What you are actually optimizing
AI visibility is not one number. It is whether your brand appears in relevant answers, how early it appears, how it is described, and whether your domain gets cited, all relative to the competitors named in the same answers.
Optimization means moving those signals on the questions that decide purchases in your category. Publishing volume without a target signal is activity, not optimization.
The loop
Step 1
Baseline with repeated samples
Fix a set of buyer questions and measure each one several times per engine. One draw per question is a coin flip wearing a dashboard.
Step 2
Read the gaps, not the score
Find the prompts where competitors are named and you are not, the descriptions that are wrong, and the domains answers cite where you are absent.
Step 3
Change what the gaps point at
An answer-first page for a losing prompt, a corrected description on a cited domain, a comparison page for the head-to-head question. Each change should map to a specific gap.
Step 4
Re-measure on the same basis
Same questions, same engines, same method. If the prompt set changed midway, the trend is measuring your edits to the tracker, not your visibility.
Step 5
Keep or iterate
A change that clears the uncertainty range is progress worth keeping. A change inside the range is a reason to iterate, not to celebrate.
Why one-shot checks stall the work
The same question can produce a different brand list on the next draw. Teams that check once see phantom wins and phantom losses, then optimize against noise until the program loses credibility.
Mention rate with a confidence interval solves the credibility problem: it tells the team how sure the measurement is, which is what makes the next change defensible.
Changes that usually come first
Every program is different, but the first round of gap-driven work tends to include the same few moves:
- Rewrite the highest-intent page to answer its question in the opening lines.
- Add question-form headings that match how buyers phrase the losing prompts.
- Publish a direct comparison page for the head-to-head question competitors currently win.
- Get present and accurately described on the two or three domains answers cite most in your topic.
- Add structured data and an llms.txt so nothing machine-readable blocks the above.