For a local owner, there is no single AI ranking to chase. In this study, Google Gemini 3.8 Flash and OpenAI GPT-6 Luna often recommended different businesses for the same question. A useful response is to keep business facts accurate wherever customers and assistants find them.
The answers give a view of what a customer might receive when asking for a dentist, salon, gym, restaurant or clinic. They do not measure bookings or prove why a business was selected. The ChatGPT prompts for your Google Business Profile show how to turn that uncertainty into practical work.
What this study asked
Reach ran the study on 2026-10-08. It sent 20 customer-style questions to each assistant through an API: Google Gemini 3.8 Flash and OpenAI GPT-6 Luna with low reasoning. These were API model responses, not a test of the consumer apps. There were 40 answers in total.
The questions covered 10 Indian cities: Mumbai, Delhi, Bengaluru, Hyderabad, Chennai, Kolkata, Pune, Ahmedabad, Jaipur and Lucknow. They covered 5 business types, with 13 English prompts and 7 Hinglish prompts typed in Roman script. Some asked for a shortlist; others asked which place to choose.
Gemini lists; Luna picks
Gemini named 131 businesses across its answers; Luna named 65. That is 6.6 names per Gemini answer and 3.3 per Luna answer, rounded to a decimal place. A longer list gives a customer more options to check. A shorter answer can make the recommendation feel more decisive, even when the evidence is incomplete.
| Measure | Gemini 3.8 Flash | GPT-6 Luna |
|---|---|---|
| Answers | Gemini 3.8 Flash20 | GPT-6 Luna20 |
| Business name mentions | Gemini 3.8 Flash131 | GPT-6 Luna65 |
| Mean names per answer | Gemini 3.8 Flash6.6 | GPT-6 Luna3.3 |
| Answers with web search | Gemini 3.8 Flash13/20 | GPT-6 Luna20/20 |
This difference does not establish that either assistant is better at finding local businesses. The prompts, model behavior and search choices all affect the output. A mention is a recommendation to investigate, not a verified assessment of service quality.
The assistants mostly chose different names
Only 21 of Luna’s 65 names also appeared in Gemini’s answer to the same prompt: 32%, rounded to a whole percentage. In 10 of the 20 prompts, the assistants shared no names. Strict name matching gives 20 shared mentions; the reported 21 includes a shortened-name match identified in the study analysis. This is directional overlap against Luna’s list, not an average agreement score.
Shared names were concentrated in familiar food destinations. Kolkata family-dinner answers shared 4 names, Hyderabad biryani answers shared 3, and Ahmedabad Gujarati-thali answers shared 2. That pattern is an observation from these questions; the study did not measure fame or prove that famous businesses have a general advantage.
If an owner appears in one answer and is absent from another, that alone cannot show a rise or fall in visibility. Save the exact question, model, date and answer before comparing results. The Gemini vs ChatGPT vs Claude guide covers the broader assistant choices; the models in this study were specifically Gemini 3.8 Flash and GPT-6 Luna.
Chain share depends on the business type
Gemini’s salon and gym answers leaned more toward chains than its dentist and restaurant answers. Luna’s gym answers also had a higher chain share than its dentist and restaurant answers; its salon and restaurant shares were equal. The assistants did not show a uniform preference across categories.
| Business type | Gemini 3.8 Flash | GPT-6 Luna |
|---|---|---|
| Salons | Gemini 3.8 Flash16/25 | GPT-6 Luna3/12 |
| Gyms | Gemini 3.8 Flash12/21 | GPT-6 Luna5/12 |
| Clinics | Gemini 3.8 Flash8/20 | GPT-6 Luna2/8 |
| Restaurants | Gemini 3.8 Flash5/39 | GPT-6 Luna5/20 |
| Dentists | Gemini 3.8 Flash1/26 | GPT-6 Luna0/13 |
The chain label is a judgment call. The study treated brands operating across cities or owned by national groups as chains, with borderline cases decided individually and not independently verified. The table preserves those labels. It cannot establish that chain ownership caused a recommendation or that every remaining business is an independent.
Luna hedged more and cited more answers
Luna included a citation in 20/20 answers; Gemini did so in 13/20. Luna had 40 citation links in total, compared with Gemini’s 95. Thus, Luna cited sources more consistently across answers, while Gemini supplied more links overall. All Gemini citation links used Google redirects, so their underlying source domains were not visible in the study files.
The supplied source analysis identifies Practo as Luna’s most cited domain. Other cited sources included Pagenearby, Tripadvisor, District, Top Rated Online and Justdial, alongside business websites. These are places an owner can check for outdated contact details and conflicting business information. A citation identifies a source used in an answer; it does not validate every claim on that page.
Approximate keyword flags marked caveat language in 14/20 Luna answers and 9/20 Gemini answers. Rating, review or listing signals appeared in 14/20 Luna answers and 7/20 Gemini answers. This supports checking those public signals. It does not reveal their weight in a model’s decision or prove that changing them will earn a recommendation.
Hinglish showed a different pattern
Luna named 15 businesses across the 7 Hinglish questions, a mean of 2.1 per answer. Across the 13 English questions, it named 50, a mean of 3.8. Overlap with Gemini was 3/15 names for Hinglish (20%) and 18/50 for English (36%). These percentages use Luna’s names as the denominator.
Gemini went in the other direction: it named 50 businesses across Hinglish answers, a mean of 7.1, and 81 across English answers, a mean of 6.2. Fewer names was therefore a Luna pattern, not a shared assistant pattern.
Only 7 Hinglish prompts were asked, and they were different questions from the English set, not paired translations. City, business type and wording changed too. The result is indicative only; it cannot isolate a language effect. An owner can try the phrasing customers use, while treating a change in the answer as a prompt-specific observation.
What a local owner can control
The answers mixed live search and model knowledge. The visible citations and review references suggest a practical place to start: make public business information useful and consistent. This is an owner workflow, not a formula for AI placement.
- Check the Google Business Profile checklist: use the correct category, current opening hours, accurate contact details and a clear description of real services. A complete profile helps customers verify the recommendation they receive.
- Read recent reviews, fix recurring service problems and reply to Google reviews with useful context. The answers referenced reviews, but this study did not test whether replies caused selection.
- Fix business information across listings. For a clinic, check the relevant Practo page; for a restaurant, check the review or discovery sites customers use. Align the name, address, phone, hours and website with the business’s actual details. Keep the business website accurate too.
A dentist owner checks the recommendation path
The Mumbai dentist prompt produced 6 Gemini names and 5 Luna names, with no shared names under the study’s matching rule. An owner seeing those answers should check both sets of sources before treating either as a competitive ranking.
Start with the clinic’s own facts: category, treatment information, opening hours and booking contact. Compare the Google profile, website and relevant listing page. Correct a mismatch, then record the question and response when checking again. A later mention may be useful evidence, but a single changed answer cannot prove what caused it.
Keep that check separate from Google Maps rank tracking. These assistant answers were not Maps rankings, and the study did not compare their names with a Maps result set.
Limits before acting on the findings
This was one run of 20 prompts per model on 2026-10-08. There were no repeat runs to measure variation. Search use was model-dependent, and names, chain labels and keyword flags involved extraction or judgment. Different name spellings can hide overlap; the overlap rule was not a verified business-identity match.
Some names could not be cross-checked against an independent source in this dataset. That does not make them wrong. Business existence, service quality, locations, ratings and claims were not independently verified. A citation also cannot substitute for checking the current details before booking.
The study did not change a profile or listing and measure the effect. It cannot promise that profile completeness, reviews, replies or consistent listings will cause an AI recommendation. Those actions help customers check facts and give search systems clearer information to use.
Start with the profile facts
Use the profile audit to find missing or inconsistent information, then work through the relevant guides. Keep a separate record of assistant answers if monitoring them. Reach can help with profile and review work; this study gives no basis for promising a particular AI recommendation.