Local Search After AI Overviews: What Happened to the Three-Slot Map Pack

Three Slots Used to Be the Whole Decision

For roughly a decade, local search had a simple shape. Someone typed a service and a place, Google drew a map with three businesses under it, and most of the calls went to those three. Owners learned to care about one thing above everything else: being in the pack. Reviews, categories, proximity and the profile itself were all levers on a single outcome, and the outcome was easy to see because you could search for yourself and count.

The question owners now bring to us is whether AI Overviews broke that. The honest answer is more useful than yes or no, and it comes from two independent datasets that, read together, describe a change most local marketing has not caught up with.

As with the rest of this series, we are not publishing survey percentages of our own. Every number on this page comes from a named public study, linked where it appears, with its limits stated. The frameworks are ours.

What the Data Actually Shows

The first dataset is Whitespark’s study of AI Overviews in local search, published in May 2025: 540 queries across six industries — plumbers, personal injury lawyers, dentists, optometrists, medical clinics and real estate agents — in Houston, Phoenix and Denver. Across all of them, an AI Overview appeared on an average of 68% of queries and a local pack on 39%.

That headline invites the wrong conclusion. The useful part is the split by what the searcher was doing:

The second is BrightLocal’s local AI visibility study, published in September 2026: about 200,000 non-branded, localized prompts for roughly 1,300 business locations over 60 days, run through Google AI Overviews, Google AI Mode and ChatGPT. Three of its findings matter here. AI Overviews named an average of 2.5 businesses per answer. Running the same prompt again produced only 20 to 33 percent overlap in the businesses named. And Google Maps mentioned a tracked business in 66% of searches, against 32 to 38 percent for the AI surfaces.

The Two-Stage Local Search

Put the two studies side by side and a shape appears that neither states on its own. A local purchase is not one search. It is a sequence, and the sequence now crosses two different surfaces.

Stage one: the shortlist

Before anyone types “dentist near me,” they ask questions. Is this tooth pain urgent. What does an implant usually cost here. How do I choose between a general dentist and a specialist. These are the informational and hybrid queries in the table, and they are answered by an AI summary nine or more times in ten, with a map almost never in sight. When that summary names businesses, it names two or three — and those names are the shortlist.

Stage two: the selection

Then the searcher switches to a request — a service and a place, or a business name they just read — and the map appears, exactly as it always did. Reviews, hours, distance, the phone button. This is the stage the map pack still owns outright.

The map pack used to be where the local decision was made. It is now increasingly where a decision made somewhere else gets confirmed. A business can hold a strong map position and still lose, because the customer arrives at the map already carrying two names, and neither of them is yours.

How much of the decision sits in stage one depends on the purchase. A locksmith at eleven at night has no research stage. A family choosing an orthodontist, a homeowner pricing a roof, or anyone hiring a lawyer has a long one. The more expensive and less frequent the purchase, the more of the decision has moved upstream of the map.

Rank Is a Position. AI Visibility Is a Frequency.

The second change is subtler and, for measurement, more disruptive. A map pack result is stable enough to be called a rank: search twice from the same place and you will usually see the same three businesses. An AI answer is regenerated. BrightLocal’s 20 to 33 percent repeat overlap means that most of the businesses named in one run are not named in the next, and a business that appears once came back in only about half of later runs.

You do not rank in an AI answer. You appear in some fraction of them. We call that fraction share of answers, and it is the only honest way to describe stage-one visibility. A screenshot showing your business in an AI Overview proves you can appear. It says nothing about how often you do, which is the thing that determines how many shortlists you are on.

This also explains a conversation we have constantly: an owner sees their business named by an assistant, a week later sees a competitor named for the same question, and concludes something broke. Nothing broke. Both businesses have a share of answers above zero, and the owner sampled once. It is the same measurement trap as the one in our report on the attribution blind spot, arriving through a different door.

Three Slots Became Two and a Half, and Sometimes None

The shortlist also got shorter. The local pack had three fixed slots. BrightLocal measured AI Overviews naming 2.5 businesses on average, and Google’s AI answers naming no business at all about 11 percent of the time. AI Mode named 3.5 and ChatGPT 4.1, but AI Overviews are the surface that appears on top of ordinary Google results, which makes the tightest list the one most people see.

Two and a half slots that rotate between runs is a very different competition from three slots that hold. The rotation is, perversely, good news for smaller businesses: a fixed three-slot pack locks out everyone in fourth place, while a regenerated answer gives any credible business a share above zero. The bad news is that “credible” is now decided by what the assistant can find and confirm about you, not by distance from the searcher.

One more finding is worth a sentence. BrightLocal found storefront businesses mentioned at a 33.6 percent rate against 29.4 percent for service-area businesses. The gap is modest, but it runs in the direction you would expect: an address is one more thing an assistant can confirm. Service-area businesses tempted to fake that presence with a grid of thin city pages should read the fake-local trap in our guide to local SEO vs national SEO first.

Where the Shortlist Gets Its Names

BrightLocal analyzed close to two million citations behind those answers. Business websites accounted for 42 percent of them, Google Business Profile about 28.5 percent and Yelp about 9.5 percent. That can look like it contradicts our September teardown of what AI assistants actually cite, which argued that a business’s own site usually confirms a recommendation rather than earning it. It does not. A source that confirms is cited constantly, because every answer that names you needs something to check you against. What the citation data adds is a warning: if the site is thin, there is nothing there to cite, and the confirmation step fails quietly.

The practical reading is a division of labor. The profile carries stage two: hours, reviews, photos, the call button. The website has to carry stage one: written answers to the questions people ask before they choose, in sentences an assistant can lift and attribute. Most local business websites were built for stage two only — a services list, a contact page, a gallery — and have nothing in them that answers a question.

What to Do About It: Two Scorecards

The response is not to abandon the map. It is to stop measuring a two-stage process with a one-stage scorecard.

Scorecard one: selection

Keep tracking map position for your plainly local queries — the ones that pass the map test in our local vs national guide. This number still means what it always meant, and the profile work behind it is covered in do I need a website if I have a Google Business Profile.

Scorecard two: share of answers

Write down ten questions your customers ask before they pick anyone — cost, urgency, how to choose, what to expect. Run each one in AI Overviews, AI Mode and ChatGPT, logged out, on a fixed day each month, and record whether you were named. The percentage of runs that include you is your share of answers. Ten questions across three surfaces is thirty samples a month, which is enough to see a trend and too few to panic over any single run.

Then close the gap the second scorecard exposes. Every question where you score zero is a page, or a section of one, that your site does not have yet. That is the work our answer engine optimization engagements and SEO services are built around, and the practical version of it for a single business is in our guide to getting recommended by ChatGPT.

What This Analysis Cannot Tell You, and Our Conflict

The limits first. Whitespark’s study is from 2025 and covers six industries in three cities; Google has expanded AI features since, so the exact percentages will have moved, even if the split by intent is the durable part. BrightLocal measured mentions, not calls or bookings, and neither study can tell you how many customers in your market do their research in AI answers rather than on review sites or by asking a neighbor. The two-stage model is our synthesis of the two datasets. It is a framework, not a finding either study reports.

Then the conflict. We sell the stage-one work this report says matters, so read it knowing that. What makes it less self-serving is that the share-of-answers scorecard costs nothing to run and does not need us, and that its result can come back telling you to stay put: a business whose purchase has no research stage — the emergency call, the same-day errand — should spend almost everything on the map and very little on content. We build with Claude Code, which makes producing answer pages fast. Fast is only worth paying for if your customers actually ask the questions first.

The closing point is the one we would most like quoted. For ten years, local visibility was a place on a map. It is becoming a probability of being named, decided before the map is ever drawn — and the businesses measuring only the place will not notice the probability falling until the phone does.