Who Google AI Mode Recommends, and Who It Leaves Out

What we measured, and why this design

Every piece of advice about AI visibility for local business rests on an assumption nobody has checked in public: that the businesses an AI assistant recommends are the businesses that already rank. If that holds, the advice is simple. Rank in Maps and the AI follows. If it does not hold, a business can be losing answers it should be winning, and no amount of attention to the map pack will show it.

So we tested it. Fifty questions, each one phrased the way a customer would type it. Five trades, ten cities, three city sizes. For each question we recorded two things on the same day: which businesses Google AI Mode recommended in its answer, and which businesses held the top twenty spots in Google Maps for the equivalent search.

That gives two groups drawn from the same demand, at the same moment, in the same place. Businesses the AI put forward, and businesses that ranked and got no mention. Every number below compares those two groups.

The trades are the ones where local search decides real money: dentists, personal injury lawyers, plumbers, Italian restaurants, auto repair shops. The cities are New York, Chicago and Houston at the top, Charlotte, Columbus and Sacramento in the middle, then Boise, Chattanooga, Stamford and Boca Raton at the smaller end.

Finding 1: an AI answer is not the map pack reshuffled

Of the 247 businesses Google AI Mode recommended across the fifty answers, 78 of them, very nearly a third, did not appear anywhere in the Google Maps top twenty for the same search. Not at rank four and not at rank nineteen either.

Of the 169 that did appear, the median sat at rank six, and a quarter of them were outside the Maps top ten. Across all 247, barely one in five held a top-three Maps spot.

Where the 247 recommended businesses sat in Google Maps	Count	Share
Maps top 3	54	22%
Maps top 4-10	73	29%
Maps top 11-20	42	17%
Not in the Maps top 20 at all	78	32%

If you have been checking your map pack position and reading it as your AI visibility, this is the number to sit with. The two overlap, and they are a long way from being the same thing.

Finding 2: the businesses it picks share a profile signature

The choice is not random. The two groups differ, sharply and consistently, on the attributes a business actually controls.

Attribute (median)	Recommended	Ranked, not mentioned	Test
Google reviews	595	269	p = 1.1 × 10⁻¹⁰, A = 0.66
Profile photos	72	39	p = 0.00017, A = 0.59
Star rating	4.9	4.8	p = 0.0017, A = 0.58
Has a website	98%	93%	p = 0.022

The A statistic is the plain-language version of the effect. Pick one business from each group at random and there is a 0.66 chance the recommended one has more reviews. That is a real separation rather than a rounding artifact of a large sample.

Two attributes are missing from that list for a reason. Publishing opening hours separated nothing, because essentially every business in the top twenty publishes hours, so it works as a floor rather than an advantage. Having a website is close to a floor too, and in only one category did it look like a lever.

Trade by trade

The pattern holds in four trades out of five, and the size of the gap swings enormously.

Trade	Reviews, recommended	Reviews, not mentioned	p	Photos, recommended	Photos, not mentioned
Home services (plumbers)	731	134	0.00036	119	51
Dentists	734	352	0.000027	38	24
Law firms	341	137	0.00066	38	23
Auto repair	332	212	0.000017	47	25
Restaurants	916	973	0.89	722	660

Home services shows the widest gap of all. A recommended plumber carried five and a half times the reviews of a plumber who ranked and went unmentioned. Auto repair had the narrowest gap on reviews and the cleanest separation on rating, where 4.8 against 4.7 was one of the most reliable differences we measured.

Finding 3: the ones picked from outside the map pack are not outsiders

The obvious objection to Finding 1 is that the 78 businesses recommended from outside the Maps top twenty must be marginal, a scattering of odd picks and small operations the ranking system correctly passed over.

They are not. We pulled the Google profile for each of them separately. Of the 78, 73 had a findable profile, and those profiles look like this:

Median	Recommended, in Maps top 20	Recommended, outside Maps top 20	Ranked, not mentioned
Reviews	595	528	269
Rating	4.9	4.8	4.8
Photos	72	66	39
Has a website	98%	97%	93%

The businesses AI Mode reaches past the map pack to recommend look statistically indistinguishable from the ones it recommends inside it, and both are a different population from the businesses that rank and get no mention. So what gets picked follows the profile more closely than it follows the ranking. A strong profile that is not winning its map pack can still win the answer.

Finding 4: restaurants behave differently, and that matters

In restaurants, review volume did nothing. Recommended restaurants had 916 median reviews and the ones left out had 973. The test returns p = 0.89, which is a polite way of saying the two groups are the same. Photo counts did nothing either.

Star rating was the one attribute that still separated them, 4.6 against 4.5, and it did so reliably.

The reason shows up in the answers themselves. For Italian restaurants, AI Mode wrote about critics, neighborhood reputation and signature dishes, and quoted city magazines, Reddit threads and food media. For plumbers and dentists it rendered Google business cards with star ratings and review counts, then reasoned from them.

The consequence is blunt. Advice built on review volume belongs to a trade, and anyone selling it as universal has not checked. A restaurant chasing its thousandth review to win an AI answer is chasing the wrong thing. A plumber doing the same is doing exactly the right thing.

Finding 5: city size changes the strength of the effect, not its direction

City tier	Reviews, recommended	Reviews, not mentioned	p
Large (NYC, Chicago, Houston)	580	340	0.021
Mid (Charlotte, Columbus, Sacramento)	642	285	0.000073
Small (Boise, Chattanooga, Stamford, Boca Raton)	554	219	0.00000054

The separation is cleanest in small cities and weakest in the largest ones. In New York or Chicago, where hundreds of businesses clear any plausible threshold, the profile attributes stop telling anyone apart and something else decides the answer. In a smaller market the profile still does most of the work, which is good news for the businesses least able to buy their way into an answer.

One more small-city detail: every business AI Mode recommended in a small city had a website. Every single one, against 94% of the businesses that ranked and were left out. It is the only place in this data where a website looked like a lever rather than a floor.

Finding 6: what the answers quote

Across 50 answers there were 291 citations, a median of six per answer. Stripped to domains and counted by how many of the fifty answers cited each at least once:

Source	Answers citing it
Google’s own business entity cards	72%
Reddit	36%
Yelp	36%
Justia	12%
Angi	12%
Facebook	8%
Instagram, Yahoo Local	6% each

Beyond those, half of all citations went to businesses’ own sites, local media and “best of” roundups.

Reddit turning up in more than a third of local commercial answers is the number worth sitting with. It is not a directory, it cannot be optimized in the usual sense, and it cannot be bought. It is people talking, and Google is reading it back to customers who ask which plumber to call.

What a local business should take from this

Four things follow directly from the data, and one thing does not follow but is worth saying anyway.

Stop reading map pack position as AI visibility. The two come apart for a third of the businesses in this sample. Check the answer itself, in your trade and your city, and check it more than once, because the answers vary between runs.

Review volume is the strongest profile lever outside dining. The gap between the two groups is a factor of two overall and a factor of five in home services. That is not a threshold to clear. It is a direction that pays across the whole range.

Keep the profile actively maintained. Photo counts separate the groups almost as reliably as reviews do. Whether photos are the cause or simply mark a profile someone tends to, the businesses getting recommended are not the ones that filled in a profile once in 2019.

Be present where the answers look. A third of these answers quoted Reddit, a third quoted Yelp, and half of all citations went to independent sites and local roundups. Being discussed away from your own property is not a layer on top of profile work. In this data it is how a business gets recommended without ranking.

Here is what does not follow. None of this tells you that being recommended in an AI answer produced a customer. The study measures visibility, not revenue. Anyone who tells you they can convert one into the other on the strength of a correlation study is telling you something the data does not contain.

What this study does not show

Stated plainly, because a study that hides its limits deserves to be ignored.

One measurement window. All of it was collected on 24 August 2026. AI Mode answers are generated per query and vary between runs, so a repeat next month will recommend a different set. The pattern should survive. The specific businesses will not.

One question template per trade. “Who is the best dentist in {city}?” behaves differently from “emergency dentist open now near me”. Comparability across cities was bought at the cost of coverage across intents.

Correlation, not causation. Nothing here shows that adding reviews causes an AI to recommend you. Review count marks businesses that are established, busy and actively managed, and any of those could be the quality that matters.

English, United States, one surface. No ChatGPT, no Perplexity, no other country, no other language.

Matching is imperfect. Businesses were matched by normalized name within trade and city at a 0.85 similarity threshold, with the near-misses reviewed by hand. A handful of edge cases in either direction would not move the headline numbers, but they exist.

Methodology in full

Questions. Five category templates, phrased as natural questions, applied unchanged across ten cities:

Category	Question	Maps search for the comparison group
Dentists	Who is the best dentist in {city}?	dentist
Law firms	Who is the best personal injury lawyer in {city}?	personal injury attorney
Home services	Who is the most reliable plumber in {city}?	plumber
Restaurants	What is the best Italian restaurant in {city}?	italian restaurant
Auto repair	What is the best auto repair shop in {city}?	auto repair shop

Cities. Large: New York NY, Chicago IL, Houston TX. Mid: Charlotte NC, Columbus OH, Sacramento CA. Small: Boise ID, Chattanooga TN, Stamford CT, Boca Raton FL.

AI Mode collection. Each question ran through the SerpApi google_ai_mode engine with the city set as a canonical Google location target, gl=us, hl=en, on 24 August 2026. We captured both the structured response and the rendered answer for every question.

Extraction. In 33 of the 50 answers, Google rendered business cards carrying name, rating, review count and address inside the answer, and those were parsed automatically. The other 17 gave their businesses in prose or in a comparison table: 12 were extracted by pattern matching on the answer’s structure and 5 were coded entirely by hand, and all 17 were checked line by line against the answer text. Names appearing only inside a quoted source snippet were not counted, because a business the assistant quotes someone else mentioning is not a business the assistant recommended.

Comparison group. The top 20 Google Maps results for each equivalent search, collected the same day via the Outscraper Maps API with language=en and region=US. 1,000 profiles, 998 of them unique, since two businesses surfaced in two city searches each.

Matching. Recommended businesses were matched to comparison profiles by normalized name, lowercased with legal suffixes and punctuation stripped, within the same category and city, accepting matches at 0.85 similarity or above. Every match and every near-miss was reviewed by hand. The 78 unmatched businesses were then looked up individually to capture their profile data.

Statistics. Two-sided Mann-Whitney U tests on the unpaired groups, chosen because review and photo counts are heavily skewed and nowhere near normal. The A statistic reported alongside each p value is U divided by the product of the group sizes: the probability that a business drawn at random from the recommended group exceeds one drawn from the other. Website presence was tested with Fisher’s exact test.

The data

Everything this study rests on, published in full, because numbers without their workings are just a claim.

• Question list, with answer format and counts, 50 rows
• Businesses recommended by AI Mode, with their Maps position, 247 rows
• Maps comparison group, 1,000 rows
• Every source the answers cited, 291 rows

If you run it again and get something different, we want to know. The point of publishing the workings is that someone checks them.