We Asked AI 567 Times to Recommend a Local Business. Gemini Named a Real One 79% of the Time — ChatGPT Only 64%.

As people increasingly ask ChatGPT, Gemini and Perplexity for recommendations instead of searching Google, we ran an independent test across 9 local-service categories to see who actually gets named in the answer — and who gets left out.

Why We Wanted to Know

Search is changing. Instead of typing "hair salon Zurich" into Google and scrolling through ten blue links, a growing number of people just ask an AI assistant — and take whatever it says at face value. That is a real shift in how customers find a business, and it raises an obvious question: when someone asks, does the AI actually know your business exists?

Nobody publishes that answer, because no AI engine discloses how often it names a real, specific business versus giving a vague non-answer. So we built a small, independent study to find out — the same kind of measurement our own product runs for individual businesses, run here across 9 categories and 4 cities at once.

How We Tested It

We picked 9 real categories — things like hair salons, yoga studios, aerial yoga, and private tutors — spread across 4 cities, plus one category (online Italian lessons) with no city at all, to see what happens when there's no location to anchor the answer.

For each category, we asked realistic buyer questions — the kind a real customer types, like "best hair salon in Zurich" or "recommend a good tutor for a high school student in Miami" — to three AI engines: Perplexity, ChatGPT, Gemini. Each question was asked 3 times per engine, and every answer was checked for one thing: did it name a real, specific business, or did it name nobody at all?

That gave us 567 individual answers to grade. We also tracked, separately, whenever an answer's first suggestion was a booking platform or marketplace (like a tutoring marketplace or a class-booking app) rather than an actual business — a different, related question about who AI defaults to recommending.

n=3 samples per query — pilot scale, well below the N=12 Beacon uses for a real business scan. More on what that limits in Limitations, below.

What We Found, at a Glance

• Gemini named a real business in 79% of its answers, Perplexity in 69%, and ChatGPT in just 64% — a 15-point gap between the best- and worst-performing engine on the exact same questions (n=189 per engine).
• Across all 567 answers collected for this pilot, between 21% and 36% named zero real businesses at all, depending on which engine was asked — meaning for a meaningful share of buyer questions, a well-run local business was completely invisible to the engine its customer actually used.
• Even within city-bound categories, invisibility ranged from 15% (Beauty in Zurich, n=135) to 50% (Aerial Yoga in Zurich, n=54) — the same three engines, the same question format, a large swing driven by category alone.
• The one category not tied to a city — italian lessons (online) — was platform-mediated in 75% of answers (n=32), versus 0%–46% for the 8 city-bound categories: removing the location anchor appears to hand the category to global platforms and apps rather than named providers.

Does the Engine You Ask Change Whether You Get Named?

Yes — by a wide margin. We asked the exact same 9-category question set to all three engines and counted how often each one named a real business versus naming nobody at all.

Share of answers naming a real business, by engine. n=189 per engine.

Which Categories Get Left Out Most?

Category mattered as much as engine did. The most-invisible category in this study went unanswered more than three times as often as the least-invisible one — same engines, same question format.

Share of answers naming zero real businesses, by category (all 3 engines pooled). Darker = more invisible.

What Happens When There’s No City to Anchor To?

One of our 9 categories — italian lessons (online) — has no city requirement: the buyer is in Zurich, but the provider can be anywhere. We tracked a different signal for this one: how often the AI’s first suggestion was a booking platform or global app rather than a specific named provider.

Share of answers led by a platform or app rather than a named business. Highlighted: italian lessons (online) (location-unbound).

At 75% (n=32), it towers over the 0%–46% range of the 8 city-bound categories. Take the city away, and the category appears to default to global platforms instead of local providers — worth knowing if your business depends on being named ahead of the marketplace that lists you.

This category is reported separately by design — it tests a different question than the 8 city-bound categories, and is never averaged into their combined figures.

The Full Breakdown

What This Means

The headline number here isn't really about which AI is "best." It's that even ChatGPT, the weakest performer in this pilot, still named a real business most of the time — and yet 36% of its answers named nobody at all. That failure is invisible to the business it happens to. There's no notification, no missed-call log. A business simply isn't in the answer, and nobody finds out unless someone checks.

That risk isn't evenly spread. Aerial Yoga in Zurich went unanswered in 50% of the answers we collected for it (n=54) — categories with fewer, less-established local competitors and thinner online listings seem to leave AI engines with less to work with, and less to say.

For a business, the practical takeaway is simple: being good, or even being the obvious local choice, doesn't automatically mean an AI engine will say your name when someone asks. Whether that's worth fixing is a separate question — but it's no longer a hypothetical one. This pilot is a first measurement, not a final one; the Limitations below explain exactly how far it should and shouldn't be trusted.

About This Study

This is a pilot, and we’d rather you know exactly where its edges are than oversell it. Four things to keep in mind:

• Pilot-scale sampling. n=3 per query, well below the N=12 Beacon uses for a real business scan — too few samples to fully average out run-to-run answer variance.
• Single snapshot in time. Data was collected August 5, 2026. AI engines change their answers week to week; this is not a trend line.
• Business names are Claude-extracted from free text, not verified against a real business registry — a mis-parsed or hallucinated name is possible.
• 9 categories is not a representative sample of all local-service categories — this is a pilot, not a census.

How to Cite This Study

For press inquiries, interviews, or to request the underlying data, contact igor@inserviss.app. The methodology here matches Beacon’s own product — Scout samples an engine multiple times per question and records who actually gets named, the same approach used for this study.