The method
How the Snapshot
is measured
It is detailed enough that you could repeat it yourself. We publish it so you can check our work instead of taking our word for it.
Retrieved → Cited → Trusted, in full
We check your robots.txt rules for the named AI crawlers (GPTBot, ClaudeBot, PerplexityBot and the rest), hosting-level bot blocks (Cloudflare's "Block AI bots" toggle, Vercel's firewall rules, shared-host security plugins), and whether your content exists in raw HTML, because most AI crawlers do not run JavaScript. Pass or fail, and you can verify it yourself with a curl command we show you in the report.
Every question is classified per engine: were you cited as a source, mentioned by name, merely retrieved, or absent, and the same for each competitor. We show raw counts, not just percentages, broken out by buyer-journey stage.
The engines lean on third-party sources: directories, review sites, articles. We aggregate every source the AIs actually cited in your market, rank them by frequency, and check your presence on each. This is where most of the fix list comes from.
How the questions get built
We do not test vanity prompts. We build a picture of your actual buyer, then write the questions they ask at each stage of deciding, and you see every one before we lock the set. The five stages are:
The set is weighted toward Comparison and Selection, which are the questions asked by buyers who are about to choose someone. Those are the answers worth winning.
How we handle variance in AI answers
Ask the same engine the same question an hour apart and you can get different shortlists, and we have measured it. So every question runs five times per engine, and the report shows how stable the result was ("you appeared in 3 of 5 runs") rather than one picked screenshot. If a result is shaky, the report says so.
Every one of those counts carries a 95% range next to it, because 3 of 5 runs on its own does not tell you much: the honest range around it is 23% to 88%. We use a Wilson score interval to work that range out, we raise the run count when a decision depends on a close call, and we treat a week-over-week move as inconclusive unless the two ranges separate. The full write-up is here, including the code, the worked numbers, and the four things the range does not cover.
Which engines, and how
Two engines by API: Claude and Perplexity, every question, five runs each. That produces a log we can reproduce, with every citation URL. One engine by hand: five of the questions are put to ChatGPT in a browser and screenshotted, because its API does not return what a real user is shown. That is the whole engine list. When another engine gets an official API that returns real answers, we add it and say so here. We do not claim coverage we cannot log. That means Google's AI Overviews and AI Mode are not in the Snapshot today. They have no official API that returns real answers. We say this up front because most of your buyers meet AI inside Google, and a Claude result tells you little about what Google shows.
Questions to ask any AI-visibility vendor, including us
What to expect, and when
You get the report in days. If you fix things, the first measurable movement typically shows at weeks 4 to 8, with durable gains over months 3 to 6, because AI engines re-crawl and re-weigh sources slowly. Anyone promising faster is making the same claim as the spam in your inbox.
There are two more limits. We make you findable when someone is already looking. We do not make people look: brand demand comes from referrals, reviews, and reputation, and nothing on our ladder manufactures it. And we do not write copy to game the model. Longer, more persuasive, fluffier text wins in AI answers today, and a vendor selling "AI-optimized copy" will offer it. We will not, because it is the kind of thing the engines are already learning to filter, and it reads as spam to the person on the other end.