One Spanish, Many Markets: AI Visibility in LatAm and Spain

Ask an AI assistant, in Spanish, which suppliers it would recommend in your category. Then ask again, in Spanish, using the word a buyer in Mexico would use rather than the word a buyer in Spain would use. Put the answers side by side. Often they share not a single company name.

That is not a translation error, and it is not the ordinary variance of a probabilistic system. It is the plainest demonstration available that language and market are two different variables, and that any measurement folding them into one reports on a market that does not exist.

Spanish is a language, not a market

Spanish is the working language of a set of markets with very little in common commercially. Different distributors, different regulators, different trade press, different forums where people ask strangers what to buy. A brand can be a default recommendation in Spain and a total unknown in Colombia while the words on the page look identical.

The web a model learned from was not written by "Spanish speakers" in the abstract. It was written in Madrid, in Mexico City, in Buenos Aires, in Bogota, by people describing different markets to different buyers. The language is shared. Almost nothing else is.

The mechanism

A model answers from what it absorbed during training and, on platforms that retrieve live pages, from what it can fetch as you ask. Both pools were produced in specific places, though they sit under one language label. Wording that reads as peninsular Spanish leans on material written for Spain. Mexican wording leans on material written for Mexico. The model is not deciding where you are. It is following the statistical company your words keep.

Once the material shifts, everything downstream shifts with it. The buying guides and trade publications a model reaches for were never one set across the Spanish-speaking world. Neither are the names it offers instead of yours. Ask in one Spanish-speaking market and you may hear about companies that do not operate in the next at all. Not weaker competitors. Absent ones.

The word for your category is not one word

Buyers do not search using the category label you use internally. They use their market's label, and Spanish supplies a different one per market. A computer is an ordenador in Spain and a computadora across much of Latin America. A phone is a movil in one market and a celular in another.

In business categories the divergence sharpens, because the vocabulary is fused to local rules and institutions. Invoicing software in Spain is discussed in the language of Spanish electronic invoicing obligations. In Mexico it is discussed in the language of the CFDI and the tax authority that defines it. Two prompts, both in flawless Spanish, both about billing software, reaching two bodies of text that barely overlap. Accounting, insurance, payroll and payments behave the same way: the category has a local name because it has a local shape.

So if your prompt set uses only one market's vocabulary, you are not measuring Spanish. You are measuring that market and calling it Spanish. The tool reports faithfully. It just does not report what you think.

Why an aggregate Spanish score means almost nothing

Suppose you are well established in Spain and unknown in the Southern Cone. A single Spanish score averages a market where you are named often with one where you are never named, and returns a middling number that describes neither. It is an artefact.

Worse, it moves for reasons unrelated to your visibility. Add a few prompts written in Mexican vocabulary and the score drops, though nothing about your brand has changed. Remove them and it rises. The score describes your prompt set, not the world, and you are reporting it upward as a fact about the world. Split the data by market and it becomes usable at once: strong here, absent there, and a specific gap somebody can be given as a job.

What this does to how you write prompts

• Do not translate from English. A translated prompt carries the assumptions of the market it was written for, and yields grammatical Spanish that nobody would actually type.
• Do not copy one market's prompts to the rest. The prompt set that works for Spain is a Spain prompt set. Reusing it in Argentina measures how Argentine material responds to Spanish phrasing, a question nobody asked.
• Write each market's prompts in that market's real vocabulary. Use the term your local sales team says on calls, the term on local competitors' pages, the term the local trade press prints. If nobody on your side speaks the market, you have just learned something else.
• Keep a small shared set as a control. A handful of prompts phrased identically everywhere gives you one axis where the only variable is the market's material, separating what changed from what merely got reworded.

The competitor list is the payload, not the score

If you could keep one output per market, keep the names. Which companies the assistants recommend in place of yours is the finding that survives contact with reality. A competitor absent from your home market but named everywhere in an export market's answers is not noise in your data. It is the point.

To be precise about what is varying

None of this is geolocation. Nothing here depends on where a user is standing, on an IP address, on maps or on proximity. What varies is the language variant of the question and the market frame it belongs to, and therefore the text the model reaches when it answers. You cannot make a model detect a user's location. You can choose whether you measure with the question your Mexican buyer would ask or the one your Spanish buyer would ask. Those two choices measure you differently.

Spanish is only the clearest case

Spanish is worth working through because the splits are impossible to ignore. The same structure exists anywhere a language spans markets, and it is quieter there, which makes it more dangerous. English spans British, American, Indian and Australian markets, each with its own vocabulary and its own trusted publications. French written for France and French written for Quebec draw on separate material and separate competitive sets. Portuguese in Europe and Portuguese in Brazil are not one market wearing one label. If your English score is really an American score, you have the same problem, and you are less likely to notice, because everything reads correctly.

What separating markets buys you, and what it does not

It does not improve your visibility. Nothing here will make an assistant recommend you more often, and any tool suggesting otherwise is describing a lever that does not exist. What it does is make your visibility legible. An aggregate hides the two facts you needed: where you are present, and where you are simply not in the conversation.

That is what PSentry is built to do: run your prompt set across ChatGPT, Claude, Gemini and Perplexity, score each language and market separately rather than blending them, surface the sources behind the answers, list the competitors named in your place in each market, and export the gaps where you sell but the assistants do not mention you. Two honest limits belong here too. Nothing about the tool promises a different answer in any market: what it gives is a reading of the market as it already speaks, taken twice inside a month rather than continuously, and that reading does not rewrite a single word the assistants produce.

A diagnosis is not a cure. But a company that discovers it has been reporting one Spanish score across markets where its position is completely different has learned something true about itself, and that is more than most brands in this category have.

Frequently Asked Questions

Do I need a separate prompt set for every Spanish-speaking country I sell in?

For every market you sell in seriously, yes. Where you have no operation, no: you get a number you cannot act on. The test is commercial, not linguistic. A distributor, a price list and a target make it a market, and a market deserves its own questions.

My aggregate Spanish score improved. Is that good news?

You cannot tell, which is the problem with the metric. The move could be growth in one market, a decline elsewhere masked by growth here, or a change in which prompts were included. Until the score is split by market, it is not a finding.

Can I just have a native speaker translate my English prompts?

Translation gives you correct Spanish, which is not the Spanish a buyer types. You need the term that market uses for your category, and that is a question for someone who sells there, not someone who speaks the language. Local competitors' pages and the local trade press are the sources.

A competitor appears in one market's answers and nowhere else. Is that a bug?

Almost certainly not. Regional players with strong coverage in their own market's publications routinely dominate the answers there and are absent elsewhere. That asymmetry names who owns the answer you want.

Does the tool detect which country a user is in?

No, and it does not attempt to. There is no geolocation involved. The segmentation is by language and market: the question is written in a given market's language variant, and the model is measured on the answer it gives.

Does any of this apply if I only sell in English?

Yes, in a milder form. If you sell into both British and American markets, one English score averages two competitive fields with different vocabulary and different publications behind them. Spanish just makes the mechanism obvious.