If I started AI search today, I would do these 5 things first

1. Collect 25 real buyer questions before writing anything

I want questions with constraints in them: budget, location, use case, risk, compatibility, timing, who the buyer is. Those are the prompts that force the model to choose between brands.

If I have access to real conversations, I use them. If I do not, I start with sales calls, support logs, Reddit, reviews and the language customers already use. The point is to begin with demand rather than with a keyword list.

2. Run the same decision prompts repeatedly and record who gets picked

One AI answer is noisy. I would run a compact set repeatedly across the models that matter to the business, then separate three things: brand mention, recommendation position and cited source.

I am looking for prompts where a competitor is consistently chosen and where the reason is visible. That is a much better work queue than a generic “AI visibility score.”

3. Inspect the fan-outs and the sources behind the answer

This is where the plan gets specific. If the model searches for a sub-question such as “best X for Y,” “X reviews,” “X pricing,” or a local inventory fact, I want to know which page wins that sub-search and whether our brand has equivalent evidence.

Fan-outs are useful because they turn an opaque recommendation into a set of smaller research tasks. Some belong on the site. Others clearly belong on a third-party source.

4. Make the important facts boringly easy to retrieve

Before publishing more prose, I would clean up the facts the model needs to compare the brand: product attributes, pricing logic, locations, availability, compatibility, proof, policies and who the company actually is.

That means useful visible copy first, then structured data where it fits, stable URLs, consistent listings and machine-readable feeds for data that changes often. We now do this for Aiso's own public dataset summary.

5. Only then create content or distribution for a specific missing answer

If the assistant cannot answer “is this suitable for X?” from our site, that is a content gap. If it can answer from our site but trusts a competitor because every independent source mentions them, that is a distribution gap. Those are different jobs.

I would write or pitch exactly what closes the gap and rerun the original prompt set after the evidence has had time to propagate.

What I would not do first

• Track hundreds of wording variants before I understand the main buyer decisions.
• Treat llms.txt as a ranking trick.
• Rewrite every page into stilted “AI optimized” prose.
• Chase a brand-visibility percentage without saving the underlying answers and citations.
• Publish content when the actual problem is that the assistant trusts a third party more than us.

The first week should leave you with a small map of losses you can explain. That is enough to start.