You have people whose job is to know what is being said about your brand in the press and on social media. Does anyone own what AI says about it?
In Mexico and Brazil, generative engines are already recommending brands by name, with judgements attached and direct comparisons to the competition. Every day, to buyers who are making a decision.
It is the only channel where your brand is discussed, and nobody on your team is reading it. It is worth understanding where the answer comes from.
The less material there is, the more each source weighs
These engines build their answers from the written record that exists about a market. In Spanish and Portuguese, that record is far thinner than most people assume.
The Common Crawl report describes the open dataset that feeds much of the training corpus behind large language models. English accounts for 40.6% of pages. Spanish, 4.6%. Portuguese, 2.5%. Two languages spoken by close to 900 million people occupy roughly a fourteenth of the crawled web.
That scarcity is not only a coverage problem. It concentrates risk. Where fifty sources exist about a brand, one outdated article carries little weight. Where four exist, it can become the official description.
Your brand appears in the answer, not in the citations
The source mix changes completely depending on the language of the question. Profound analysis of 3.25 billion citations across seven models and 14 countries, counting each country’s prompts only in its national language.
In Portuguese, AI Overviews concentrated 65% of its social citations on YouTube, Instagram tripled to 17%, and Reddit fell to 7%. In Spanish, TikTok jumped to 16%, five times its English baseline. ChatGPT did the opposite and held Reddit at between 51% and 76% of its social citations in every country measured, reaching 71% in Brazil.
That last figure matters, because Reddit is a near-monolingual American platform. Between 43% and 49% of its web traffic comes from the United States. When a Brazilian buyer asks about a category, much of the social material behind that answer is opinion written in English, mostly by North Americans, about a market that is not theirs.
No brand put it there or approved it. In practice, it functions as a third-party reference in front of a buyer who is evaluating options.
The language AI prefers is not Latin America’s
Engines do favor sources in the language of the query. Weglot research covering more than 1.3 million citations found that, for localized queries in Mexico, 96% of AI Overview citations came from Spanish-language sources.
But the shift towards local domains is the weakest of every language measured. Japanese-language queries sent 26% of citations to Japanese domains; French queries, 16% to .fr domains; and Spanish queries in Spain, just 7% to .es domains. Argentina, in the same test, came in lower still.
In reputation terms, the description of your brand in Spanish is being assembled from sources that are probably not Latin American. With prices from another market, regulation from another country, and cases that do not apply.
One caveat on the source. Weglot sells website translation, so the direction of its conclusion is not a surprise. The magnitudes do match independent academic work on cross-lingual retrieval.
What a communications team can do
- Audit before you argue. Run a fixed set of category questions in Spanish and Portuguese across several engines and from locations inside each market, recording which brands are named and with what adjectives. A query run from New York tells you nothing about what a buyer in Guadalajara sees.
- Treat the corpus like the press kit. If the record about your brand in Portuguese is four documents, the priority is not optimising those four. It is that forty exist. Local media, sector, and institutional sources, and up-to-date entries in reference repositories.
- Correct the record, not the answer. You cannot ask a model for a correction notice. When a description is wrong, the fix happens in the sources that the model draws on. It is slow, and it is the only route available.
- Publish your own data, attributed and dated. Research by Aggarwal and colleagues, presented at KDD 2024, found that adding statistics and authoritative sourcing raised a page’s visibility in generative answers by up to 40%. A figure with a sample, a country and a date is a sentence a model can carry. An adjective is not.
- Add the scenario to the crisis protocol. What do we do if an engine states something false about the brand in front of thousands of buyers, in a language the global team does not read? Better to have the answer before the incident.
What changes in your communications plan
For twenty years, digital reputation was defended in a list of links you could see, measure, and contest.
Now it is condensed into a paragraph written differently every time, in the language of whoever is asking, from a record that in Latin America is thin and mostly written by other people.
Everything starts with finding out what these engines are already saying about you.
How we approach this at Sherlock
We audit what the main engines say about a brand in Spanish and Portuguese, querying from inside each market rather than from a remote office. From there, we work both ends, the owned content in each language and the local coverage that feeds that record.
That discipline has a name. It is called GEO, or generative engine optimization, and it means working so that a brand is described accurately inside the answers AI generates, not only in a list of links.
Our consultants are based in the countries where the question is being asked, and Broadminded, our research team, produces the proprietary data a model can cite. Start with an audit, not a campaign.