Some things about AI citations over two months of measuring

I build IvaBot, a tool for checking visibility in AI search, and I watch citations across my own and client accounts. These are random observations from the last two months, i put in one place. I am posting them to hear what works and what does not for other people doing the same thing, because i'm still learning.

1. Publish a specific number, a result from your own measurement or a direct quote. The model pulls that kind of material into an answer because it cannot assemble it from anywhere else. - this works the best!

2. Leave answers inside discussions on Reddit, YouTube and similar platforms, and write each one so it stands on its own. The model often pulls a single comment rather than the thread around it, and a comment that answers the question completely can become the cited source by itself for long period of time. Threads with recent activity and real discussion underneath come up more often.

3. Measure on the prompts a customer would type, such as the best tool for a given job or so on. The answer then depends on where the brand actually stands, because the wording carries no hint toward any company.

4. Fix one exact brand string before the first measurement and keep it unchanged between runs. The trend line stays comparable only while the same thing is being counted. Example from a client: on a fintech account the mention count moved from 488 to 159,512 between two runs after the string reverted to the short brand name that a large cosmetics company also uses, absorbing around 157,000 mentions belonging to them. Nothing was published and the site was untouched, and on the chart it read as the best month in the account's history. So pay attention whar are you tracking :)

5. Read the tone of every mention alongside the fact of it. A negative source and a missing page call for different work, and the count reads both the same way. Several accounts had solid citation rates while the pages being cited carried negative reviews. And it was really bad for local SEO, untill i found importance of tracking sentiments.

6. When the same prompt loses the brand across two consecutive checks, open the answer and find which source took that place. Two losses in a row mark a real shift, which makes reading the full answer worth the time. From there the work is a page of your own that answers that prompt better, or a mention inside the source that is already cited.

The tone check went into IvaBot this month, straight after the fifth point.

I would be glad to hear what has worked for you and what has not, especially from anyone measuring this across more accounts than I am.