The Consensus Gap
Four AI assistants were asked the same local question in 48 US markets. 70.3% of the businesses they named were named by only one of them. Being recommended by AI is not one outcome. It is four separate ones.
Analysed and written by Abdullah İskifoğlu, founder of LocalSeen. Derived from data collected 2026-08-15/16.
of the businesses named by AI were named by only one of the four assistants
were named by all four, which is 16 businesses out of 553
markets had no single business that all four assistants agreed on
Our first study asked how many local businesses AI assistants name at all. The answer was about one in eight. That number quietly assumes something it never tested: that the four assistants are broadly looking at the same short list.
They are not. Running the same published dataset through a second question, we find that 389 of the 553 businesses that any assistant named were named by exactly one of them. Not ranked lower by the others. Not mentioned in passing. Absent.
How many assistants named the same business
Every business that at least one assistant named, counted once per market, grouped by how many of the four named it.
What we found
The assistants disagree more than they agree. The closest pair, ChatGPT and Gemini, share 30.1% of their combined picks. The furthest apart, Claude and Google AI Overviews, share 8.1%. There is no pair anywhere in the matrix that agrees on even a third of its list.
Tracking one assistant is not a sample, it is a different question. A business owner who watches only ChatGPT sees 204 of the 553 businesses AI recommends in these markets, and is blind to the other 63.1%. Even Google AI Overviews, the most generous engine by a wide margin, misses 32.4%.
Google AI Overviews is the outlier, not the consensus. 72.5% of the businesses it names are named by no other assistant. It is doing something structurally different from the chat models: it reads a local results page, while they answer from what they have learned. Treating it as a proxy for the others is the single most expensive mistake in this data.
The less the assistants know, the less they agree. In coffee shops, 55.9% of picks were solo. In hair salons it was 88.3%, with not a single business all four agreed on across all eight cities. Categories with strong third-party coverage pull the models toward each other. Thin categories leave each model to guess separately.
And the uncomfortable reading. If the four assistants converged on the same businesses, AI visibility would be a hard but legible game with clear winners. They do not converge. That means a business can be the top AI recommendation in its city and have no idea, because it happens to be the pick of the one assistant nobody checked.
How much any two assistants agree
Of every business either assistant in a pair named, the share both of them named. Pooled across all 48 markets.
| Pair | Named by both | Named by either | Agreement |
|---|---|---|---|
| ChatGPT and Gemini | 78 | 259 | 30.1% |
| ChatGPT and Claude | 62 | 233 | 26.6% |
| ChatGPT and Google AI Overviews | 81 | 497 | 16.3% |
| Claude and Gemini | 29 | 195 | 14.9% |
| Gemini and Google AI Overviews | 65 | 442 | 14.7% |
| Claude and Google AI Overviews | 35 | 430 | 8.1% |
What you miss by watching only one
“Only this assistant” is the share of all 553 AI-recommended businesses that this engine alone was the one to name. “Blind spot” is the share it never named, which is what a single-engine tracker cannot see.
| Engine | Businesses named | Only this assistant | Blind spot |
|---|---|---|---|
| Google AI Overviews | 374 | 271 (72.5%) | 32.4% |
| ChatGPT | 204 | 64 (31.4%) | 63.1% |
| Gemini | 133 | 34 (25.6%) | 75.9% |
| Claude | 91 | 20 (22.0%) | 83.5% |
Agreement by category
Eight cities pooled per category. The pattern holds in every city we looked at, which is why the category column is the one worth reading.
| Category | Businesses named | Named by only one | Named by all four |
|---|---|---|---|
| hair salon | 77 | 88.3% | 0 |
| plumber | 95 | 78.9% | 0 |
| dentist | 91 | 75.8% | 1 |
| law firm | 80 | 71.3% | 6 |
| gym | 108 | 58.3% | 3 |
| coffee shop | 102 | 55.9% | 6 |
How we measured
No new data was collected for this page. Every number here is derived from the dataset we already published, which means you can reproduce all of it yourself without an API key.
- The question
- Given the businesses at least one AI assistant named for a buyer query, how many of the four assistants named each one? Agreement is measured between engines, not against Google Maps. The share of the Maps roster that gets named at all is the first study.
- Source
- The published AI Local Visibility Index dataset, collected 2026-08-15/16: 48 markets, 8 US cities by 6 categories, four engines, three fixed question phrasings each. The collection rules are on the index page and were frozen before that data was gathered.
- Unit of counting
- One business in one market is one row. The same chain appearing in two cities counts twice, because the question is asked per market. A business named repeatedly by one engine still counts once for that engine.
- Denominator: 553, not 556
- The dataset lists 556 named businesses. Matching names across engines requires normalising them, and normalising reveals three pairs that differ only in capitalisation (in gym/Nashville, coffee shop/Columbus and coffee shop/Tucson). After deduplication 553 distinct businesses remain. We report the lower figure because the higher one would count the same business twice in an agreement measure.
- All four engines, all 48 markets
- The dataset flags one market (gym/Denver) where the AI Overviews block was not visible on the results page, yet records 8 businesses named by AI Overviews there. The flag tracks the block, not whether an answer was captured, so the market is included in full. Excluding it would discard real data and move the headline by less than a point.
- Agreement between a pair
- Of every business either engine named, the share both named. This is a Jaccard index, chosen because the engines name very different numbers of businesses and a simple overlap count would flatter whichever one talks most.
- Reproducing it
- The derivation script is
web/scripts/derive-consensus.ts. It reads the published JSON and prints the headline figures on this page. Cost to run: nothing.
Every market
All 48 rows, so the headline can be recomputed from the parts.
| Category | City | Named by AI | By only one | Solo share | By all four |
|---|---|---|---|---|---|
| coffee shop | Austin, TX | 10 | 6 | 60.0% | 0 |
| gym | Austin, TX | 17 | 12 | 70.6% | 1 |
| dentist | Austin, TX | 9 | 8 | 88.9% | 0 |
| plumber | Austin, TX | 11 | 8 | 72.7% | 0 |
| dentist | Denver, CO | 14 | 11 | 78.6% | 0 |
| hair salon | Austin, TX | 16 | 15 | 93.8% | 0 |
| law firm | Austin, TX | 7 | 7 | 100.0% | 0 |
| gym | Denver, CO | 16 | 12 | 75.0% | 0 |
| coffee shop | Denver, CO | 13 | 8 | 61.5% | 1 |
| plumber | Denver, CO | 13 | 8 | 61.5% | 0 |
| law firm | Denver, CO | 9 | 7 | 77.8% | 0 |
| hair salon | Denver, CO | 9 | 9 | 100.0% | 0 |
| dentist | Portland, OR | 10 | 5 | 50.0% | 1 |
| plumber | Portland, OR | 12 | 10 | 83.3% | 0 |
| gym | Portland, OR | 15 | 7 | 46.7% | 0 |
| coffee shop | Portland, OR | 13 | 6 | 46.2% | 2 |
| law firm | Portland, OR | 13 | 7 | 53.8% | 2 |
| dentist | Nashville, TN | 11 | 6 | 54.5% | 0 |
| hair salon | Portland, OR | 8 | 8 | 100.0% | 0 |
| plumber | Nashville, TN | 12 | 10 | 83.3% | 0 |
| gym | Nashville, TN | 9 | 4 | 44.4% | 0 |
| coffee shop | Nashville, TN | 12 | 8 | 66.7% | 1 |
| law firm | Nashville, TN | 9 | 7 | 77.8% | 0 |
| dentist | Columbus, OH | 10 | 9 | 90.0% | 0 |
| hair salon | Nashville, TN | 11 | 10 | 90.9% | 0 |
| plumber | Columbus, OH | 13 | 12 | 92.3% | 0 |
| gym | Columbus, OH | 13 | 6 | 46.2% | 0 |
| hair salon | Columbus, OH | 8 | 5 | 62.5% | 0 |
| dentist | Raleigh, NC | 13 | 9 | 69.2% | 0 |
| law firm | Columbus, OH | 10 | 7 | 70.0% | 1 |
| coffee shop | Columbus, OH | 12 | 6 | 50.0% | 2 |
| plumber | Raleigh, NC | 10 | 8 | 80.0% | 0 |
| gym | Raleigh, NC | 12 | 8 | 66.7% | 0 |
| law firm | Raleigh, NC | 13 | 9 | 69.2% | 1 |
| coffee shop | Raleigh, NC | 18 | 12 | 66.7% | 0 |
| hair salon | Raleigh, NC | 10 | 8 | 80.0% | 0 |
| dentist | Tucson, AZ | 12 | 12 | 100.0% | 0 |
| gym | Tucson, AZ | 13 | 6 | 46.2% | 1 |
| plumber | Tucson, AZ | 11 | 9 | 81.8% | 0 |
| law firm | Tucson, AZ | 8 | 6 | 75.0% | 0 |
| hair salon | Tucson, AZ | 7 | 5 | 71.4% | 0 |
| coffee shop | Tucson, AZ | 14 | 7 | 50.0% | 0 |
| dentist | Boise, ID | 12 | 9 | 75.0% | 0 |
| gym | Boise, ID | 13 | 8 | 61.5% | 1 |
| law firm | Boise, ID | 11 | 7 | 63.6% | 2 |
| plumber | Boise, ID | 13 | 10 | 76.9% | 0 |
| coffee shop | Boise, ID | 10 | 4 | 40.0% | 0 |
| hair salon | Boise, ID | 8 | 8 | 100.0% | 0 |
What this does not tell you
- Disagreement is not the same as being wrong. Nothing here says one assistant picks better businesses than another. It says they pick different ones, which is a claim about coverage, not quality.
- Name matching is strict. A business counts as named only if its name appears as a phrase. An assistant that describes a business without naming it does not register, and the strictness applies equally to all four, so it affects the levels more than the comparison.
- It inherits every limit of the source study. One country, one language, one collection window, cheap model tiers, and a Google Maps roster capped at depth 100. Those are documented on the index page and none of them are fixed by asking a second question of the same data.
- Chat models are sampled, AI Overviews is not. The chat models were asked three times per question and AI Overviews once, because it repeats itself. More samples give the chat models more chances to name something, which if anything works against the finding that they name fewer businesses than AI Overviews.
- One point in time. Models change. This is what four of them did over two days in August 2026, and it is a reason to keep measuring rather than to treat these numbers as fixed.