A one-off check of AI answers describes one afternoon. A fixed question set asked on a set date each month shows whether you are gaining or losing ground.
Write a fixed set of 30 to 50 buyer questions without your brand name, ask them in the same AI assistants on the same date each month, three times each, and count the answers that name or cite you. Torke calls that share AI Retrievability, and keeping the questions fixed is what makes one month comparable with the next.
Most attempts to track AI search visibility start with someone asking ChatGPT a few questions, taking screenshots, and asking different questions a month later, which produces two readings that can't be compared because the questions changed between them. Answers from AI assistants also vary from one run to the next, so a company that appeared on Tuesday may be missing on Wednesday without anything having changed. Tracking visibility over time needs a measurement that holds still while the market moves around it.
We call this measure AI Retrievability: the share of a fixed set of buyer questions, asked across named AI assistants on a set date, in which your company is named or cited. Other AI visibility vendors use the term too, so what we add is the method, a question set written in the buying group's words and kept stable, so that the score compares from one month to the next. It turns Machine Recall, the one-off check of whether assistants name you unprompted, into a series you can report on.
The question set is the instrument, so it is written once and changed rarely. Take 30 to 50 questions from real sources: your sales call notes, the first emails of recent buyers, the questions your support team answers and the problems listed on your industry page. Spread them across the members of the buying group, since an operations lead, an engineer and a procurement manager ask different things, and across the stages of a purchase, from early questions about the problem to late questions comparing named suppliers. Keep your brand name out of every question, because branded questions belong to Answer Fidelity.
Run the full set on the same date every month, in the same assistants, from fresh sessions with no history, three times per question. For each answer record two things, whether your company is named and whether one of your pages is cited as a source, because a citation sends a buyer to you while a mention only tells them you exist. Record the competitors named as well, since your score means little without the share your competitors hold on the same questions.
Take 40 questions, four assistants and three runs, which is 480 answers a month. If you are named in 72 of them, your AI Retrievability is 15 percent, and if 24 of those answers also cite one of your pages, your citation rate is 5 percent. If three months later the same 480 answers show 108 mentions, or 22.5 percent, the questions didn't change, so the movement belongs to the market and to your work rather than to the instrument.
The method has limits we state in every report. A model update can move every company's score in the same week, so a change shared by you and your competitors is a change in the assistant rather than in your position. Random variation matters too: with 480 answers and a score around 15 percent, chance alone can move the result by about three percentage points either way, and more than that because repeated runs of one question are not independent, so smaller moves shouldn't be reported as progress. The question set also ages, so we review it once a year and run the old and new sets side by side for one month when it changes, which keeps the series readable.
For a marketing leader, a stable question set turns AI visibility from a folder of screenshots into a number that can sit in the monthly report next to accepted leads, with its uncertainty attached. It also shows where to work, because the questions on which competitors are named and you are not form the content plan, ranked by the buying group member who asks them. In a category where search volume reads as zero, this is often the earliest measurable sign that the vocabulary you are building is reaching buyers.
A measurement becomes useful only when it is repeated the same way, which holds for AI answers as much as for any survey or panel. Pull 40 questions from your last quarter of sales calls this week, strip out your brand name, and run a first baseline across four assistants, so that next month's run has something to be compared against.
Each answer stands on its own, so you can forward one to a colleague, and an assistant asked the same question can quote it as it is.
We use 30 to 50 questions across four assistants, with three runs each. With fewer questions, a change from one month to the next can't be told apart from the normal variation between runs.
A mention names your company in the answer. A citation links one of your pages as a source, which sends the buyer to your site. We track both, because mentions without citations mean other pages are describing you.
Monthly, on the same date, with the same questions and assistants. Weekly runs mostly measure variation between runs, while quarterly runs can miss a model update that shifts every company's score at once.
Tell us what your team sells, who buys it and where deals stop moving. You'll get a written reply saying which way of working fits you, or whether a Diagnostic makes sense, and if neither does, the reason.