Which brands does artificial intelligence recommend in Uruguay?

How the verticals and the questions were chosen. The 35 verticals are the ones an average Uruguayan consumer has to decide on at some point in their life: where to study, which bank to use, which health plan to choose, where to shop, whom to hire. Purely industrial or niche verticals, where the decision isn't made by an ordinary person asking an AI, were left out.

The prompts aim to reproduce real questions from people who want to know which option is better, the kind found on forums and social media: “which university is best for engineering?”, “which health plan suits me if I have young kids?”. A first draft of questions per vertical was generated with a language model from the six defined intents; I then reviewed them one by one, discarded the ones that didn't sound like a real person, corrected the vocabulary to Rioplatense usage and checked that no ranking prompt named a brand. The final bank of 692 questions is the result of that manual curation.

The prompt bank. 692 prompts across 35 verticals (about 20 per vertical), written the way a real consumer would ask them: “What's the best health plan in Montevideo?”, “Which courier should I use to bring purchases from the US?”, “Where should I study engineering in Uruguay?”. Each vertical combines six intents: open recommendation, attribute, use case, decision, geographic and informational. 90% of prompts are in Rioplatense Spanish and 10% in English, to measure the language difference.

Distribution by vertical. The six intents add up to the 631 brand-free prompts share of voice is computed over; the “Name a brand” column is the 61 head-to-head prompts that stay out.

One example prompt per vertical, brand-free, taken from the bank:

How a mention is counted. In the August and September cuts, an extraction variant counted some brands more than once within the same answer. The effect reaches 4.0 share points in the published top 5s. In 6 categories with a statistical tie, the brand listed first changes. From the October cut on, each brand counts once per answer. I publish the September cut recalculated with that rule in the next version of the study. A response that names five brands adds one mention to each of the five. A brand's share in a vertical is its mention total divided by the sum of every verified brand's mentions in that vertical. That is why a response naming many brands weighs more in the vertical's total than one naming a single brand.

Run-to-run stability. Each prompt runs three times per model. For each vertical, the most-mentioned brand in run 1, run 2 and run 3 is taken, and the vertical counts as stable if all three coincide. With the three models pooled, 18 of 35 verticals coincide (51.4%); by model, ChatGPT 42.9%, Gemini 45.7% and Google AI Mode 42.9%. Of the 11 verticals with a clear lead (see “Margin of error”), 11 coincide across all three runs. The 17 verticals where the most-mentioned brand changes between runs are all technical ties.

Margin of error. For each vertical, a 95% confidence interval (Wilson method) is computed on the leader's share of the vertical's total verified mentions; if the runner-up's share falls inside that interval, the difference isn't statistically significant. The interval treats each mention as independent, even though each prompt runs three times and a single response can name several brands; the true interval is therefore somewhat wider than the computed one, and any error runs toward counting fewer ties, not more. In 11 of 35 verticals the leader has a clear lead; in the other 24, at least two brands are in a technical tie.

The rule that avoids mention bias. 631 prompts name no brand, and share of voice is only calculated over those. The remaining 61 are head-to-head comparisons (“ORT or Universidad de Montevideo?”) analyzed separately and never added to the rankings: no brand gains share for having been named in the question itself.

The measurement. Each prompt runs 3 times per model per corte against ChatGPT, Gemini and Google AI Mode, queried via API against the real product, with active search when the model uses it, with no login or history: the default answer any person would get. The study measures brands' current representation in the AI layer as of the corte date, not the training corpus and not consumer preference. I set Uruguay as the search location for every query. I send each prompt in its own language, Spanish or English.

Capture per model. Since August 31, 2026, I collect responses through DataForSEO, an independent data provider, so that the capture can be audited by third parties and does not depend on AuraMetrics' own infrastructure. The provider does not expose the exact version of the model queried. Dates and times in UTC; the study's corte dates are defined in UTC.

Brand verification. Every mention is checked against a hand-reviewed registry of over six thousand brands and aliases, with whole-word matching and per-brand aliases (for example, “Banco República”, “BROU” and “brou.com.uy” count as the same entity). Nothing classified automatically without human review can appear in a published ranking. Verticals with fewer than 100 verified mentions aren't published as a ranking.

The gap and the residual: two different numbers. The gap is a direct per-response classification: does the text cite any verified brand, Uruguayan or not? The extraction residual is a different thing: the share of responses whose text contains some string from the brand registry but in which no brand was attributed. In corte 3 it averages 12.6 points, with a low of 0.6 in HR & headhunters and a high of 34.5 in Tourism & hospitality. That residual mixes two things the automatic string search cannot tell apart: spurious matches (common words that coincide with an alias) and real mentions the extractor did not capture. That is why every published gap figure should be read as an upper bound. From corte 3 on, the residual is audited with a manual sample of 100 responses, classified by the author; the result is published in this section. The detail per vertical:

Declared limitations. The study currently has two points in time, a week apart: variation between cortes is reported as an observation, not a trend. The English segment is small (2 to 3 prompts per vertical) and is presented only as a presence comparison, never as a ranking. In the most fragmented verticals (dining, tourism), the #1 spot holds under 7% share: the finding there is the fragmentation, not the leader. The cadence is monthly; the series will show which moves stick.

Note on the first cut. The first cut is dated August 24, 2026. In that cut, 83% of the ChatGPT and Gemini answers came through the direct OpenAI and Google APIs. The models were gpt-4o-mini and gemini-3.1-flash-lite. Those calls carry no search location. The rest of that cut, and every Google AI Mode answer, was captured through DataForSEO. I did not keep the parameters of those tasks. I cannot verify which location they were sent with. The phrase “against the real product” describes the capture from August 31, 2026 onward. From that date, every query goes through DataForSEO, set to Uruguay and in the prompt's language. I keep the first cut published as it was measured. Every comparison with that cut carries this caveat.

By design, the bank includes 31 prompts (4.9% of the brand-free prompts) that explicitly ask for a Uruguayan brand or company, concentrated in consumer packaged goods (10), b2B software & cybersecurity (5) and apparel (4). The remaining 600 are general questions that let the answer be local or foreign. In those three verticals the leadership of Uruguayan brands is favoured by the wording of part of the questions; the breakdown is the “Ask for a Uruguayan brand” column of the distribution table. The October 2026 cut ran the full v1 bank, 692 prompts. Each of those 31 prompts has an approved neutral pair, with the same vertical, intent and language. The neutral pairs run from the November 2026 cut on. I measure them separately. They do not count toward the ranking. The comparison between each prompt with a nationality adjective and its neutral pair starts with that cut.

Correction note. Correction (September 2026): the “vs previous cut” variations published between September 4 and 9 were calculated against a partial August extraction. I recalculated them against the complete August cut. The rankings of each cut remain the same.

September cut note. In the September cut, the ChatGPT podium for consumer packaged goods includes Reddit in second place, with 14 mentions. Reddit is a source ChatGPT cited in its answers. Starting with the October cut it is classified as a source and stays out of the brand rankings. The top 5 for the category and its leader remain the same.

Replicability. The full methodology is published on this page, with one example prompt per vertical in the distribution table. The complete prompt bank is available to academic researchers on request, with a citation commitment.