MUM (Multitask Unified Model)

What is MUM?

MUM (MUM) is an AI model Google announced in May 2021. Google describes it as able to both understand and generate language, trained across 75 languages and many tasks at once, and multimodal, meaning it can work with text and images.

It is often written about as a ranking update. Google's own documentation says otherwise.

MUM uses the T5 text-to-text framework and is 1,000 times more powerful than BERT.

Pandu Nayak, Google, MUM: A new AI milestone for understanding information, Google Blog

What does MUM actually do in Google Search?

Google has named a short list of uses. Its ranking systems guide cites COVID-19 vaccine information and featured snippet callouts. In March 2022 Google said it would use MUM to detect a wider range of personal crisis searches and to improve spam protection in languages with little training data.

• Improving searches for COVID-19 vaccine information.
• Improving featured snippet callouts.
• Detecting personal crisis searches so help resources can be shown.
• Improving spam protections across languages.

MUM can better understand the intent behind people's questions to detect when a person is in need, which helps us more reliably show trustworthy and actionable information at the right time.

Pandu Nayak, Google, Using AI to keep Google Search safe, Google Blog

What evidence is there that Google uses MUM?

The evidence label is Confirmed, with an important limit. Google lists MUM in its ranking systems guide but in the same entry says it is not used for general ranking. A 2022 Google post made the same point, contrasting MUM with RankBrain, neural matching and BERT, which do help rank results.

So the confirmed fact is that MUM exists and powers specific features. Any claim that MUM ranks ordinary web pages is not supported by Google's public statements.

It's not currently used for general ranking in Search but rather for some specific applications such as to improve searches for COVID-19 vaccine information and to improve featured snippet callouts we display.

Google, A guide to Google Search ranking systems, Google Search Central

What should sites do about MUM?

Do not build a strategy around MUM as a ranking system. The one confirmed use that touches ordinary pages is featured snippet callouts, the short highlighted value Google sometimes shows above a snippet.

The testable step is narrow: where a page answers a factual question with a specific value, such as a date, a number or a name, state that value plainly in text next to the answer. If Google cannot find it in text, it cannot highlight it. This is a proxy for one MUM application, not a way to influence MUM generally.

Which Laurelin audit checks test for MUM?

Because MUM is not a general ranking system, these checks test the narrow surfaces Google says it affects. They are proxies.

Key answer value not stated in text beside the answer: the new check for this page, testing whether a featured snippet callout has a clear value to extract.

Question heading with no direct answer: a snippet needs a direct answer before a callout is possible.

Important facts only in images: values locked in images are hard to extract as text.

Key info only in video: the same problem for video.

Poor auto-translated content: MUM's cross-language uses make low quality translations more visible, not less.

What are the key dates for Google MUM?

• 2021-05-18: Google announces MUM (source)
• 2022-02-03: Google says MUM is not used to help rank results like RankBrain, neural matching and BERT (source)
• 2022-03-30: Google announces MUM for personal crisis detection and spam protection (source)