The general, well-documented pattern from natural language processing research is that multilingual language models trained predominantly on high-resource languages, meaning languages with abundant text data available for training, such as English, tend to show measurably reduced understanding and performance on genuinely low-resource languages with comparatively little training data available. This is an established, broadly reproduced finding in academic NLP literature, not a Google-specific secret. But it’s important to be precise about what this means for Google Search specifically today: Google announced BERT’s application to Search back in 2019, and BERT is one historical component of Google’s language-understanding stack, not the current, sole, or most advanced system Google uses for multilingual query understanding. Google has not published specific per-language ranking-quality degradation figures tied to BERT’s training data composition, and treating a 2019-era detail as though it describes Google’s present-day multilingual search quality would be inaccurate.
The general mechanism: transfer learning and low-resource languages
BERT and similar transformer-based language models are typically trained on large text corpora, and in multilingual variants, that training data is drawn across many languages simultaneously, but the volume of available text varies enormously by language. English, along with a handful of other widely-published, widely-digitized languages, has vastly more training text available than many of the world’s languages, which often have comparatively little digitized text, especially in the specific formal or informational registers that make for useful training data. This asymmetry is a well-documented, general phenomenon in NLP research: models exhibit measurably better representation, disambiguation, and understanding capability for high-resource languages than for genuinely low-resource ones, because the underlying statistical patterns a model learns are fundamentally shaped by how much and how varied its training exposure to a given language was.
This is a real, independently verifiable finding from the broader field of computational linguistics and machine learning research, not something specific to Google’s implementation. It shows up across essentially every large language model with multilingual ambitions to varying degrees, and it’s a known, actively studied challenge in NLP research generally, with ongoing academic work specifically focused on improving low-resource language performance through techniques like cross-lingual transfer learning, back-translation-based data augmentation, and targeted low-resource fine-tuning.
Why BERT specifically isn’t the right frame for Google’s current multilingual search quality
Here’s where precision matters most for anyone answering this question about Google Search today. Google’s 2019 announcement about BERT described it as a significant improvement to how Search understands the nuance and context of queries, initially rolled out for English and later extended to more languages. That was a real, meaningful step forward at the time, and the general low-resource-language limitation described above plausibly applied to that 2019-era system to some degree, consistent with the general NLP finding.
But Google’s language-understanding systems have continued to evolve substantially since 2019. Google introduced MUM (Multitask Unified Model) in 2021, which Google has explicitly described as multilingual by design, built specifically to transfer knowledge and understanding across languages more effectively than prior approaches, partly in direct response to the kind of language-representation gap the low-resource problem describes. Google has continued to develop and apply subsequent language and AI systems to Search since then as well. Treating 2019-era BERT as if it were still the primary or sole determinant of Google’s current multilingual query understanding overstates its role in the present system and understates how much Google’s language infrastructure has moved on.
Google has not published specific figures quantifying ranking-quality degradation by language tied to any of these systems’ training data composition, whether for BERT specifically or for its successors. Any claim asserting a specific percentage or measurable ranking-quality gap for a named low-resource language, attributed to BERT’s training data, should be treated as unverifiable extrapolation rather than documented fact, since Google has not disclosed that level of per-language performance data.
A hypothetical scenario to illustrate the practical stakes
Imagine a hypothetical publisher, “Example Regional News,” producing content in a genuinely low-resource language for a specific regional audience. Suppose the site publishes two hypothetical articles on the same topic: one written with clear, explicit structure, plainly labeled headings, unambiguous phrasing, correct hreflang tags pointing to the right language and regional variant, and another written in a looser, more colloquial style with vague headings and no language-targeting markup at all. Hypothetically, if Google’s language-understanding systems have somewhat less refined representation of that low-resource language generally, consistent with the broader NLP pattern described above, the first article’s explicit structure would be expected to give Google’s systems more to work with regardless of any underlying model limitation, while the second article would be relying more heavily on nuanced language interpretation to be understood correctly at all. The scenario illustrates why, for a low-resource-language site, explicit structure and correct language signals matter more, not less, than they might for a high-resource-language competitor.
Practical implication
For practitioners working on multilingual or non-English-language SEO, the honest, current-state answer is a general expectation rather than a documented specific: content in genuinely low-resource languages may face somewhat less refined language understanding from Google’s systems than content in high-resource languages, consistent with the general, well-established NLP pattern, but this should be understood as a directional expectation about the field generally, not a documented, quantified Google-specific penalty tied to a 2019-era model. It’s also reasonable to expect this gap has likely narrowed to some degree as Google’s systems (MUM and beyond) have continued to develop with multilingual capability as an explicit design goal, though Google hasn’t published data confirming the magnitude of that improvement either.
The practical response for a low-resource-language site isn’t to treat this as an unfixable ranking ceiling, but to lean even more heavily on the fundamentals Google’s guidance emphasizes everywhere: genuinely well-written, expert, locally accurate content in the target language, clear on-page structure that doesn’t rely purely on nuanced language-model interpretation to convey meaning (clear headings, explicit structure, unambiguous phrasing), and correct hreflang/language-targeting implementation so Google’s systems can correctly match the right language version to the right query in the first place, since misconfigured language signals are a more common, more fixable cause of poor multilingual performance than any underlying model limitation.
