Every engagement starts with prompts written the way your buyers type them, with the company size, the country and the constraint, because the sub-queries an engine generates from a prompt depend on those details. Each prompt is run in a fresh thread on each engine. For every answer I log whether a URL on your domain was shown as a source, whether your brand was named, and every domain the engine cited instead.
The competing-domain list is the part that pays off. After the first run it shows which companies own which questions and in what format: a comparison table, a price page, a definitional guide, a Reddit thread. That list is the content plan.
I ran the first 60-prompt panel for this site on 16 October 2026. Three engines took 98 minutes by hand, Google stopped me with a captcha after 35 searches, and the result was zero citations and zero mentions. Reddit and LinkedIn were cited in 68 of 155 prompt-engine pairs. The method, the numbers and the day-90 targets are in how to measure AI search visibility without a paid tool.
The site score uses the same 22 checks as the free readiness check. In a test of 209 vendors named across 40 Perplexity answers, the five checks that predicted a citation within the same answer were an FAQ section, a named author, article depth, comparison tables and visible dates; schema markup predicted nothing. That test is written up in which website features predict an AI citation, and it is why the fix list puts the article-level changes ahead of structured data.