Ran answer simulation on a cluster of related questions. I get cited for 'best X for small business' but not 'best X for startups' — nearly the same intent. Why is AI citation this inconsistent, and is there anything I can do about the near-misses, or is it just model randomness I have to accept?
Some of it is genuine model variance, but the pattern you describe usually points at something fixable. If you're cited for 'small business' but not 'startups', it's often that your content literally uses the first phrase and its context, and never speaks to the second audience's language or specific concerns.
Models retrieve and cite based on semantic match to the exact framing of the question. 'Small business' and 'startups' feel identical to you but carry different associated language (runway, scaling, funding vs. established, local, budget). If your page only speaks one dialect, you'll get cited for one and miss the sibling query.
What to do: on that page, explicitly address the startup angle — a short section, the right vocabulary, a specific claim aimed at that reader. Then re-run the simulation. You're not fighting randomness, you're filling a genuine coverage gap the tool just exposed.
Added a startup-specific paragraph with the right vocabulary and re-ran it a few days later. Now cited for both variants. So it really was a coverage gap, not dice. That's a genuinely useful workflow.
