In Catalonia a quiet technical fault in the foundations of AI search is doing political and commercial work for us. What looks at first like translation noise turns out to be a systemic retrieval failure with real consequences for who is counted as an authority, who finds customers, and how whole communities see their own culture reflected online. This investigation is based on reporting by David Carrasco Pamies for Search Engine Land. See the original piece for screenshots and the full experiment. (Search Engine Land link: https://searchengineland.com/multilingual-regions-ai-search-future-478282 and author David Carrasco Pamies.)
Catalonia as a stress test for AI search
Catalonia is an unusually clean lens on the problem because two languages share a single geography. That makes it easier to see when an algorithm is treating language as a stand-in for market or jurisdiction rather than as a neutral signal about user intent. From a Barcelona IP address, the same queries in Catalan and Spanish produced markedly different pools of sources in Google AI Overviews and in ChatGPT, not merely stylistic differences but different facts, priorities, and cited authorities.
The mismatch is not hypothetical. Google acknowledged that Catalan search results had been downgraded in January 2023, after widespread complaints. Reporting on that acknowledgement is available at Catalan News. The issue reappears now in the synthesis layer that AI-overviews and chat interfaces use. In other words, an old pipeline problem is being amplified by new layers that generate answers in natural language.
How the test was run
The experiment paired Catalan and Spanish queries from a residential Barcelona IP and compared results across two surfaces: ChatGPT (fresh, logged out sessions) and Google web search when an AI Overview was generated. The queries targeted five intent types: a politically charged factual query about independence, a transactional local-business query, a cultural traditions query (Sant Jordi), a hyper-local regulatory query, and deliberately ambiguous Catalan inputs to test language identification. Sessions were repeated to check stability. The full methodology and example queries are documented in the Search Engine Land report.
Four persistent patterns
1. Vocabulary and source plurality diverge
When the same political question is asked in Catalan and Spanish, the outputs are not just phrased differently. The retrieval pools differ. Spanish answers tended toward legalistic framings anchored in the 1978 Constitution and sources such as BBC and Spain oriented outlets. Catalan answers surfaced regional outlets like El Punt Avui and VilaWeb and Catalan Wikipedia. This is not mere translation bias. It is a language-conditioned filter that creates parallel information ecosystems inside the same city.
2. Commercial retrieval shifts, and the engine doubts the minority language
A commercial search for accountants in Barcelona returned different commercial universes depending on whether the query was Catalan or Spanish. More tellingly, Google autocorrected a Catalan commercial query into something absurdly different and offered zero paid ads against the Catalan results. That absence of SEM bidding becomes a feedback signal: models trained on engagement interpret the lack of commercial activity as evidence that Catalan queries are not commercially serious, and they weight retrieval accordingly.
3. Cultural authority gets reassigned
The Sant Jordi examples show how the system assigns custodianship of a tradition based on language. Catalan queries tend to cite regional and municipal authorities; Spanish queries often cite national tourism portals or commercial sites. In consecutive sessions the exact institutional citations switched, but the structural pattern held: language determined who the system considered the owner of the answer.
4. Language identification was already broken before LLMs touched it
The experiments expose a brittle language-identification layer. Google Translate once marked the Catalan phrase Tradicions de Sant Jordi as Occitan. Google Search has suggested showing results in Spanish for queries written in clean Catalan about Catalan-only subjects, such as receptes de calçots. Crucially, the behaviour is inconsistent across sessions, which makes the problem harder for content owners to diagnose and fix. The irregularity is what breaks predictable optimisation.
Why this matters beyond Catalonia
Catalonia is a leading indicator because its geography removes some confounds. The underlying mechanisms are general: corpus weight, training-data composition, and semantic collapse in embeddings. Where one variant has much more data, models default to that variant. The same effect will show up within nominally monolingual countries when sub-national regulations, civic practice, or market behaviour differ between regions. Think of California versus Texas on data privacy, or Quebec versus Ontario on provincial law. As synthesis replaces link-first answers, those defaults become consequential.
The doom loop for minority languages
There is a self-reinforcing cycle. Models that perform poorly on a language produce low-quality generated content in that language. That content gets indexed and then feeds into the next training set. Studies and reporting have already documented significant proportions of newly created content that is machine generated on platforms such as Wikipedia and in vulnerable-language editions. See the MIT Technology Review coverage of the so called doom loop in vulnerable-language Wikipedias and related research. When high-signal local sources are scarce, generated noise becomes a larger fraction of the available corpus and the cycle accelerates.
What institutions are doing
Faced with this risk, some institutions are moving to protect local knowledge. The English Wikipedia community voted to prohibit LLM-generated article content on March 20, 2026. Regional projects and government-backed initiatives in Catalonia and the Basque Country are funding local models and language-resource projects to assert linguistic sovereignty. Those moves are precautionary and pragmatic: when global models underperform for a language, the only durable fix may be to build local capacity.
Practical steps for brands and local publishers
This is not just an academic fault. It is a commercial problem you can and should mitigate now.
– Treat sub-national jurisdictions as distinct products. Create state, regional, or municipal variants that carry explicit authority signals rather than relying on a national parent page.
– Encode jurisdiction in structured data. Use Schema.org areaServed at the granularity that matters and mark up regulator names, license numbers, and region-specific identifiers.
– Ground entities in knowledge graphs. Populate Wikidata with explicit jurisdictional and language properties (for example P1001 for jurisdiction) so AI retrieval can find deterministic hooks outside on-site copy.
– Audit with the model as the user. Run the same diagnostic prompts you would for international markets but applied to the jurisdictions you care about. If an AI overview collapses the target jurisdiction into a larger neighbour, you have a fragmentation problem.
– Watch market signals. Low SEM spend in a regional language is itself a signal the model can use. Invest where it matters: entity pages, local citations, and paid visibility can reintroduce a commercial signal the retrieval stack respects.
What this means for policy and industry
This is a systems problem that requires both technical fixes and market discipline. Engineers need to repair language-identification layers and make retrieval pipelines sensitive to jurisdictional markers. Publishers and brands need to stop assuming that language alone will be respected as a market signal. And regulators and civil society should take seriously the risk that automated content generation can hollow out minority-language information ecosystems.
The Catalan case should be a wake-up call, not a curiosity. It exposes the assumption that language equals market and that a query string reliably signals jurisdictional intent. Neither assumption holds. The result is a retrieval architecture that privileges larger, well represented corpora and pushes minority-language realities toward a default majority narrative. For businesses and institutions that care about local authority and accurate representation, the task is now to make those jurisdictional signals explicit and hard for AI systems to ignore.
Sources and further reading
This dispatch is based on reporting by David Carrasco Pamies for Search Engine Land. Read the full report at Search Engine Land: https://searchengineland.com/multilingual-regions-ai-search-future-478282. See Catalan News on Google Search history with Catalan results: https://www.catalannews.com/culture/item/google-acknowledges-issue-showing-fewer-catalan-language-links-on-results-page. For broader context on vulnerable-language content on Wikipedia and the risks of automated generation, see MIT Technology Review coverage and the academic literature referenced in the Search Engine Land piece.
