Ask an AI chatbot for the best running shoe, the best customer support tool, or the best place to buy a used car, and you will get an answer. Ask it the exact same question five minutes later, and you may get a different answer. Ask it a hundred times, and you will get a spread of answers that looks less like a recommendation and more like a roll of the dice.
That is the problem a newly published protocol called the Dice Roll Method sets out to solve. Published on October 15, 2026, the protocol is a standardized way to audit how large language models recommend brands. Instead of asking once and treating the answer as fact, it treats every AI answer as a single sample in a much bigger picture. The goal is to turn a fuzzy, random-feeling process into something a business can actually measure, track, and act on.
It sounds technical. But it points at something much bigger. AI is quietly becoming the world's most powerful recommendation engine, and right now almost nobody knows how to check what it is saying.
Traditional search gave every business a fair, boring, knowable deal. You ranked for a keyword, you appeared on page one, you got clicks. The rules were imperfect, but they were stable. You could check your position once a week and have a reasonable idea of where you stood.
Generative AI broke that. A language model does not return a ranked list. It writes a sentence. And that sentence is assembled fresh every time, influenced by temperature settings, context, phrasing, and a dozen internal factors that shift from one query to the next. Two people typing the same request can get two different brand lineups. The same person typing it twice can get two different lineups.
This is the dice roll the protocol is named after. Every query is a throw. One throw tells you almost nothing. But a thousand throws tell you the shape of the odds.
The Dice Roll Method is, at its heart, a repeated-query auditing protocol. Rather than evaluating a model's brand recommendations with a single prompt, it repeats the same query many times and studies the pattern that emerges. That shift, from one answer to a distribution of answers, is the whole idea.
Under this approach, the questions an auditor asks change:
By standardizing these measurements, the protocol gives businesses something they have never had in the generative era: a repeatable scoreboard. Not "the AI said we're good today," but "the AI mentions us in roughly this share of runs, with this much consistency, in this position."
Plenty of companies already test AI outputs. What has been missing is agreement on how to test. Without a standard, every vendor, agency, and internal team invents its own method, picks its own number of repetitions, and reports results that cannot be compared to anyone else's.
A shared protocol changes the conversation. It makes results comparable across time, across brands, and across different AI systems. It also makes the results harder to fake. When everyone measures the same way, a vendor cannot quietly run three queries, find one favorable answer, screenshot it, and call it a win.
Consumers increasingly start product research with a chat prompt instead of a search box. That makes the AI's answer the first, and sometimes only, impression a brand gets. If a model does not mention you, you are not on the shelf. If it mentions you inconsistently, you are on a shelf that keeps vanishing.
The Dice Roll Method formalizes an uncomfortable truth: in a generative world, brand visibility is a probability, not a position. Businesses will have to manage it the way they manage any other probabilistic system, with sampling, monitoring, and statistics.
Expect an entire industry to grow around this. AI visibility auditing, brand mention tracking, and generative performance reporting are all natural extensions. The moment a method becomes standardized, it becomes teachable, certifiable, and contractable. That is when it stops being a clever trick and starts being infrastructure.
It also reshapes job roles. Marketing teams will need people who understand sampling and variance, not just copywriting. Data teams will be asked questions about prompt sets and run counts. The line between SEO and data science will keep blurring.
When you measure a system repeatedly, you can start asking harder questions. Does a model consistently favor the biggest brands? Does it systematically overlook smaller or newer players? Do recommendations shift depending on how a question is phrased, or on who is asking?
Repeated-query auditing makes those patterns visible. Visibility is the first step toward accountability. You cannot debate fairness in AI recommendations until you can measure what the recommendations actually are.
For anyone whose revenue depends on being recommended, the message is blunt: stop spot-checking and start sampling.
Pick the handful of prompts that matter most to your category. Run each one many times. Record which brands appear and where. That baseline becomes your reference point. Without it, you have no way of knowing whether a change you made helped, hurt, or did nothing.
Being mentioned once is close to meaningless. What matters is your share of mentions relative to competitors across many runs. This is the generative equivalent of market share, and it is the number executives will eventually ask for.
A brand that appears in half of all runs is not the same as a brand that appears in all runs half the time. Low consistency means unpredictable exposure, which makes revenue forecasting harder and marketing investments riskier. Consistency is often worth more than raw frequency.
As the method gains traction, AI visibility targets will start appearing in agency agreements and vendor scorecards. "Maintain top-three mention share across our core prompt set" is a very different clause than "improve search rankings," and businesses should prepare for that shift before it arrives.
The stakes go beyond marketing departments. If AI assistants become the default way people choose doctors, lawyers, insurance, or schools, then the statistical shape of AI recommendations becomes a public interest question.
Repeated-query auditing gives regulators, journalists, and researchers a practical tool. Rather than arguing about whether a model is biased in the abstract, they can run a query many times and show what the distribution looks like. That is far more persuasive than a single screenshot, in either direction.
It also cuts against manipulation. If a brand's visibility is measured as a distribution rather than a single answer, it becomes harder to game with one-off tricks. You cannot fake a thousand runs.
The Dice Roll Method is not glamorous. It is a measurement protocol, and measurement protocols rarely make headlines. But they quietly decide who wins. Search engine optimization became a multibillion-dollar discipline because someone agreed on how to measure rankings. Generative AI visibility is heading for the same fate.
The organizations that get ahead will be the ones that accept a strange new reality early: the AI does not have a single opinion about your brand. It has a distribution. And the only way to know it is to roll the dice, over and over, and count.