What a ChatGPT-style AI assistant recommends to customers in Riga
We asked an OpenAI model with web search real customer questions about 8 kinds of local business in Riga and recorded every answer: what the AI searched for, on which sites, which businesses it named and what it said about them. 902 answers, 229 questions in total.
902 answers · 229 questions · 8 niches · 20–25 October 2026
By: localup.lvPublished:
Key findings
902 of 902 answers started with a web search — 3.3 searches per answer on average. The AI does not answer from memory; it answers from what it finds.
50% of the AI's searches were in English or mixed when the customer asked in Russian (759 of 1522); for Latvian questions, 462 of 1419 (33%).
38% of searches looked for a business's official website (1150 of 2989).
69% of answers had directory or business-list pages in their search results (618 of 902); review and rating sites: 171 of 902 (19%).
96% of the links to businesses the AI named went to the business's own website, not to a directory (1860 of 1937).
2.2 businesses named per answer on average; 281 of 902 (31%) of answers named three or more.
73% of mentions in "padel clubs" went to the three most-named businesses; in "lawyers", 17%.
72% of answers to the same question on another day named that question's most-named business again (average over 103 questions). Answers change.
20% of answers quoted a price (182 of 902); a rating: 100 of 902 (11%); opening hours: 91 of 902 (10%). For vets, 53 of 108 (49%) of answers gave opening hours.
27% of answers had a link to a source in the text (242 of 902). In the rest the AI used search results without showing them.
Every number comes from the published dataset; method and limits below.
The customer asks in Russian — the AI searches in English
Each number below is the share of searches the AI ran while answering questions in that language (language detected from the words of the search).
| English | 589 of 1522 (39%) |
|---|---|
| Mixed (English + another language) | 170 of 1522 (11%) |
| Russian | 496 of 1522 (33%) |
| Latvian | 267 of 1522 (18%) |
| English | 306 of 1419 (22%) |
|---|---|
| Mixed (English + another language) | 156 of 1419 (11%) |
| Latvian | 953 of 1419 (67%) |
| Russian | 2 of 1419 (0%) |
The search language is not the answer language: we asked for the answer in the language of the question. 433 of 2989 (14%) of searches were restricted to one site (site:).
Directories find you, your website convinces
Which kinds of sites appeared in the AI's search results (10.1 different sites per answer on average). One answer can count in several rows.
| Websites of the businesses the AI named | 893 of 902 (99%) |
|---|---|
| Directories and catalogs | 618 of 902 (69%) |
| Booking, delivery and service platforms | 345 of 902 (38%) |
| Media, blogs and guides | 295 of 902 (33%) |
| Review and rating sites | 171 of 902 (19%) |
| Social networks | 28 of 902 (3%) |
| Map place pages (Google Maps etc.) | 16 of 902 (2%) |
Unrelated pages (forums, encyclopedias, scientific papers, PDF documents, search pages) were in the results of 896 of 902 (99%) of answers; we count them in no row.
Platforms the AI saw most often
Directories, review sites, platforms and media the AI's search returned most often (businesses' own websites are not named).
| viss.lv · Directories and catalogs | 161 of 902 (18%) |
|---|---|
| 1188.lv · Directories and catalogs | 144 of 902 (16%) |
| ss.lv · Directories and catalogs | 129 of 902 (14%) |
| zl.lv · Directories and catalogs | 129 of 902 (14%) |
| top-rated.online · Review and rating sites | 99 of 902 (11%) |
| 1189.lv · Directories and catalogs | 97 of 902 (11%) |
| fresha.com · Booking, delivery and service platforms | 97 of 902 (11%) |
| skydive.lv · Media, blogs and guides | 89 of 902 (10%) |
| getapro.lv · Booking, delivery and service platforms | 72 of 902 (8%) |
| liveriga.com · Media, blogs and guides | 68 of 902 (8%) |
| topklinikas.lv · Directories and catalogs | 61 of 902 (7%) |
| oldtownriga.com · Media, blogs and guides | 57 of 902 (6%) |
By niche: how many businesses the AI names, and how concentrated
Each niche page shows what customers asked, where the AI searched and what to do.
| Niche | Answers | Different businesses | Three most-named (share of mentions) | Russian questions: searches in English or mixed |
|---|---|---|---|---|
| Dentists | 120 | 53 | 81 of 220 (37%) | 66 of 216 (31%) |
| Beauty salons | 120 | 83 | 52 of 281 (19%) | 96 of 175 (55%) |
| Car repair | 120 | 48 | 66 of 248 (27%) | 31 of 167 (19%) |
| Cafés | 108 | 42 | 105 of 240 (44%) | 166 of 196 (85%) |
| Lawyers | 108 | 91 | 40 of 233 (17%) | 57 of 168 (34%) |
| Massage | 98 | 40 | 64 of 240 (27%) | 81 of 151 (54%) |
| Vets | 108 | 35 | 86 of 240 (36%) | 84 of 204 (41%) |
| Padel | 120 | 20 | 209 of 288 (73%) | 178 of 245 (73%) |
Businesses are counted by name, matched to the website the AI gave; name variants without a link may count separately. Business names are not published.
Price, rating, opening hours
Answers in which the AI gave a specific price, rating, review count or opening hours.
| Price | 182 of 902 (20%) |
|---|---|
| Rating (stars or “…/5”) | 100 of 902 (11%) |
| Number of reviews | 88 of 902 (10%) |
| Opening hours or “24 hours” | 91 of 902 (10%) |
How we collected the data
Period and size: 20–25 October 2026, 10 repeated runs of our free check with 8 real Riga businesses from 8 niches. In each run the AI answered customer questions about each business's niche in Latvian and Russian (for cafés also in English). Checks that failed have no answers (4 of 80).
Model: the answers come from the OpenAI model gpt-6-luna with the web_search tool, called through the OpenAI API. This is similar to ChatGPT with search, but not the same as the ChatGPT app, which may use another model and settings plus the user's location, language, memory and personalisation.
The unit is an answer: one question asked in one run. The same question on another day is a new, independent answer and counts again. An answer reused from the cache (same question, word-for-word the same text) counts once: 10 of 912 records were such duplicates, leaving 902 answers. Shares are per answer, so questions asked more often weigh more.
Site type: directories, review sites, platforms, maps and media come from our catalog registry and a hand-checked list; “websites of the businesses the AI named” are domains that some answer gave as a named business's link; an article-shaped page on an unknown site counts as media; everything else is “other sites”. Search language is detected from the words (the same classifier our check uses). Prices, ratings and opening hours are found in the answer text with regular expressions.
What this data does not prove
- This is not a measurement of the ChatGPT app. It is how one OpenAI model behaves through the API on our questions.
- The sample is small: 8 businesses, one per niche, and questions we wrote the way customers ask. Only café questions were also asked in English (14 answers).
- Search results are not the same as “read”: we see what the search returned and what the answer named or cited, not what the model actually weighed.
- Site types and languages are detected automatically; some sites may sit in the wrong group.
- The data covers the period above. AI models and search change, so results at another time may differ.
- We do not publish the names of the test businesses or of the recommended businesses — only the aggregate.
The study by niche
- How a ChatGPT-style AI picks a dentist in RigaDentists · 120 answers
- How a ChatGPT-style AI picks a beauty salon in RigaBeauty salons · 120 answers
- How a ChatGPT-style AI picks a car repair shop in RigaCar repair · 120 answers
- How a ChatGPT-style AI picks a café in RigaCafés · 108 answers
- How a ChatGPT-style AI picks a lawyer in RigaLawyers · 108 answers
- How a ChatGPT-style AI picks a massage studio in RigaMassage · 98 answers
- How a ChatGPT-style AI picks a vet in RigaVets · 108 answers
- How a ChatGPT-style AI picks a padel club in RigaPadel · 120 answers
See what the AI answers about your business
In the free check the AI answers real customer questions about your niche, and you see which businesses it names and why.
No sign-up · Works without a website too