Half of consumers now ask AI for business recommendations. Two-thirds of local businesses appear in none of the answers.

The demand side: BrightLocal's numbers, and their dates

The dates matter, so first things first: BrightLocal's Local Consumer Review Survey 2026 was published 2026-02-11, and its AI cut, “Half of consumers are asking AI for business recommendations”, on 2026-03-10. This is not this week's news — it is the best demand-side measurement in local search, and we are pairing it with a supply-side measurement we took today. Method, stated on BrightLocal's own page: n = 1,002 US adult consumers via a SurveyMonkey age-representative panel; the AI questions run on the 455 respondents who used AI for local business recommendations in the past 12 months. Re-read on 2026-10-02: every figure in the table below is still on the page (its dateModified reads 2026-04-15), and the ranking sentence reads "the third most used tool for local business recommendations, behind only Google and Facebook".

The supply side: measured 2026-08-03, recounted 2026-10-02

Every AskedAbout audit asks ChatGPT and Perplexity the questions a buyer actually types ("I'm looking for a med spa in Austin, who do you recommend and why?") and records the verbatim answers. On 2026-08-03 we cut the full production corpus: 179 completed audits of real local businesses (med spas, dentists, personal-injury law, home services) and 1,060 answers. On 2026-10-02 we recounted it from a fresh export: 183 audits and 1,084 answers, so four audits were added in two months. We count only the unbranded answers, the ones whose prompt does not contain the business's name, because the third prompt of every audit names the business and would guarantee a hit. The 08-03 cut matched names on word boundaries and hand-rejected false positives. The 10-02 recount applies the same word-boundary rule by script, with no hand pass, and sets it beside the audit's own name judge, which also accepts short forms of a name ("Sanjiva Medical Spa" for Sanjiva Med Spa). The exact-name rule is the floor; the judge's count is the ceiling.

The two counting rules disagree on 25 of 723 answers, and every one of them runs the same way: the judge says named, the exact-name rule says absent. In each case the engine wrote a variant of the name (a short form, an extra word, a different spacing), so the exact-name rule under-counts and the judge's figure is the better estimate of how often a buyer would recognise the business in the answer. Ten businesses move from zero to at least one named answer under the judge. Recount artifact: content/news/supply-side-recount-2026-10-02.json.

The vertical split is new to this page. Dentists are named in 9.6% of their unbranded answers and 38 of 48 practices are named in none; the other three verticals sit between 16.4% and 17.9%. Our dental guide carries the dental cut on its own, with the recommendation-only questions counted the way an audit scores them. Audits run from 2026-06-12 to 2026-09-09.

Put the two halves together and the arithmetic is uncomfortable: the channel 45% of consumers now use does not surface two-thirds of the local businesses we have measured by exact name, and three in five by the looser count. This is not a long-tail phenomenon where everyone gets a small slice. The distribution is concentrated absence: 122 of 183 businesses appear in none of their buyer answers, 30 appear in exactly one, and 6 appear in every answer we sampled. The engines are not spreading recommendations around; they are repeating a short list, an effect we measured directly in our brand-reappearance study (already-surfaced businesses get re-surfaced at 3.5× the rate).

The least-quoted number is the most actionable one

Everyone quotes BrightLocal's 45%. Almost nobody quotes the 88%: among consumers who get an AI recommendation, 88% fact-check it — 51% to see whether a review is legitimate, 37% to see the source. An AI mention is not the end of the funnel; it is a referral into your reviews and the pages the engine cited. That means the citation layer — which sources the engine linked when it named you — is not SEO trivia, it is the surface a majority of your AI-referred customers actually inspect before calling. Being named by an engine that cites a thin directory page is a different outcome from being named next to your own site.

Limits, on both halves

• BrightLocal's numbers are self-reported survey data — n = 1,002 via an online panel, and the AI-specific figures rest on the 455-respondent subsample. Surveys measure what people say they do; the 6%→45% jump partly reflects AI answers becoming ambient in interfaces people already used.
• Our corpus is 4 verticals of US local SMBs on two engines with fixed prompts. 183 businesses is a real sample, not the economy, and only four were added between the 08-03 cut and the 10-02 recount, so the two readings are the same corpus counted twice, not a trend. Verticals differ; the table above and our 10-business benchmark show the spread.
• The two datasets measure different populations — BrightLocal sampled consumers, we sampled businesses. Pairing them is arithmetic about a channel, not a joint study.
• We have a commercial interest. We sell the audit that produces the supply-side number. Every figure above is reproducible from the corpus, and the method (unbranded-only denominator, word-boundary matching, the judge count beside it) is published so it can be checked.

The takeaway

For a local business, the question stopped being “should I care about AI search” — 45% of your customers are already asking, and three in five businesses like you are invisible when they do. The first step costs 60 seconds: check which engines name your business and which sources they cite. For agencies, the demand-side survey plus a client-by-client supply-side measurement is the pitch deck writing itself — the 5-pack exists for exactly that.