MUM, short for Multitask Unified Model, is an AI model Google announced in May 2021 that can understand and generate language, read images alongside text, and carry what it learns in one language across to others. Google uses it for a handful of specific jobs in Search rather than as a general ranking system.
How MUM works
MUM is built on the same transformer architecture as BERT, but Google described it as far more capable: trained across 75 languages and on several tasks at once, so it builds a broader picture of a topic than a model that only predicts missing words. Two traits set it apart.
- It is multilingual. Knowledge learned from sources in one language can inform answers in another. In practice this mainly matters for topics where good information exists in only a few languages.
- It is multimodal. It can take an image and a text question together, which is the idea behind asking Google Lens about a photo and adding words to refine it. A multimodal model treats the picture as part of the query, not as an attachment.
Google’s public examples of MUM in use have been narrow. It was used to identify hundreds of names for COVID-19 vaccines across dozens of languages so that people searching in any of them found official information. It has powered features that suggest related topics and refinements for broad queries, and it helps with spotting searches for personal crisis information so that helpline details can be shown. Google’s own list of ranking systems has said MUM is not used for general ranking. At the time of writing (October 2026), much of the attention has moved to Gemini-based features such as AI Overviews, which are a separate system.
Why it matters
MUM matters less as something to optimise for and more as a signal of direction. Google has spent years moving from matching words to understanding what a person is trying to get done, and MUM was designed for the long, multi-step questions people actually have: planning a loft conversion in a Victorian terrace, comparing two business bank accounts, or working out whether a rash needs a GP appointment.
For a UK business, the practical lesson is that pages written around a single keyword phrase compete poorly with pages that cover the real task a searcher faces. If a homeowner in Bristol needs to know about party wall notices, planning permission and costs before hiring a builder, the page that explains the sequence clearly is the one these systems are built to recognise.
Common mistakes
- Treating MUM as a ranking update. There was no “MUM update” to recover from. If traffic dropped in 2021 or 2022, look at the core updates from that period instead.
- Machine-translating pages to chase multilingual reach. MUM’s cross-language ability is Google’s tool for understanding, not an invitation to publish thin translated copies for audiences you do not serve.
- Ignoring images. Visual search is real. Stock photos with empty alt text give a multimodal system nothing to connect to your products or services.
- Writing for the model. Nobody outside Google can see how MUM weighs a page. Content written to “please MUM” usually ends up vague.
How to act on it
Work out the full job behind your main searches. For each important page, list the questions a customer has before, during and after the one the page targets, and answer the ones that belong there. Understanding search intent at that level is more useful than any tactic aimed at a specific model.
Use original photographs of your work, products and premises, with descriptive file names and alt text, so an image search for a specific item can lead somewhere. Keep each page focused on one intent and link to the next logical step rather than cramming everything into one page. If you want help planning content around how people really search, that is the core of my content SEO and strategy work.