Ask ChatGPT to recommend a web design company in the Dominican Republic and it will name eleven. Ask Google's AI the same question, in the same language, on the same afternoon, and it will name five. Between those sixteen companies, exactly one appears on both lists.

We know because we ran it. Over two days in September 2026 we put 26 questions to four AI assistants — ChatGPT, Google AI Mode, Google's AI Overview, and Gemini — twice each, in fresh sessions, in English and Spanish. That's 208 runs. Between them they named 369 distinct Dominican businesses, and we checked every one against its own site, its Google listing, industry registers and local news.

The finding that surprised us most wasn't about any individual business. It was that "how visible is my business in AI" turns out not to be a question with one answer. It has four, and they barely agree.

Zero businesses were named by all four platforms

In the service-sector portion of the study — web design, immigration lawyers, real estate agents, wedding photographers, dive shops, car rental, vets — 117 businesses were named across 64 runs. Here's how they distributed:

• Named by all four platforms: 0

• Named by three: 16

• Named by two: 27

• Named by only one: 74 — that's 63%

Nearly two-thirds of the businesses AI recommended existed on exactly one platform's map of the market. Not ranked lower elsewhere. Absent.

The individual queries are starker than the average:

• Web design: ChatGPT and Google AI Mode named 16 companies between them. They shared one.

• Immigration lawyers: 13 firms. One shared.

• Car rental in Punta Cana: 11 companies. One shared.

• Dive shops in Bayahíbe — a village with maybe a dozen — produced nine names across two platforms. They agreed on two.

The clearest example is our own business

We should disclose something: DR Web Studio, the company that commissioned this research, appears in its own dataset. So does the awkward result.

Across Google's surfaces, DR Web Studio was named ten times — ranked first, second, fourth and fifth depending on the surface and the run. Across five ChatGPT runs covering 36 separate company mentions, in both English and Spanish, it was never named once.

That's not a story about our rankings. It's the cleanest illustration in the study of the point: the same business, the same website, the same week — prominent on one platform's answers and entirely absent from another's. If we had only checked ChatGPT, we'd have concluded we were invisible. If we'd only checked Google, we'd have concluded we were doing well. Both conclusions would have been wrong, because each was a quarter of the picture.

(For completeness, because it matters: the Google fieldwork ran on the account that manages our Business Profile. When we re-ran the English query with personalisation switched off, our position dropped from first to fifth — so part of that ranking was an artefact of being signed in. In Spanish it stayed first under the same control, for a reason we'll come back to.)

Two Google products don't agree with each other

You might expect the divide to run between companies — OpenAI's index versus Google's. It doesn't. Google's own surfaces disagree with each other about as much as they disagree with ChatGPT.

Asked to recommend boutique hotels in the Dominican Republic, Google AI Mode named Tortuga Bay, Amanera, Eden Roc, Casa Colonial and Hodelpa — all large or luxury properties. Gemini, on the identical question, named Natura Cabana, Punta Rucia Lodge, Mahona, Takuma and El Valle Lodge — five small owner-operated properties. Two Google products, one question, not a single business in common.

They don't even behave the same way. Across repeat runs of the same query, Google's AI Overview was frozen — several answers came back byte-for-byte identical. Google AI Mode, same company, same query, same day, replaced half its list between runs. One product is a lookup; the other is regenerating an answer each time.

Meanwhile, underneath both of them, Google's ordinary local pack — the map results that have been there for years — was 100% stable across all eleven repeat captures. Not one entry, rating, review count or quoted snippet changed. For a business trying to measure where it stands, the old map listing is a real position and the AI answer above it is a moving target.

Language changes which system you're in

The sharpest split we found wasn't between platforms at all. It was between languages, on one platform.

Ask Google in English for a web design company here and you get map cards: ratings, review counts, street addresses, Call and Directions buttons. Ask in Spanish and you get something structurally different — pure web citations. No cards. No ratings. No addresses. No local-business layer anywhere in the answer.

Those aren't two versions of one result. They're two different ranking systems. In English, visibility runs through your Google Business Profile. In Spanish, it runs through your website. That's also why our own Spanish position held steady under the personalisation control while the English one moved four places — there was no Business Profile signal in the Spanish answer to personalise in the first place.

Gemini takes this further. Asked in English for a Dominican web designer, it named seven companies. Asked in Spanish, twice, it named zero — and instead produced a genuinely useful buying guide, with local payment gateways and a warning about domain ownership, that routed the reader to no Dominican supplier at all. A well-written, helpful answer containing no route to anyone in the market.

What actually predicts agreement

The platforms do converge sometimes, and the pattern is consistent enough to be useful.

They agree where businesses have both a dense Google Business Profile and a real website. Real estate, wedding photography and dive shops each produced four or five shared names across platforms. In hotels — an unusually well-documented sector, where 83% of the businesses named had their own site — six properties were named by all four surfaces.

They diverge almost completely where the market is fragmented or professional, and where the documentation is thin. Web design, law and car rental produced one shared name each.

So the driver isn't the platform. It's how well the market documents itself. Where businesses are thoroughly described in public — a claimed listing, a working site, consistent naming — the AI layer converges on the same handful. Where they aren't, each platform assembles a different market from whatever fragments it happened to reach, and those fragments differ enormously: ChatGPT read agency websites directly and quoted their published prices; Google read its own business index; Gemini, on the same question, effectively read out a directory page without citing it.

What we'd do with this

Three things, in order.

Check all four, not one. Search your own business name and your main service category on ChatGPT, Gemini, Google AI Mode and an ordinary Google search. It takes twenty minutes. The odds are strong that the results differ more than you expect — 63% of businesses in our data existed on only one.

Check twice, on different days. Run-to-run overlap averaged 0.63 across the study, and one platform swapped half its list between two runs minutes apart. A single check is a snapshot of a system that is not stable. Any number you're told about your "AI visibility" from one run is noise.

Fix the boring layer first. The one surface that didn't move at all was Google's local pack — and that's the one you can directly influence, for free, in about thirty minutes. Below the AI answers, the fundamentals still decide who gets found.

An honest note on what this study is and isn't

This is a snapshot: 208 runs over two days in September 2026, from the Dominican Republic, on one set of accounts. AI systems change. Our sample is curated rather than random — we chose queries a real customer might type, which means it over-represents well-known businesses and makes these findings conservative rather than inflated. We measured what was published, not what it earned; nothing here says anything about revenue.

And one limitation worth stating plainly: the Google runs were made from an account that manages one of the businesses in the dataset. We ran a de-personalisation control, we've reported what it showed, and the cleanest check remains ChatGPT — which has no connection to that account and never named the business at all.

We're publishing the method alongside the findings because anyone can re-run it. That's rather the point.

At DR Web Studio we build the kind of site AI systems can actually read — fast, bilingual, with real text rather than images, and a Google listing wired to it properly. If you'd like us to check what the four assistants currently say about your business, contact us for a free consultation and we'll look and tell you honestly what we find.