How MUM works
MUM stands for Multitask Unified Model. Google announced it in 2021 as a way of understanding what a question means, not as a new rule for ranking pages. It was trained across many languages at once, so something it learns from material written in one language can inform an answer given in another. It is also multimodal, meaning it is built to handle images alongside text rather than treating them as separate worlds.
The problem it was made for is the long, messy question that keyword matching cannot answer — where someone describes their situation and expects the engine to work out the steps in between. Answering that means joining facts that sit on different pages, in different languages, and sometimes in a picture. Google has said it applies MUM to particular features gradually rather than switching it on across all results, and it has never published a full list of where it runs.
Why MUM matters
MUM matters for the direction it points in, not for anything you can configure. Search keeps moving away from matching strings of text and towards interpreting intent, and each model in that line — from Hummingbird onwards — has made exact phrasing less important and the completeness of an answer more important. A page that covers a subject properly now competes for wording its author never thought of.
The multilingual side matters if you sell outside one language. Google has described MUM as able to carry understanding from one language into another, which means a subject thinly covered in a smaller language is less of a dead end than it used to be. For a Nepali business publishing in English for customers abroad and in Nepali at home, that is a reason to write each version properly rather than running one through a translator.
Common mistakes with MUM
The common mistake is treating MUM as something to optimise for. There is no MUM tag, no MUM setting and no report that shows it working. If a proposal promises MUM optimisation, it is selling a name rather than a method.
The second mistake is reading multimodal as an instruction to pile images onto every page. Images earn their place when they show something words cannot — a product from the angle a customer worries about, a process caught mid-step — and they need real file names, alt text and captions before anything can interpret them at all.
How to act on it
Write for the whole question rather than for the search box. Answer the thing that was asked, then answer the obvious follow-up in the same place, because connected explanation is exactly what these models are built to interpret. Keep the language plain: a system designed to extract meaning has no reason to prefer padded prose, and neither does a reader.
Then check what you are actually being found for. Query data in Google Search Console will show questions you never targeted, and those gaps are the next thing worth writing. Where the constraint is depth of content rather than a technical fault, that is what content-led SEO is for.
