Four AIs, one question, four different answers
Four leading AI engines chose the same top local business in 4 percent of shared tests. Their disagreement often starts with the businesses and pages each one considers.
By Anthony Rinaldi, Yext Research
An analysis of 491,972 shared local query-market cases from February through August 2026.
AI search · local recommendations · citation mix · business profiles · reviews · schema
Each engine starts with a different consideration set. Inside those sets, ratings, review depth, recent activity, and clear business data help shape which name reaches the top.
What to do. Make every location easy to find on its website, profile, and key directories. Then keep reviews, replies, photos, and posts current.
A consideration set is the group of businesses an engine names for one local question. Agreement measures whether at least three engines place the same business first.
The study used unbranded local recommendation prompts in the form “What are the top 10 [category] near [neighborhood]?” The same prompt and place were tested across all four engines.
The panel covered 3,569 keywords and 2.05 million local questions. The core comparison uses the 491,972 cases asked of all four engines across 124,820 market observations each month. Cases with fewer than four engines were excluded from the consensus comparison.
The same question rarely produces the same winner
A factual question may converge on one answer. A subjective local recommendation can vary even across repeated runs of the same model. Across the common panel, all four engines chose the same top business in 4 percent of cases. They chose four different top businesses in 48 percent.
A full split happened 12 times as often as a full match. Sample 491,972 shared local query-market cases.
The examples show the range the aggregate hides. One query has no shared first pick, two have partial agreement, and one is unanimous. The same location sometimes appeared under different names across engines. The matching step resolved those variations before distinct businesses were counted.
Auto sunroof shops near Winnetka
Nineteen businesses appeared in 26 results. Each engine chose a different name first.
Waterproofing near Sedalia
Affordable Foundation Repair ranked first in two engines, second in one, and fourth in one.
Oil changes near Findlay
Valvoline ranked first in three engines and third in the fourth.
Tire shops near Killeen
Discount Tire ranked first in all four, but the lists still contained 22 distinct businesses across 29 results.
Agreement can change from none to unanimous while each list still contains many different businesses. Four observed query examples with name-and-place deduplication.
Across the four lists, a query produced an average of 23 different businesses. Each engine returned about 7 businesses on average, even when asked for 10.
Why candidate pools matter
A companion Google local search surfaced about 37 businesses in a typical market. Each AI engine named about 7, and the union across all four averaged 23. Agreement generally fell as the broader option pool grew, although category patterns did not move in perfect lockstep.
| Market density | Overall | Retail | Financial | Services | Healthcare | Hospitality | Food |
|---|---|---|---|---|---|---|---|
| All markets | 18.3% | 25.6% | 21.8% | 20.4% | 17.0% | 14.1% | 9.6% |
| Rural | 23.7% | 30.2% | 29.0% | 19.4% | 22.7% | 17.0% | 12.9% |
| Low density | 17.3% | 22.4% | 19.7% | 21.1% | 16.8% | 14.2% | 9.9% |
| High density | 14.6% | 18.3% | 17.8% | 20.4% | 13.5% | 11.4% | 7.1% |
| Very high density | 12.5% | 22.4% | 12.5% | 11.3% | 11.9% | 11.0% | 7.5% |
Agreement is highest in rural markets and lowest in very-high-density markets overall. Retail stays higher than food at every density. Percent of cases where at least three engines chose the same top business.
The next table uses the same rows and categories. It counts the different businesses named across all four lists.
| Market density | Overall | Retail | Financial | Services | Healthcare | Hospitality | Food |
|---|---|---|---|---|---|---|---|
| All markets | 23 | 18 | 17 | 27 | 23 | 27 | 30 |
| Rural | 21 | 17 | 16 | 29 | 21 | 25 | 25 |
| Low density | 24 | 18 | 17 | 27 | 24 | 27 | 30 |
| High density | 25 | 19 | 18 | 27 | 24 | 28 | 32 |
| Very high density | 25 | 17 | 20 | 28 | 24 | 28 | 37 |
Food reaches 37 different businesses in very-high-density markets, where agreement is 7.5 percent. Category differences show pool size is one part of the pattern. Average unique business names across the four returned lists.
Each engine cites a different mix of pages
Gemini puts the largest share on brand sites. OpenAI puts the largest share on directories. Anthropic cites review sites far more than the other three. Perplexity is the most even across brand sites and directories. These shares describe sources named in answers, not every source or signal the engines may use.
| Engine | First-party | Local | Directories | Industry | Reviews | Social |
|---|---|---|---|---|---|---|
| Gemini | 41.9% | 11.9% | 28.8% | 12.7% | 1.7% | 2.9% |
| Anthropic | 33.1% | 8.2% | 26.0% | 10.5% | 20.5% | 1.7% |
| Perplexity | 31.6% | 6.4% | 32.5% | 16.6% | 9.4% | 3.6% |
| OpenAI | 26.9% | 8.1% | 40.9% | 20.1% | 1.9% | 2.1% |
No single page type carries the same share across all four engines. Reported share values by engine and citation class. Totals vary by one tenth of a point because of rounding.
Consideration and ranking reward different work
Clear location data is associated with being named. Reviews and ongoing activity help once a business is in the set. Posts, photos, responses, and site structure can support both stages.
The overall and category results support the same pattern. Clear location data helps a brand enter the set, while fresh activity supports its position. Relative lift values from matched profiles.
Rating still shapes the order inside a returned list. It works as one signal among several, not a universal rule.
A high rating is common at the top, yet the highest-rated business in a returned list ranks first only 40 percent of the time. Top-pick rating measures from the matched profile analysis.
Engines can value ratings and still disagree because each engine ranks a different candidate set.
Consensus profiles show depth across many signals
Unanimous top picks are tied to a broad web presence. Their own website appears among cited pages in 91 percent of cases. Key directories also appear more often when engines agree.
of unanimous top picks have their own website among the cited pages
The ratios compare how often each directory appears in unanimous answers with split answers. Own-site share and citation-frequency ratios from the consensus analysis.
The matched profiles show the profile gaps behind that web presence.
| Profile signal | Unanimous | Majority | Typical field |
|---|---|---|---|
| Median reviews | 568 | 531 | 19 |
| New reviews per month | 12.2 | 10.4 | 1.9 |
| Posts in the past year | 13.1 | 15.5 | 6.0 |
| Photos | 187 | 258 | 88 |
| Reviews answered | 62% | 55% | 33% |
| Local-business schema | 56% | 48% | 30% |
| Profile completeness | 95% | 93% | 80% |
| Average star rating | 4.2 | 4.2 | 3.7 |
Review depth and recent activity create the largest visible gaps. Star rating moves less than review count, responses, posts, photos, and schema. 15,472 unanimous matched profiles and a typical field of 6.6 million profiles. The majority count was not reported.
Build for consideration, then keep earning rank
Make every location legible
Publish complete location pages, claim the profile, add a clear description, and use local-business schema.
Prioritize the category sources engines cite
Keep broad directory coverage complete, then focus on authoritative sources for each category.
Keep proof fresh
Build review depth, answer reviews, add photos, and post often enough that each location looks active.
A local scan can come later. Search your category in several engines and record which businesses each one considers.
The next tests should show the missing counts
- Publish the case count behind each category and density cell, and define each density band.
- Report citation counts, source-class rules, model versions, and the collection date for each engine.
- Show the lift model, comparison group, controls, and confidence range for every profile signal.
- Break the lift and benchmark tables out by Retail, Financial, Services, Healthcare, Hospitality, and Food.
- Track the same cases over time. Show how pools and agreement change after model updates.
Questions readers ask
How often do four engines recommend the same local business+
Why do the engines disagree+
Does a larger market always reduce agreement+
Do ratings decide which business ranks first+
Which profile fields are linked with being named+
How the study was measured
- The panel
- 3,569 distinct local keywords were tested from February through August 2026. The broad run covered 2.05 million local questions. The common comparison includes 491,972 query-market cases asked of all four engines. Cases with fewer than four engines were excluded from the consensus comparison.
- The prompt
- The core panel used unbranded local recommendation prompts in the form “What are the top 10 [category] near [neighborhood]?” The same prompt and place were tested across all four engines.
- The engines
- Google Gemini, OpenAI, Perplexity, and Anthropic were included. The study treated the first-ranked business as the top choice.
- The market field
- The analysis reports 124,820 market observations per month. Results are grouped into Rural, Low density, High density, and Very high density markets.
- The categories
- Retail, Financial, Services, Healthcare, Hospitality, and Food are reported separately in the market tables.
- Companion local search
- A traditional Google local search estimated the broader local field and matched businesses named in AI answers to public profile attributes. Name and place resolved variations in how the same business appeared across engines.
- Agreement
- Unanimity means all four engines ranked the same business first. The heatmap measures cases where at least three engines ranked the same business first.
- Profile analysis
- Ratings, reviews, review responses, posts, photos, schema, profile completeness, descriptions, attributes, and location pages were evaluated for matched profiles.
- Benchmark cohorts
- The unanimous profile table uses 15,472 matched profiles. The typical field contains 6.6 million profiles. Query-level agreement and profile-level benchmarks use different analysis grains.