Guide / Free method / October 2026
Ask the questions your buyers ask, not questions about your name. Ask each one several times in ChatGPT, Gemini and Claude, in a fresh chat with memory off, and note whether the assistant searched the web. Then count how often you are named, who is named instead and which sources the answers lean on. One chat proves nothing. This is the method we use, cut down to what you can run yourself in an afternoon.
Assistants do not have one fixed answer. They write a new one each time, they disagree with each other, and they change when their sources change. In our CRM study, a company outgrowing spreadsheets asked which CRM to buy. ChatGPT named Salesforce in 87% of its answers. Gemini named it in 17%.
The situation matters as much as the assistant. Pipedrive came up in 99% of answers about a first CRM for a small business, and in 8% of answers for enterprise sales teams. Even the order of a list shifts from one run to the next, so where you appear in a single answer says almost nothing. Whether you appear at all is what holds up.
A single chat shows you one draw from that spread. It can make you look safer, or more absent, than you are.
Source: Kesterley industry reports, August 2026, answers recorded through the assistants’ developer interfaces. CRM report.
Five to eight buying situations, phrased the way a buyer would type them, without your name. Marketing science calls them category entry points: the moments in which buyers start looking. For a CRM vendor: “Which CRM suits a 20-person sales team moving off spreadsheets?” For a payroll provider: “Who can pay our staff in five countries?” Buyers rarely ask for you by name before they have a shortlist, so these questions show whether you are on it.
“What does [company] do, and who is it for?” and “What are the downsides of [company]?” They show the story the assistant tells about you, which is a different matter from whether it recommends you. Count them separately.
Use a new chat with memory and personalisation switched off. ChatGPT calls this a temporary chat; Gemini and Claude have similar modes. Never mention your company before the question. Your own history with an assistant changes what it tells you, and your buyers do not share that history.
Buyers use all three, and the answers differ more than most teams expect. Use exactly the same wording in each.
Five fresh chats per question and assistant. With fewer, you cannot tell a pattern from luck. Six questions make 90 answers, about two hours including the log.
An answer with links or sources was written with a live web search. One without was written from what the model learned in training. Where the setting exists, run each question once with search off and once with it on. The two move at different speeds: search answers can change within weeks when the sources change, memory answers over months.
One row per answer: the question, the assistant, whether it searched, whether you were named, who else was named, anything it got wrong about you and the sources it cited. Use our answer log template or any spreadsheet.
Count, per question and assistant, how many answers named you. Three out of fifteen is a number you can track. “It mentioned us once” is not.
Then look at who was named instead. Those brands are your real competitors in AI answers, and they are not always the ones on your sales team’s list. Check what the assistant got wrong about you: old prices, discontinued products, the wrong kind of customer. A wrong claim usually traces back to a page somewhere. In answers written with a search, the cited sources show where the assistant forms its view of your category: review sites, comparison articles, your own pages or a competitor’s.
On noise: with five answers per assistant, a difference of one or two answers is usually luck, while a gap like zero against twelve out of fifteen is not. The next section shows the arithmetic.
The method borrows from marketing science. Research at the Ehrenberg-Bass Institute shows that brands grow by being thought of in more of the situations in which buyers start looking for a product. These are called category entry points, and a brand’s standing is measured across many of them, never with a single question. Step 1 applies the same idea: the buying situations are the sample, so the result describes how your category is bought rather than a list of prompts someone happened to like.
Each answer is treated as one draw, the way a survey treats one respondent, which is why every question is repeated. Small samples carry wide margins. Three named answers out of fifteen mean a true rate somewhere between 7% and 45% at 95% confidence. Zero out of fifteen means between 0% and 20%; twelve out of fifteen, between 55% and 93%. Those two ranges do not overlap, so that gap is real. Three against five out of fifteen is not.
Answers from memory and answers written with a live search measure different things. Memory reflects what a model absorbed in training, the machine counterpart of what buyers remember. Search reflects what it can find today. Averaging the two would mix measures that move at different speeds, so they are kept apart, in this check as in our audits.
Sources: Romaniuk, J. and Sharp, B. (2004), Conceptualizing and measuring brand salience, Marketing Theory 4(4). Sharp, B. (2010), How Brands Grow, Oxford University Press. Romaniuk, J. and Sharp, B. (2016), How Brands Grow Part 2, Oxford University Press. Romaniuk, J. (2023), Better Brand Health, Oxford University Press. Intervals by the Wilson score method: Wilson, E. B. (1927), Journal of the American Statistical Association 22. Our full rules are in the methodology.
Short on time? Take your two most important buying questions and ask each three times in each assistant, in fresh chats with memory off. That makes 18 answers. It will not give you a rate you can trust, but it tells you quickly whether you are in the conversation at all, and who is.
It cannot tell you whether AI answers move your sales: it shows whether you are in the conversation, not what the conversation is worth. It is a sample, and your buyers may see something different, because assistants personalise answers and change between versions. It also covers only half of the picture. The other half is whether an AI agent can read your website at all, which our free website check tests in about a minute.
A manual check is the right start. When the result matters for a budget or a board discussion, the sample needs to be bigger and the method stricter. Our audit records about 1,000 answers across 8 to 16 buying situations of your category, keeps memory and search answers apart, compares you with the competitors you name, checks every claim about you against your own website and traces the sources behind the answers. It costs €399 and arrives in five business days.
You can, and it is worth doing, but it answers a different question. It shows the story the assistant tells about you. It does not show whether the assistant recommends you when a buyer describes a problem without naming anyone.
Memory and personalisation, location, the model version and plain chance. That is why the method uses fresh chats and repeats every question.
Monthly, with exactly the same questions, so the numbers stay comparable. Check again after a launch, a rebrand or major press coverage.
Not by itself. Being named when buyers ask is a precondition for being chosen, not proof of it. We call the measure AI availability and treat it as a leading indicator.
Kesterley, October 2026. The full measurement rules are in our methodology.