BERT and MUM: Natural Language Understanding in Search

BERT

Google announced BERT for Search in an October 25, 2019 post by Pandu Nayak. The key points from it:

• Scope at launch: “BERT will help Search better understand one in 10 searches in the U.S. in English, and we’ll bring this to more languages and locales over time.”
• Where it helps: “Particularly for longer, more conversational queries, or searches where prepositions like ‘for’ and ‘to’ matter a lot to the meaning”.
• The example: for “2019 brazil traveler to usa need a visa”, the word “to” matters: “It’s about a Brazilian traveling to the U.S., and not the other way around.”
• Featured snippets: Google said it was “using a BERT model to improve featured snippets in the two dozen countries where this feature is available”.

BERT’s contribution is context: reading a word in relation to the words around it, so small words like “to” and “for” are read as part of the meaning.

MUM

Google introduced MUM on May 18, 2021 in a post by Pandu Nayak. From it:

• MUM “is 1,000 times more powerful than BERT”, and “not only understands language, but also generates it.”
• “It’s trained across 75 different languages and many different tasks at once”.
• “MUM is multimodal, which means it can understand information from different formats like webpages, pictures and more, simultaneously.”
• Google described the plans as future work: “We’ll bring MUM-powered features and improvements to our products in the coming months and years.”

Google’s current ranking systems guide is the place to check how MUM is used today: not for general ranking, but for specific applications, such as improving searches for COVID-19 vaccine information and improving featured snippet callouts.

Where AI Overviews Fit

AI Overviews and AI Mode are a separate layer. Google’s May 2024 announcement says AI Overviews are made possible “by a new Gemini model customized for Google Search”, and doesn’t attribute them to MUM. Visibility in those features is covered in AI Overviews.

What This Means for Content

Better language understanding moves the work from matching strings to answering questions:

• Answer the question as asked, including the small words that define it. A page about visas for travelers to the U.S. should be clear about who is traveling where.
• Cover the topic, not just the phrase. Systems that understand concepts can match a page to related phrasings, so a page that covers the subject thoroughly can serve many of them; see pillar pages and topic clusters.
• Write naturally. Unnatural keyword insertion was never good writing; use the words a reader would use to ask.
• Structure answers clearly, with headings that ask the question and a direct answer beneath, which also helps featured snippets and related questions.
• Serve longer, conversational queries, the kind Google said BERT helps with; see long-tail keywords.

What You Can’t Optimize

There’s no BERT or MUM setting, markup or keyword density. Google’s ranking systems guide describes them as systems Google uses, not as signals sites provide. What you control is the page itself: content that reads well and answers the question, the same thing that served readers before these systems existed. Broader natural-language topics are covered in NLP and search, and entities in entity SEO.

Measuring the Effect

You can’t isolate BERT or MUM in your data, but you can see whether your pages match a wide range of natural queries:

• Query diversity in Search Console: the number of distinct queries a page receives impressions for, including longer and question-form ones.
• Long-tail performance: clicks and impressions for longer queries over time.
• Featured snippet and related question presence for your target questions.

Treat these as signs that your content answers questions well, not as BERT metrics.

Use the queries behind a page as the test: if its impressions come from questions the page doesn’t answer, or from a reading of its question that ignores a word like “to” or “for”, the answer on the page is what to rewrite.