How AI Engines Choose Norwegian vs English Sources

The single biggest factor in whether ChatGPT, Google AI Overviews or Perplexity cites a Norwegian-language source or an English-language one is the language of the query itself, not the topic, and not some fixed national default. But "it depends on the query language" is only the headline; the actual citation rates vary enormously by engine, ranging from roughly 85% local-language citation on Google AI Overviews down to roughly 52% on Grok for equivalent non-English prompts, according to the largest published multilingual citation study available (Temso AI, "Lost in Translation," April 2026, 7,058,891 citations across 350,000 responses, four models, six non-English languages, twelve countries).

No study in that dataset or any other currently tests Norwegian directly. The closest available proxy is Swedish, a fellow low-resource Germanic Nordic language, and this article is explicit about that substitution everywhere it applies, rather than presenting Swedish findings as if they were Norwegian ones.

Why Does English Even Compete With Bokmål in the First Place?

Because Norway's English proficiency is exceptionally high by global standards, which is precisely what makes the language question interesting rather than settled. Norway ranks 5th globally on the 2025 EF English Proficiency Index, with a score of 613 against a global average of 488. Norwegian users are genuinely capable of searching, and being satisfied with an answer, in either language. That capability is exactly why source-language selection in Norway is a live strategic question rather than an assumption: high English proficiency makes English content commercially viable, but it does not, on its own, prove that AI engines default to English sources for Norwegian queries. The two claims are frequently conflated in casual market commentary, and this article treats them as separate throughout.

What Actually Determines Whether an AI Engine Cites a Norwegian or English Source?

Temso AI's study found that the dominant driver of source-language selection in AI search is the language the question is asked in, not the topic of the question. Ask an engine something in Norwegian, and it is measurably more likely to cite a Norwegian-language source than if you ask the identical question in English, though the size of that effect differs hugely by platform.

Using Swedish as the nearest proxy, local-language citation rates for non-English prompts were: Google AI Overviews 85.1%, Microsoft Copilot 60.4%, ChatGPT 57.4%, and Grok 47.1% (compared with 43.7% for English on the same Grok queries, essentially near parity). Across all six tested non-English languages combined, the spread was Google AI Overviews 85.4%, Copilot 76.7%, ChatGPT 70.2%, Grok 51.7%. Temso's own explanation for why Germanic languages including Swedish and Dutch underperform some Romance languages in this dataset: "Spanish and French have substantially larger volumes of web content than Swedish or Dutch," a volume effect rather than a linguistic one.

Why Swedish Is the Best Available Proxy, and Where That Proxy Breaks Down

Sweden and Norway share close linguistic ancestry, comparable population size, similarly high English proficiency, and comparably thin web corpora relative to English. But a proxy is not a substitute for direct evidence, and two independent research efforts, working separately, reached the same conclusion: no controlled Norwegian-language citation study currently exists. A defensible Norway-specific GEO programme should include a first-party bilingual retrieval test (running fifty to a hundred semantically equivalent Bokmål and English prompts across the major engines) before making firm claims about Norwegian-specific citation behaviour, rather than treating the Swedish proxy as final.

One piece of Norway-specific evidence does exist, even without a formal citation study: a 2026 comparison of ChatGPT Search and Perplexity for Norwegian-language queries, published by the Norwegian outlet aivett.no, found that ChatGPT Search performs "somewhat better on Norwegian and Scandinavian content," with Norwegian news sources such as NRK, Aftenposten and Dagbladet more consistently represented in its results. Perplexity, by contrast, "prioritises English-language sources even for Norwegian queries," producing answers the reviewer described as internationally oriented rather than locally relevant. That single practitioner comparison points in the same direction as the Swedish proxy data (ChatGPT more locally weighted than some competing engines), which is a useful, if informal, form of corroboration.

Why Does Platform Choice Matter as Much as Language Choice?

Because the AI-assistant market itself is fragmenting quickly, which means optimising for one engine's citation behaviour increasingly means optimising for a shrinking share of actual usage. A global tracker (Similarweb GenAI, web traffic only) shows ChatGPT's overall share falling from roughly 86.7% in early 2025 to roughly 64.5% by January 2026, with Gemini rising to around 21.5% and Grok and DeepSeek each picking up single-digit shares. A separate, differently-constructed panel focused on B2B referral traffic (Goodie, March-April 2026) found a different but directionally consistent picture: ChatGPT at 62.6%, Claude at 18.5%, Gemini at 10.6%, Perplexity at 7.3%, and Copilot at roughly 4%. These two datasets are not directly comparable, one tracks general web traffic globally, the other tracks B2B referral behaviour specifically, and neither is Norway-specific. What they agree on is the direction: no single engine holds the kind of dominant share that would justify a single-platform GEO strategy, in Norway or anywhere else, and the language-citation differences documented throughout this article compound that fragmentation rather than existing independently of it.

Does Industry Category Change the Answer?

Yes, and the swing is large: a 41-percentage-point spread between the most locally-cited and least locally-cited categories in Temso's dataset. Locally-facing verticals cite local-language sources heavily even for AI queries: K-12 education at 76.9%, accounting and audit at 68.0%, auto repair at 65.3%. Globally-oriented verticals skew back toward English even on local-language prompts: hotels and hospitality at 35.5%, restaurants at 38.9%, higher education at 46.9%.

Category Type	Example Vertical	Local-Language Citation Rate
Locally-anchored	K-12 Education	76.9%
Locally-anchored	Accounting / Audit	68.0%
Locally-anchored	Auto Repair	65.3%
Globally-oriented	Higher Education	46.9%
Globally-oriented	Restaurants	38.9%
Globally-oriented	Hotels / Hospitality	35.5%

For a Norwegian export sector like maritime, energy or aquaculture, this table is a warning against assuming a Bokmål-only content strategy will win citation share. A shipping-technology buyer researching in Norwegian may still land on English-heavy sourcing for the more globally-benchmarked parts of their question, which argues for the bilingual, dual-layer content architecture referenced later in this piece rather than a single-language build in either direction.

Why Is Norwegian Underrepresented in AI Training Data to Begin With?

This is where a structural, not just behavioural, explanation matters. Professor Erik Velldal of the University of Oslo's Department of Informatics has stated plainly that "Norwegian only makes up only 0.1 percent of the language amount ChatGPT is trained on," and that the exact training composition is not fully disclosed. English, by comparison, makes up roughly 49.7% of all websites tracked by Web Technology Surveys (2024) and an estimated 46 to 60% of the text corpora used to train large language models generally. The National Library of Norway's own published research notes there are more than one hundred times as many English Wikipedia pages as Norwegian ones.

Academic research on multilingual citation behaviour, independent of Temso's commercial study, reinforces the same underlying dynamic. MegaWika (arXiv 2307.07049) found that across forty-nine non-English Wikipedias, 48% of citations pointed to English-language web sources and only 33% to same-language documents. A separate AI-news-intermediary study found English Wikipedia to be the single most-cited source globally, across languages. The pattern researchers describe is retrieval-augmented systems re-localising toward the query language at the search layer, while the underlying model's background knowledge remains English-heavy regardless.

Bokmål vs Nynorsk: A Second Layer of the Same Problem

Norway has two official written standards. Bokmål is used by roughly 86 to 87% of the population and primary-school pupils as their primary written language, concentrated in urban and eastern regions; Nynorsk serves a smaller base, roughly 7.5 to 11%, concentrated in western and rural mountain districts, with both standards legally required in public broadcasting and government communication. University of Oslo research evaluating Norwegian question-answering datasets found that most commercial language models perform measurably better on Bokmål than Nynorsk, and continue to struggle with Norwegian-specific commonsense reasoning and factual accuracy in both standards. For a GEO build, Bokmål is normally the correct first target on volume and model-performance grounds; Nynorsk is added deliberately, where a client's own audience geography, public-sector obligations or brand positioning justify it, not by default.

What Do Norway's Own Language Model Programmes Tell Us?

Norway is not treating this gap passively. NorwAI, based at NTNU, and the Language Technology Group (LTG) at the University of Oslo, working with the National Library of Norway, have built a genuine domestic research infrastructure around it.

Does Content Freshness Change Which Language Gets Cited?

Recency is a factor independent of language, and it interacts with the thin-corpus problem in a way worth naming explicitly. AI engines generally weight recently-published, regularly-updated content more heavily in retrieval, consistent with broader GEO measurement findings that a meaningful share of AI bot hits target content published within the past year. For a thin-corpus language like Norwegian, this cuts two ways. On one hand, it means a well-structured, recently-published Bokmål source has a genuine opportunity to outweigh an older, more established English competitor purely because the corpus it is competing against is smaller and more thinly updated. On the other hand, it means the "early-mover advantage" described elsewhere in this research is not permanent; as more Norwegian content gets published and indexed, the relative weight of any single well-structured source declines, and stale Norwegian content ages out of citation relevance at least as fast as English content does.

The practical implication is a maintenance commitment, not just an initial build. A Norwegian entity page or FAQ set published once and left untouched for two years is competing against a growing, increasingly current corpus, even if it was genuinely well-structured at launch. Scheduled content refreshes, at minimum an annual review of core entity and FAQ pages, keep a Norwegian source from quietly losing the thin-corpus advantage it started with.

What Should a Bilingual Content Strategy Actually Do About This?

Four practical implications follow directly from the evidence above, rather than from general bilingual-SEO best practice imported without adaptation.

First, entity definitions and FAQ content should be written in Bokmål first for local-intent queries, then layered into English for export-facing and comparison queries, rather than machine-translated from one into the other; University of Oslo's own findings on model performance gaps make quality of the source language a citability factor in itself.

Second, category matters as much as language. A Norwegian export business (maritime, energy, professional B2B software) should expect its more globally-benchmarked content to be cited from English sources even on Norwegian-language prompts, and should build English-language technical and comparison content deliberately, not as an afterthought to the Bokmål build.

Third, engine choice is not neutral. If ChatGPT genuinely retrieves more Norwegian-source content than Perplexity, as the aivett.no comparison suggests, a Norwegian-audience GEO programme should weight ChatGPT-specific Norwegian content work accordingly, while still tracking Perplexity separately rather than assuming parity across engines.

Fourth, treat the Swedish-proxy data as a working hypothesis, not a settled fact, and commission a direct Norwegian bilingual retrieval test before making firm claims to a client about exactly how their category performs. Our book Cited or Silent walks through how to design and run that kind of first-party citation test across multiple engines and languages as part of a documented 90-day GEO roadmap.

---

Frequently Asked Questions

Do Norwegian AI searches automatically default to English sources?

No, and both Claude's and ChatGPT's cross-validated research on this topic explicitly reject a universal default. Source selection depends on the engine, the query's category, the size of the local-language corpus on that topic, and how recently sources were published. Locally-anchored categories cite Norwegian sources heavily; globally-benchmarked categories skew back toward English even on Norwegian-language prompts.

Is there a citation study specifically for Norwegian, or is all of this really about Swedish?

As of this research, no controlled citation study has tested Norwegian directly. Swedish is used throughout this article as the closest available proxy, given the two languages' shared low-resource status and linguistic proximity, and that substitution is stated explicitly rather than presented as Norwegian-specific data.

Should a Norwegian brand publish in Bokmål, Nynorsk, English, or all three?

Bokmål first, for the large majority of use cases, given its larger population share and stronger model performance in current research. English should be built deliberately for export-facing and internationally-benchmarked categories, not as an afterthought. Nynorsk is added only where a specific audience, public-sector obligation, or brand position justifies it.

Why does ChatGPT seem to handle Norwegian content better than some other engines?

A 2026 Norwegian comparison (aivett.no) found ChatGPT Search "somewhat better on Norwegian and Scandinavian content" than Perplexity, which leans toward English sources even for Norwegian-language queries. This is one practitioner comparison, not a large controlled study, but it points the same direction as the broader (Swedish-proxy) citation data showing ChatGPT with meaningfully higher local-language citation than some competing engines.

This article discusses AI-search research findings and market data; it is not technical implementation guidance for a specific website. Figures reflect publicly available research as of mid-2026.

---