MUM (Multitask Unified Model)

What MUM is

MUM (Multitask Unified Model) is a multimodal language model Google announced at I/O on May 18, 2021 — built on its T5 architecture, trained across 75 languages, and able to process text and images together rather than as separate problems. The practical stake: MUM powers the features that answer complex, multi-step needs directly in Google's interface, which means journeys that used to take eight of your pageviews can now resolve in one enriched results screen.

Google claimed MUM is "1,000 times more powerful than BERT." Treat that as Google's marketing framing of model scale, not an independently verified benchmark — no external party has measured it. What's verifiable is where MUM actually shipped, and that list is more instructive than the number. In the model lineage, MUM is the step after BERT: from understanding a query to reasoning across languages, formats, and subtasks at once.

The documented examples

Google's launch scenario: a hiker who has climbed Mt. Adams asks how to prepare for Mt. Fuji in the fall. Answering requires comparing two mountains, reasoning about seasonal conditions, and inferring implicit subtasks (fitness, gear, weather) — plus surfacing Japanese-language sources a US searcher would never find. That's the "multitask" in the name.

The first confirmed production use was narrower and real: in mid-2021, Google used MUM to identify over 800 name variations of 17 COVID-19 vaccines across some 50 languages within weeks — a cross-language entity task that would previously have taken months of manual curation. The second flagship: multisearch in Lens (rolled out from 2022), where you photograph a floral shirt and type "socks with this pattern," combining image and text in one query.

The rollout timeline that actually happened

Worth pinning down, because MUM coverage in the trade press ran far ahead of shipped reality. May 2021: announcement at I/O, explicitly framed as early-stage. June 2021: first production use, the vaccine-name matching. September 2021: Search On event previews multisearch and "Things to know." April 2022: multisearch opens in US beta; "multisearch near me" follows later that year. After that, the MUM label quietly fades from Google's communications as Gemini-era branding takes over the generative story. So roughly eighteen months from announcement to its known feature set — a useful calibration the next time a model announcement promises the end of SEO by Thursday.

MUM capabilities and where you can actually see them

How to check MUM-era exposure for your site

1. Search five of your money topics and inventory the SERP furniture: "Things to know," topic refinement chips, video key moments. Each module is a subtask Google extracted — note which ones your content covers and which it skips.
2. In Search Console, segment queries of six-plus words and compare their CTR trend against short queries over 24 months. Complex-query CTR erosion with stable rank is the signature of answers being absorbed into the results page.
3. Take product-category pages and run their hero images through Google Lens yourself. If Lens can't identify your product or matches a competitor's, your visual inventory is invisible to multisearch.
4. If you publish in multiple languages, spot-check whether your strongest-language content has an equivalent for queries where Google surfaces translated or cross-language results. Our AI Overviews citation guide covers the adjacent synthesis surfaces.
5. Re-run this audit quarterly. MUM-backed features ship incrementally with no changelog for your niche.

Common mistakes

• Treating MUM as a ranking update. It's a model powering specific features, not a core-update-style reranking. Google explicitly said at launch it wasn't being used for ranking search results generally — attribute traffic shifts carefully.
• Repeating the "1,000x" claim as established fact. It's Google's own characterization. Quoting it uncredited in client decks will bite you when someone asks for the benchmark.
• Ignoring images because "we're a text site." Multisearch makes images queryable. Original, well-marked-up images are now answer inventory, not decoration.
• Answering only the literal query. MUM-era features decompose intent into subtasks. A "Mt. Fuji hiking" page that skips gear, season, and fitness gets its gaps filled by someone else's content in the same SERP.
• Assuming English-only competition. Cross-language transfer means the definitive Japanese resource can inform what English searchers see. Your competitive set is bigger than your language.

FAQ

Is MUM live in search right now?

Yes, in specific applications: vaccine name matching (2021), multisearch (2022), "Things to know"-style features, and video topic understanding. It is not a blanket replacement for the ranking stack.

How is MUM different from BERT?

BERT understands language within a query or passage. MUM adds multimodality (images + text), 75-language transfer, and multitask reasoning — generating and connecting information, not just interpreting it. Both sit on the NLP stack the previous entry maps out.

How do I "optimize for MUM"?

You don't, directly. You cover topics at subtask depth, publish original images with accurate markup, and treat your best content as multilingual inventory. Those are the inputs MUM-powered surfaces consume.

Is MUM the model behind AI Overviews?

No — AI Overviews run on Gemini-family models. MUM was the 2021 architectural step in that direction; Google has since moved to newer models for generative features, while MUM-built features persist.