Why we ask every question six times.

The same question, a different answer

An AI assistant writes its answer one word at a time, choosing each word from a range of likely options. The companies that build these models describe the setting that controls this choice, usually called temperature. Anthropic explains that higher temperatures produce more varied outputs, including variation in the answers themselves, and that even at the lowest setting results are not fully deterministic. Google describes temperature in the Gemini documentation as the degree of randomness in how the next word is selected.

Most people use these assistants through their own apps, where the settings are chosen by the provider and not shown to the user. What a business owner sees is the effect: ask for a recommended accountant in a town on Monday and on Tuesday, and the list of names can change. Some businesses appear every time. Others appear once and then vanish.

Answers can also change because the assistant searched the web. When an assistant looks something up before it replies, the pages it finds may differ from one run to the next, and so may the businesses it names.

Why a single screenshot misleads

A screenshot of one answer is a sample of one. If your business is named in it, you cannot tell whether that happens every time or only once in a while. If you are missing, you cannot tell whether you are never mentioned or were simply unlucky on that run.

Both mistakes are expensive. A business that sees itself named once may decide there is nothing to fix. A business that sees itself missing once may spend money on changes it did not need. We would rather give you a measurement than an anecdote.

What six runs show

Asking the same question six times turns a yes-or-no into a count. Named in all six answers means the assistant treats you as a steady part of its answer. Named in two or three means you are in contention but not secure. Named in none means the assistant does not connect you with that question at all, at least for now.

The same counting works for sources. If an assistant cites the same directory or review site in most runs, that source matters for your question. If a source appears once, it is weaker evidence of where the assistant looks.

Six is a practical choice, not a statistical law. It is enough to separate a steady pattern from a one-off, and few enough that we can run every question by hand on three assistants. A free check with three questions is 54 answers, each one read and recorded.

How we run the questions

Each run starts a fresh conversation, so an earlier answer cannot influence the next one. We use the exact wording agreed with you every time, because a small change in phrasing is a different question. We keep the location of the question fixed when it mentions a town, and we note the date of every run.

For each answer we record which businesses were named, in what order, and which sources were cited or linked. Where an assistant shows no sources, we record that too. The result is a table you can read in a minute: who was named, how often, and on what evidence.

What repeated runs do not show

A check is a snapshot. Assistants are updated, their search results change and new businesses open. That is why the paid audit repeats the same questions after 30 days rather than treating one measurement as permanent.

Repeated runs also cannot show what every customer sees. People use different apps, accounts and settings, and some assistants adapt to the person asking. We measure under consistent conditions so that changes between two checks mean something, and we say plainly that the numbers describe our runs, not every possible conversation.

Finally, a count is not a promise. If your business moves from one mention in six to five in six, that is a real change in what we observed. It does not mean an assistant will always recommend you, and we will never tell you otherwise.