How we measure this
Written so that someone who wanted to disagree with us would know exactly where to aim. If a number here is wrong we would rather hear it than defend it.
What exactly is being measured?
Whether an AI assistant, asked to recommend a provider in a category and a place, names a given real business. The Q3 2026 release measures unprompted recall, meaning the assistants answered from their own training knowledge with web browsing switched off. It does not measure whether a business is cited when an assistant browses the live web first, which is a separate and usually more forgiving number that we will report separately rather than blending in.
Which businesses are in the sample?
906 real businesses inside a 20 mile radius of St. Louis Park, Minnesota, collected from the Google Places API and then filtered to those with a verifiable local presence. They span 18 categories, chosen because they are the service categories most small businesses in this metro compete in.
What were the assistants asked?
Three buyer-style questions per category, phrased the way a person actually asks: the best provider in the market, a recommendation near the market, and which ones people trust. Each question was put to two models, Anthropic Claude Haiku 4.5 and OpenAI GPT-4o-mini, at temperature zero, for 120 prompts in total. A business counts as recalled if the answer names it or its website domain.
How is a business counted as not recalled?
It was never named across any of the answers for its category. Name matching is deliberately generous: we count a match on the business name or its domain, and we accept partial name matches. That bias means the true share of businesses assistants cannot name is, if anything, slightly higher than we report.
What are the known limitations?
Four, stated plainly. Models change without notice, so a release is a snapshot rather than a constant. Businesses with very generic names can match by coincidence and be scored as recalled when they were not. Temperature zero reduces but does not remove run-to-run variation. And categories with small samples are pooled rather than published as their own percentage, so no headline rests on a handful of businesses.
Where does the website data come from?
A separate crawl of those same businesses' own websites, fetching each site once and recording whether it loaded, whether it presented a valid certificate, and whether it declared a mobile viewport. Sites that did not load are excluded from the security and mobile percentages, because a timeout is an unreachable site rather than an insecure one.
Is any individual business identified?
No. Every figure published is an aggregate across a category. We do not name a business as not recalled, and we do not sell the list. The raw dataset available for download contains category-level counts only.
Can this be reused?
Yes, with attribution, under a Creative Commons Attribution 4.0 licence. Journalists and researchers can request the underlying question set and the exact model versions by email.
Questions about the method, or want the raw question set? Email us.