Local AI search has barely been measured. Google's documentation for its AI features states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that page never mentions local results, Business Profiles or Google Maps at any point, in a document written to tell site owners what to do about AI. Any guide telling a Minneapolis contractor to optimize a Business Profile for AI Mode is extrapolating well past what that source actually says.

This is the condition of the entire category...the advice is confident and the evidence behind it is thin. Headline numbers fall apart the moment somebody opens the study they came from.

What Google Has Actually Documented

Google has shipped local AI features the advice mostly walks past...three of them are dated and checkable.

  • Maps Grounding: Generally available 26 September 2025, giving developers "more than 250 million places" with daily-updated Maps information and reviews.
  • Ask Maps: Announced 12 March 2026, drawing on "over 300 million places" and a contributor community Google puts above 500 million people.
  • Gemini: Announced 10 June 2026, reading a Business Profile for its owner rather than feeding that profile into public answers.

Google publishes no error rate for Maps grounding. All three do a different job from the one being sold...the pitch is that a tidy profile changes what an assistant tells a stranger about a business, and Google has documented no such path from one to the other.

The Number Everyone States Backwards

Sterling Sky measured 5,943 unique businesses appearing in AI local packs against 18,330 in regular three-packs across the same query set. Ten seconds of arithmetic settles it...that is 32% as many, which is 68% fewer. Several 2026 publishers restate the finding as "32% fewer local businesses", a far milder claim that happens to be wrong by half, and it is the single most repeated statistic in this category. Sterling Sky used Places Scout. Its other findings matter more.

  • Frequency: AI local packs appeared on roughly 7% of the tracked keyword set.
  • Shape: One or two businesses named in place of the familiar three, with no call button.
  • Membership: Different businesses from the ones the traditional three-pack showed on the same queries.
  • Spread: 88% of 322 markets named fewer unique businesses in the AI version.

Membership matters more than the count...local pack rank tracks something other than AI visibility, and that is the part sitting under a division the restatements skipped.

The 68 Percent Misread

Search Engine Journal reports that "AI Overviews are currently appearing in 68% of local business-type queries", which is the most repeated line about AI Overviews in local search...it takes an average built across three intents and pins it to one of them. Whitespark ran 540 manual queries across Houston, Phoenix and Denver, six verticals, split three ways by intent. Collection ran in Q2 2025.

  • Local intent: AI Overviews 15%, local packs 93%.
  • Informational intent: AI Overviews 92%, local packs 6%.
  • Hybrid intent: AI Overviews 97%, local packs 17%.

Averages hide the split, because 68% is an average across all three intents. Intent does all the work. On the queries a plumber actually cares about, the map pack showed 93% and the AI Overview 15%...a finding about informational search gets filed under local and sold back as urgency.

Then the second large study says something else again. Local Falcon's 60,000-search grid test, covering 4,423 businesses across roughly 21,000 distinct search points, puts AI Overview prevalence at 40.2% overall. Whitespark and Local Falcon disagree by roughly 28 points on the headline, and published estimates for local business queries run from 15% to 68% depending on how "local" is defined...that range is the honest state of the measurement.

Inside that same test, a place name in the query makes an AI Overview less likely. Local Falcon puts it at 35.0% on queries carrying a location name against 46.1% on queries without one, a gap of 11.1 points running downward. SE Ranking holds the other half of that same edit. Re-running a location-named query returns 53 to 56% domain overlap, against 34 to 37% for a query like "restaurants near me". Either figure quoted without the other tells half the story...naming the place buys fewer AI Overviews and steadier ones.

Fast Adoption, Slower Trust

Consumer behaviour moved faster than the measurement did. BrightLocal surveyed 1,002 American adults and filtered its AI questions down to the 455 who had asked an AI tool for a business recommendation in the previous twelve months. Fieldwork ran in 2026 and the sample was national.

  • Asked AI: 45% of the full sample, against 6% a year earlier.
  • Tool split: ChatGPT at 31% and Google AI Mode at 23% among those 455 users.
  • Trust: 63% trust the answers, and 88% say they check them.

An AI answer works as a shortlist. Verification is where the two surveys in circulation stop agreeing. Only 33% of GatherUp's respondents are "always clicking through on the links and sources generated by AI products", against BrightLocal's 88% who fact-check by verifying review legitimacy or checking sources. Ever-checks and always-clicks are different questions...the industry quotes them interchangeably.

Pew Research tracked actual browsing for 900 American adults across 68,879 Google searches in March 2025, of which 12,593 produced an AI summary. Users clicked a traditional result on 8% of visits where a summary appeared, against 15% of visits without one, and clicks on a source cited inside the summary ran at 1%. Pew's remains the only panel-based, non-vendor dataset within reach...and none of it is local.

How Many Reviews, Honestly

Every published page gives a different number. Seven vendors produced seven answers with no shared study between them...which is what estimation looks like when it is dressed as measurement.

  • Thresholds: 20, 25, 50, 80, 100, 150, 200.
  • Top-three rank: 240.
  • Cross-citations: None, and no shared dataset or lineage between any of them.

One vendor with a named study reports 270 queries on one page and 405 on another, and its own tier data reverses at the top, falling from a 79% mention rate in the 100 to 199 band to 69% above 200. Others ran tests of their own, on 48 queries and 15 dental clinics, or on an unpublished analysis of 50,000 businesses...each one stands on its own.

Only one controlled experiment exists...and it points somewhere else entirely. A pre-registered algorithm audit randomized hotel attributes across twelve models and 3,024 choice sets per model, then ranked what actually moved the pick. Star rating carried 35.7% of the decision weight, statistically alongside price at 33.8%, with sustainability at 13.1% and review volume at 9.4%. The paper ranks valence above volume by nearly four to one. Moving from 3.9 stars to 4.7 was worth 31.6 percentage points, while moving from 45 reviews to 2,100, a forty-six-fold increase, bought 8.3.

Volume works as a credibility weight on the rating rather than as a gate...getting picked depends just as much (if not more) on the rating as on the count. Its design was synthetic hotel cards, one vertical, twelve models. Its authors are explicit that the audit measures choice among an already-retrieved set of five rather than retrieval...whether the same weights hold for a St Paul dentist is unknown.

Inside ChatGPT's Local Pipeline

OpenAI's help centre names exactly two search partners, Bing and Shopify, and the phrase it uses is "sometimes partners". Google is absent from that list, along with Google Maps, Business Profile and the Places API, and what the absence supports is narrower than it looks. Documentation stops well short of a Business Profile as a profile, a licence to Google's local data appears nowhere, and whatever Google-origin information arrives does so second-hand. Separately, two independent inspections of ChatGPT's own JSON describe a provider labelled serp scraping the embedded Google Maps panel, and one researcher puts Google dependency at 94% across a hotel-only sample...a licensed feed in either direction remains undocumented.

  • Route in: A partner feed of local business records that ChatGPT indexes directly.
  • Record fields: Seven per business, id, name, address, coordinates, phone number, website and platform URL.
  • Status: Beta, tested with approved partners that OpenAI leaves unnamed.
  • Business Profile: Absent from the specification, and the application route at chatgpt.com/merchants takes a feed rather than a profile.

Two parties get placed in that pipeline...each claim comes from somewhere other than OpenAI. Foursquare claimed the ChatGPT search partnership on its own LinkedIn account in December 2024, unacknowledged by OpenAI and unobserved in any 2026 trace. Yelp's licensing deal was reported by Axios on 23 July 2026, and its Request a Quote flow matches the feed spec exactly...a fit rather than a confirmation.

Foursquare's supposed 60 to 70% share of ChatGPT's local results traces back through two unlinked hops to one Spanish-language test of 50 prompts, five cities, GPT-4o, May 2025, and that model generation has been replaced several times over since. OpenAI's own inclusion guidance is thinner still. Allow OAI-SearchBot to crawl the site, and let the host and/or content delivery network allow traffic from its published IP addresses.

Minneapolis and St Paul Apart

Distance is one of three factors Google names for local results, and its own help page lists them as "relevance, distance, and popularity". That is the mechanism the separate-markets claim rests on. Google itself has never said Minneapolis and St Paul are separate markets...that step is a vendor inference drawn from 34 unnamed client engagements with no published data behind it. Assistants work at a coarser grain still, and the two vendors publishing a location parameter disagree about how fine it should be.

  • Location floor: An area over 3 sq km holding at least 1,000 users.
  • Suburb size: Edina and Bloomington at roughly 40 and 90 square kilometres of land.
  • OpenAI's tool: Free text city and region, with no coordinate field at all.
  • Perplexity's contrast: Latitude and longitude, against OpenAI's worked example of Minneapolis and Minnesota as plain strings.

That reference implementation flattens Edina, Bloomington and Woodbury into one market...testing at that grain is still waiting to happen. Exactly one original Twin Cities test sits in the published record, 96 queries run in Minneapolis in July 2026, once, by a competing vendor, with the dataset held back and the run still unrepeated. A comparison between two adjacent suburbs remains unpublished, and St Paul remains unmeasured.

Variance Inside a Single City

Anybody claiming to have spotted suburb-level differences from a handful of manual checks is reading run-to-run variance, and the closest thing to a same-query, different-place experiment says why. SE Ranking, which sells AI visibility tracking and published no raw data or testing dates, re-ran location-named queries across five cities.

  • Same city: 46 to 49% of the same URLs came back on a re-run.
  • Distant cities: 41 to 44% came back a thousand miles apart.
  • Net difference: About five points for place, against roughly fifty inside one place.

No Report For Business Profiles

Google shipped AI reporting for websites in 2026...the Business Profile side is still waiting.

  • Search Console: Generative AI performance reports shipped in June 2026, covering websites only.
  • Business Profile: No equivalent report of any kind arrived alongside them, then or since.
  • Ask Maps: Glenn Gabe's March 2026 write-up observes impressions flowing into ordinary Business Profile analytics with no separate breakout, and Google has stayed silent on the point either way.

First-party reporting on it exists nowhere...a local business tracking its own standing in the local AI surfaces is working blind. Any vendor selling a local AI visibility score is selling third-party sampling rather than Google data, worth knowing before signing anything. Third-party sampling also happens to be all that exists, and vizmeta HQ at hq.vizmeta.com runs it free and monthly across ChatGPT, Claude and Google AI, logging what each one says about a business over time. It is sampling...the same caveat applies to it.

Second-hand numbers need their caveats printed next to them. SOCi's widely quoted finding that ChatGPT recommends only 1.2% of local businesses comes from a study of 2,751 multi-location brands across more than 350,000 locations...one shop with a single address reading that figure is reading a statistic about franchise estates.

Two Assets From One Invoice

Chamber membership buys two different assets...and most of the advice conflates them. In Whitespark's 2026 survey of 47 local search experts, quantity of structured citations ranks 40th for AI search visibility, while quality of unstructured citations, a category the report defines to include industry associations, ranks 4th. Thirty-six rank positions separate them.

A directory row is the structured one...member spotlights, awards and news posts on indexable pages are the unstructured one, and the Minneapolis chamber runs its member directory on ChamberMaster/GrowthZone software, where four things are true of it today.

  • Directory page: Roughly 130 kilobytes of served HTML carrying no business names at all.
  • Member links: Absent from that HTML, because JavaScript injects the listings after load.
  • Member pages: Server-rendered, carrying LocalBusiness markup in the older microdata form rather than JSON-LD.
  • Sitemap: 1,525 URLs, 1,503 of them under /list, quietly doing every bit of the discovery work alone.

Any reader can run that same check on their own chamber in about ten minutes, and should...the other chamber platforms remain unaudited and may behave differently.

Test It From the City

Checking this takes about an hour the first time through...every step of it is free.

  1. Write the exact prompts a customer would use, naming Minneapolis or St Paul in each.
  2. Run every prompt five times, because the answer moves between runs.
  3. Record the whole answer rather than a position, since who else got named is the useful part.
  4. Repeat the set with device location switched on, then off, which is the default state.
  5. Log the date, the model version and the city, then repeat the identical set next month, keep every raw answer, and compare the two months side by side.

Repetition is the method...one answer describes one run and stops there, and the volatility figures above are the reason why.

Where This Stays Unknown

The honest list is short, four claims the published evidence currently leaves unsupported.

  • Profile data: Whether it reaches AI Overviews or AI Mode at all, since Google's documentation never says so.
  • Field edits: Whether changing one changes an assistant's answer, because no controlled before-and-after test has been published.
  • Factual accuracy: How assistants perform on hours, addresses and service areas, which the published work leaves unmeasured.
  • St Paul: What any of this looks like there, in a city that no published dataset has ever covered and that the one Twin Cities test never reached.

Useful work here is narrower than the category admits...get the record consistent everywhere it appears, keep the site crawlable, then measure the same prompts from the same city every month and watch the line move. That is a floor rather than a strategy, and it is the only honest place to start.