is platform drift, not sampling". Drift is their word for the AI tool changing underneath the measurement. Their own one line summary of it is "You can't measure your way past drift."
The post also lists what actually changes. The companies quietly update the models, add new tools, and re-tune the machinery that writes the answers, and none of that is visible to a business or in its control. The web those models read is changing at the same time. So the movement left in a monthly number, after the ordinary randomness has averaged out, is mostly the tools, not the businesses in the answers.
Read it with its limits. Profound sells this kind of tracking, so this is a company writing about its own product, and the headline that suits it best is the one saying once a day is enough. The half that matters here is the half that does not flatter the product: what is left moving is the tools themselves. The test also spread across hundreds of questions at once, which smooths a lot away. One business watching one short list of questions gets less smoothing, not more.
Do the AI companies really change the software that writes the answers?
Yes. Anthropic, the company that makes the Claude AI app, keeps a public page listing which versions of Claude it has switched off, and the date it switched each one off.
That page says it in one line: "As safer and more capable models launch, Anthropic regularly retires older ones." Then it lists them. Claude Sonnet 4.5 was marked for retirement on 30 September 2026, with the switch-off listed for 30 November 2026, and six older versions were switched off during the first half of 2026 (Source: Anthropic, 2026).
Two limits on reading too much into that one list. It covers the versions other software makers build on, and the Claude app a person types into follows its own release schedule. And Anthropic is one company of four here: the makers of ChatGPT, Perplexity, and Gemini each run their own schedule in their own way. The point is not the exact dates on one page. It is that the thing answering a customer's question in November is not quite the thing that answered it in October, and nobody sends a business a note when it changes.
How much does the answer move from one day to the next?
Only about half the business names stay put from one day to the next. Three researchers put the same questions to several AI apps for almost two months, among them ChatGPT, Perplexity, and Gemini. They compared each day with the day before. Between 45% and 59% of the business names showed up on both days (Source: Schulte, Bleeker, and Kaufmann, 2026).
That work was published in April 2026 by Julius Schulte, Malte Bleeker, and Philipp Kaufmann, in a paper called "Don't Measure Once". The range runs from 45% to 59% because they tracked four separate groups of questions, and the groups did not behave alike. The paper also says the spread around those figures was wide, meaning some days moved much more than others. Its own summary of the problem is that answers "can vary across runs, prompts, and time, making one-off observations unreliable".
Their advice is the useful part for anyone holding two monthly numbers. One reading on one day is the wrong thing to look at, and a business is better described by a range than by a single figure. For a number about one business they write that "Rolling aggregation over two to four weeks is therefore recommended", which in plain words means add several weeks of checks together instead of trusting one.
The limits are real and the paper states them. It covered 8 questions in each of four shopping areas, 32 questions in all, in German, in Switzerland, between 24 January and 20 March 2026. That is a small set of questions in one language and one country. What carries over is the direction, and it is a strong one: the daily list is not steady, so two readings a month apart are two picks from something that moves on its own.
What makes last month's count and this month's count worth comparing?
Nothing about the check changes except the month. The same questions, word for word. Each question put to each app the same number of times. Each app counted on its own. The date written beside every count.
The question list is the part most reports quietly change, and it is the part that matters most. Write down the questions a real customer would type, such as how much a new boiler costs in Leeds, or who can clear a blocked drain tonight, and then never change that list again. A count taken over a different list of questions is a different measurement wearing the same label.
The number of tries is the next part. Putting a question once gives one reply, and one reply settles nothing. In January 2026 SparkToro, the research firm run by Rand Fishkin, published a test in which 600 volunteers put 12 questions to the most popular AI tools in the United States, among them ChatGPT and Claude, 2,961 times in all. The same list of businesses came back twice less than 1 time in 100 (Source: SparkToro, 2026). So each question gets put at least three times, and five is better.
Counting each app on its own matters for the same reason. ChatGPT, Perplexity, Gemini, and Claude do not behave alike, so one blended figure across all four hides which app a business is missing from. A business can climb on one and fall on another in the same month, and a single averaged number shows neither.
What would count as proof that a business improved at getting named by AI?
The strongest evidence anybody has is a rise in the count that holds for several months, on a question list that never changed. That is worth having. It is still not proof that the business caused the rise. A review of 45 studies on getting named by AI found no way of doing it with a proven effect that lasts and works across different AI apps.
That review was published in July 2026 by Olivier Martinez, who read the research on getting picked by AI from 2023 onward. The conclusion is that no technique reviewed shows "a stable, longitudinal, cross-platform causal effect on organic discoverability" (Source: Martinez, 2026). In plain words: nothing in that literature has been shown to work the same way over time and across different AI tools. So when a report says a change caused a rise, it is going further than the published research can.
What is left is still worth having, as long as it is read honestly. A rise on a question list that never changed, counted the same way, holding for several months in a row rather than one, is worth paying attention to. A single month's jump is not. And a move from zero to anything at all is a different kind of news than six to eight, because zero means a business was missing from every reply, which is a real problem with a real job attached.
There is also no good score to aim for. The only honest comparison is the same count next month, on the same questions, which is why the date beside the number is doing as much work as the number.
What should a business owner do with last month's number?
Put it next to this month's number only if the check was run the same way twice. Same questions, word for word, and each question put the same number of times. If either of those changed, the two numbers cannot be compared, and the honest move is to write a question list today and start again from there. A comparison that is broken and gets repeated to people is worse than no comparison at all.
Then keep it boring on purpose. Write the question list down and leave it alone. Put each question to ChatGPT, Perplexity, Gemini, and Claude at least three times, and five is better. Count how many replies name the business, app by app. Write the date beside each count. Wait for several months of that before reading anything into a move.
It is worth the trouble because customers really do ask this way now. In a survey of 1,002 US adults, 45% said they had used AI to find a local business in the past year, up from 6% a year earlier in the same survey (Source: BrightLocal, 2026).
And put two questions to any report that hands over a rise from last month. Ask whether the question list is the same list as last month, word for word. Ask how many times each question was put. If either answer is unclear, the rise is a number with nothing underneath it, and the honest thing to do is write a question list and begin again.
One more thing, because the count on its own does not fix anything. If a business is named in none of the replies, the job is not comparing months, it is getting named in the first place, and that is a different job with different work attached. The related answers listed with this page are the ones that cover it.
Where does that leave a business owner?
Honestly, one month's rise on its own does not prove a business improved. It starts to mean something when the count is taken the same way twice. Same questions, word for word. Each question put to each app the same number of times. Each app counted on its own. The date written beside every count. That is what AI Says Who, the service behind this page, does.
We research the questions customers really use, freeze that list so later months stay comparable, put them to ChatGPT, Perplexity, Gemini, and Claude several times each, and write the date beside every count. No single blended score, no position, and no claim that one month's move proves a business improved. Gemini is asked with no live web search, so it reports what it already knows about a business and never shows a link. The first check is free, and the button below starts it.
Free. No card, no call.