Choosing a baseline that will not embarrass you later
The first run of an AI visibility programme almost always looks bad, because nobody has optimised for a surface they were not measuring. Set expectations before the first report, and frame the initial numbers explicitly as a baseline rather than a performance verdict. The metric that matters is the delta from that baseline against named competitors, not the absolute value.
Choose competitors deliberately. Three to five direct rivals with comparable size gives a fair comparison; including a category giant you cannot realistically match makes every subsequent report read as failure. Track the giant separately as context if leadership wants it.
Document the methodology alongside the numbers - prompt count, engines covered, run date, and any prompt changes. Six months in, this is what makes the trend defensible when someone asks why a quarter moved.
Segmenting by intent to find the real problem
Aggregate citation share hides the diagnosis. Split it by prompt category and the picture becomes actionable. Strong branded citation with weak category citation means people who already know you find you, but you are absent from discovery - a content coverage problem. Strong category citation with weak comparison citation means you are visible but losing the evaluation - a differentiation and evidence problem. Weak everywhere usually means a technical or crawler-access problem upstream.
Each pattern implies a different quarter of work, which is why the segmentation is worth the small extra effort at logging time. Recording the prompt category as a field when you log a run costs nothing and makes every subsequent analysis possible.
Report the segments separately to stakeholders as well. 'We are cited in sixty percent of comparison prompts and twelve percent of category prompts' communicates a strategy; a single blended percentage communicates nothing anyone can act on.