The Rosetta Stone carries one decree, carved three times: hieroglyphic, Demotic, Greek. For most of recorded history the world could read the Greek and nothing else. The decree existed. Its authority existed. But to anyone who couldn’t read the third script, the other two were stone.
That is roughly the position of a US agency running AI visibility for a client that sells into Latin America. The audit exists. The reporting is real. It’s just carved in one script, and the buyer is reading a different one.
The same brand, two different answers
Generative engines don’t retrieve from one fixed index of “the truth about your client.” When a model answers a live question, it resolves that question in the language it was asked. A prompt in Spanish pulls Spanish-language sources, surfaces Spanish-language competitors, and reconstructs the category from a body of content the English prompt never touched.
So the same brand can hold a defensible position in English and functionally not exist in Spanish. Not ranking lower. Not appearing.
This is not a translation problem. Translation assumes there’s one answer that needs to be restated in another language. What actually happens is that two separate answers get generated from two separate retrieval paths, and only one of them is on the agency’s dashboard.
Where the split actually happens
Three mechanisms compound, and none of them show up in an English-only audit.
The entity declaration is monolingual. JSON-LD is how a site tells machines explicitly what it is — which organization, what it does, what it sells — instead of letting the model infer it from visible text. Most sites that have it have it once, on the English tree. A localized /es subtree that inherits the design but not the structured declaration leaves the model inferring the entity from translated marketing copy. Inference is where competitors get substituted for you.
The competitive set flips. The companies a model names in an English answer about, say, industrial logistics software are not the companies it names in Spanish. Regional players with dense Spanish-language content outrank global brands with thin translated pages, because the retrieval pool changed. Agencies benchmark against the English competitive set — the one the client’s US procurement team knows — and never see the regional set that’s actually taking the recommendation.
Nobody’s tooling covers it. Our review of the GEO landscape found 30+ tools operating globally and none built for Spanish-language analysis. Profound, the best-capitalized company in the category at $58.5M raised, operates in English for enterprise accounts. The gap isn’t hidden. It just isn’t on anyone’s roadmap, so agencies inherit a blind spot they didn’t choose and can’t see the edges of.
Why this is worse now than it was as an SEO problem
Multilingual SEO had a forgiving failure mode: bad hreflang meant your Spanish page ranked below where it should, on a results page the user was still going to scroll. The user could recover. The link was there.
Generative answers don’t have a scroll. Pew Research found in 2025 that roughly 1% of users click a link inside an AI response. There is no page two to be buried on. There’s the answer, and there’s everything the answer didn’t mention, and the second category is indistinguishable from not existing.
That’s what makes language a revenue variable rather than a localization checkbox. McKinsey projects $750B in sales flowing through generative engines by 2028. A brand that resolves cleanly in English and ambiguously in Spanish isn’t capturing half of that exposure — it’s capturing the half its agency happens to be measuring.
What we’re not claiming
We’re not publishing a delta number here. We could tell you that Spanish-language answers diverge from English ones by some precise percentage across some sample of domains, and it would be a great headline. We don’t have that dataset yet at a size worth publishing, and a made-up figure would be exactly the kind of thing this article is arguing against.
What we can say is structural, and you can verify it yourself in about twenty minutes, which is the point of the next section.
The four checks, on any client, today
1. Ask the buying question in both languages. Not “what is [client]” — the actual question a buyer asks before a purchase: which vendor does X for a company of Y size in Z country. Run it in English. Run it in Spanish. Compare the named companies, not the sentiment.
2. Check whether the Spanish tree declares itself. View source on /es or the localized domain. Is there JSON-LD, or did the localization pipeline carry the copy and drop the structured data? This takes thirty seconds and fails more often than you’d expect.
3. Confirm the AI crawlers can reach the localized paths. GPTBot, ClaudeBot and PerplexityBot each crawl on their own cadence, and each has to be allowed to. A robots.txt written for the English site, or a localized subdomain that never got the same treatment, produces exactly the invisibility described above — silently, with no error state anywhere.
4. Convert the gap to dollars before the renewal call. A finding stated as “weaker Spanish visibility” gets filed. The same finding stated as monthly revenue exposure, with the traffic, conversion rate and order value that produced it, gets a budget line. Percentages describe. Dollars decide.
The third script
The reason the Rosetta Stone mattered wasn’t that it said something new. It’s that the same authority existed, in parallel, in a script someone could finally read.
Your client’s authority probably already exists. The case studies, the technical depth, the reason a buyer should pick them — it’s all there, carved in English. The question isn’t whether it’s true. It’s whether the engine answering a Spanish-language buyer can read it.
GEO (Generative Engine Optimization) is the discipline of making a brand legible and citable to generative engines like ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. It’s to LLMs what SEO is to search engines — a complement, not a replacement: SEO positions you in search results, GEO determines whether a model names you in the answer. Where SEO measures rankings and clicks, GEO measures citation and, at Clicon, the revenue exposed when that citation goes to someone else.
Sources: Pew Research Center (2025), click-through behavior on AI-generated responses. McKinsey, generative commerce projection to 2028. Clicon competitive review of the GEO tooling landscape (2026); Profound funding per public reporting.
We run Lotus on your client’s real domain, show you the bleed in USD, and hand you the artifacts ready to deploy. In both languages. No third script required.