Why AI Local Recommendations and the Google Map Pack Disagree: A Cross-Vertical Look at the Citation Gap

Ranked first in one place, absent from the other

For two decades, "being found locally" meant one thing: appearing in the block of three business listings with a map that Google shows above the organic results, the Local 3-Pack that most people call the map pack. Winning it mattered, and it still does. Local searchers click those pack results far more often than the organic or paid links beneath them.

The problem is that a second surface now sits alongside it. When a buyer opens ChatGPT, Perplexity or Gemini and asks for the best med-spa nearby, or a plumber who can come today, the assistant returns a short list of names it composed. That list is not the map pack. It is generated by a different system reading a different set of sources, and the two do not have to agree. A business can be the first result in the map pack and go entirely unmentioned in the AI answer written for the same query. Being absent there is not the same as ranking fourth in a list; it is being left out of the reply the buyer actually read.

What the Google map pack actually reads

The map pack is a local-search product with a well-documented set of inputs. Google assembles it from three broad signals its own guidance names: relevance (how well a profile matches the query), distance (proximity to the searcher), and prominence (how established and well-regarded the business is, including its reviews). The primary object those signals attach to is the Google Business Profile, the structured listing a business claims and maintains, cross-checked against Google's local index of the wider web.

This has a practical consequence. Most of what moves a local pack ranking is inside a system a business can see and manage: the profile's categories, hours, service area, photos, and the volume, recency and rating of its reviews, kept consistent with the name, address and phone number that appear elsewhere online. Local SEO, as a craft, is largely the craft of feeding that one well-understood machine cleanly. It is knowable, and it is measurable.

What AI local recommendations read instead

A general AI assistant does not query the map pack and hand you the result. It composes an answer from a different foundation, and the difference is the whole story.

The size of the gap, and how firmly it can be stated

How large is the disagreement? The most direct attempt to measure it comes from BrightLocal, which reported that a business ranking in Google's top local-pack results has less than even odds of also appearing in AI local recommendations, and that earning visibility in ChatGPT's local recommendations is far harder to obtain than ranking in the map pack, on the order of a thirty-times visibility gap in its sample.

That is a striking figure, and it should be handled with matching care. It rests on a single vendor's proprietary study rather than on independent, peer-reviewed measurement or multiple corroborating datasets. The direction it reports, that classic local ranking and generative recommendation are pulling apart, is consistent with the broader classic-versus-generative divergence documented across search. The precise magnitude is not yet settled fact. The gap is real and worth acting on, while the exact multiplier is an early estimate to verify against your own market, not a number to quote as if it were audited.

This is also why the responsible response is to measure your own gap rather than to import someone else's headline. A ratio observed across one vendor's national sample may look very different for a specific med-spa in a specific city on a specific engine.

Why the gap matters now, not later

A disagreement between two systems only matters to the extent buyers use the second one. The evidence says they have started to, quickly.

BrightLocal's consumer survey reported that 45 percent of consumers had used an AI tool such as ChatGPT, Gemini or Perplexity to find a local business recommendation in the trailing year as of its 2026 edition, up from 6 percent a year earlier. That specific jump is a large single-source swing and belongs in the emerging column until a second source confirms it. The surrounding adoption picture, however, comes from disclosed, survey-based work: Pew Research Center found that 49 percent of US adults now use chatbots, up from 23 percent in 2023, and that 42 percent of chatbot users use them specifically to search for information.

Two cautions keep this grounded rather than breathless. First, use is not the same as trust: Pew found only 29 percent of US adult chatbot users trust the information they get "a lot" or "some," and general trust in online reviews has drifted down from an 84 percent peak in 2016 and 2017 to roughly half of consumers today. Second, the map pack is not going anywhere. The point is not that one surface replaced the other; it is that a buyer population is now split across two systems that name different businesses, so being strong on only one of them means being invisible to a growing share of the people asking.

The same gap, read across four verticals

The mechanics of the citation gap are general, but the stakes differ by vertical because the buyer's risk and the trust signals differ. Four of our lead and expansion verticals illustrate the range without needing a separate statistic for each.

• Med-spa and aesthetics: the buyer often cannot verify from any listing the thing that actually governs safety, the supervising-physician structure and practitioner credentials, and non-board-certified practitioners performing aesthetic procedures is a documented patient-safety concern. A pack ranking says nothing about that; an AI answer that synthesizes reviews and third-party mentions may weigh it differently again, so a credential-legible presence has to be built for both surfaces, not just the profile.
• Dental: patients vet a new practice through reviews, location and insurance before booking, and narrative reviews measurably shift which provider is chosen. The map pack surfaces the profile and its rating; an assistant may lean on review text scattered across aggregators, so the practice that reads well in the pack can still be summarized unfavorably, or omitted, in the answer.
• Home services: the map pack captures a disproportionate share of "plumber near me" clicks, but a homeowner increasingly asks an assistant for a recommendation too. A contractor who owns the pack yet is thin across the wider web the model reads pays twice, once in lost AI-answer presence and again if the gap pushes them back toward paid lead marketplaces.
• Restaurants: reputation moves revenue most for independents, where a one-star Yelp increase raised revenue by 5 to 9 percent in the foundational study, and it is precisely those independent, review-driven signals that a general assistant synthesizes differently from how the map pack ranks them.

Reading your own gap: measure both surfaces, assume neither

Because no engine publishes who it names or why, the only way to know your standing is to measure it directly on each surface and compare, rather than to infer one from the other.

On the map pack side, the read is relatively stable: sample your priority queries from your service area and record where the business appears across the pack over time. On the AI side, the read has to respect the non-determinism. That means freezing a panel of the real questions your buyers ask, running each across a chosen engine many times, and reporting how often the business is named as a rate with a confidence band, stamped with the engine, the locale and the date. A single query on a single day is a coin toss, not a measurement.

Put the two reads side by side and the citation gap stops being a general claim about the industry and becomes a specific, dated fact about your business: strong pack, thin AI presence; or the reverse; or a gap that varies by question. That comparison is the starting point for any plan, because it tells you which surface is actually costing you buyers before a dollar of work is scoped.

What closes the gap, and what cannot be promised

If the two systems read different data, the work is to become legible to both, not to game either. Three levers do most of the closing.

The first is a clean, unambiguous entity: one consistent identity, name, address, phone and category across the profile, the site and the third-party sources these systems read, so no engine has to guess who you are or confuses you with someone else. The second is content and structured data an engine can lift a sentence from and attribute, since an assistant can only name what it can extract and trust. The third is third-party corroboration on the sources these engines actually weight, because presence in a general assistant is earned across the wider web, not inside a profile a business controls alone.

What cannot be promised is a citation. These engines are non-deterministic and their source weighting is undisclosed, so no assistant's answer on a given day is guaranteed. The work is to make the business genuinely readable and trustworthy to both systems, measure the position on each, and move it.