Ask an assistant the same question twice and the list of brands can change with nothing on the web having changed. The methodology page puts it this way: responses "can differ by time, wording, geography and model". So a single number from a single run is a photograph of a moving thing.
What a change needs to look like
- One prompt moving on one assistant means little. That is the noise.
- The same direction across several prompts, on more than one assistant, over three or more checks, means something.
- Control prompts settle the question. Keep two or three prompts you do nothing for. If your targeted prompts improve and the controls do not, your work did it. If both move together, the assistants changed.
How many prompts
Ten to twenty is enough for a small company to see a trend, if they are the right ten: category questions ("best accounting software for a ten person agency"), comparison questions, and problem questions where no product is named. Add two or three branded prompts to see how answers describe you, but keep them out of the headline number. Our guide on finding the questions buyers ask covers how to choose them.
Count the right things
Count mentions and citations separately; an answer can name you without linking to you and link to you without naming you. Leave failed answers out of the denominator: in our check on 2026-10-21, Google AI Overviews returned nothing for 3 of 25 questions, and counting those as "not mentioned" would have pulled the score down for no reason. Show the sample size next to every percentage, so 2 of 5 is never mistaken for 40% of the market.
