Write for an engine that understands, not one that matches
Structure and markup refreshed for current answer engines; the original analysis is preserved.
At I/O this month Google announced MUM, a language model it calls far more capable than BERT — multimodal, trained across many languages, built for complex questions. It is not deployed and changes nothing in rankings today. What it confirms is the direction: from matching keywords to understanding meaning. Write for the second, because the lexical tricks that fooled the first lose value as the engine reads more like a person.
the short answer
Google announced MUM at I/O this month — a language model it presents as the successor to BERT, multimodal and trained across ~75 languages. It is not deployed and changes nothing today. The value is a direction: the engine is moving from matching keywords to understanding meaning. Write for substance and natural language; the lexical tricks lose value as the engine reads more like a person.
key takeaways
- At I/O this month Google announced MUM, a language model it presents as the successor to BERT and far more capable — multimodal, trained across roughly 75 languages, built for complex multi-step questions.
- It is not deployed. Google called it still in testing, with no launch date, and said it wants to evaluate it carefully first. There is nothing live to optimise for today.
- The value of the announcement is a direction, not a task: the engine is moving from matching keywords to understanding meaning, the journey BERT began in 2019.
- The tactics that lose value are lexical tricks — keyword stuffing, density targets, exact-match contortions — because an engine reading for meaning is harder to fool by supplying more of the right words.
- What keeps and grows in value is substance: real depth, genuine expertise, natural language, the topic treated as a whole. Write for a reader, because that is what the engine increasingly approximates.
from matching to understanding
The left box is the engine SEO grew up gaming; the right box is the reader it is turning into. You do not need the newest model to be live to know which way to write — the arrow has pointed one way for years, and the announcement just lengthened it.
The announcement, in four parts
What Google announced this month; the direction of travel it confirms; what that asks of your content; and why the brief barely changes. Open each part.
01 What Google announced this month
At its I/O event this month Google introduced MUM, the Multitask Unified Model, presented by the head of Search as the next major step in how the engine understands information. The headline claim was scale: Google described MUM as roughly a thousand times more capable than BERT, the language model it rolled into Search in 2019. It is built on the same underlying Transformer architecture, so it is an evolution of that line rather than a departure, but it extends it in two directions that matter. It is multimodal, designed to understand text, images, and video together rather than text alone, so a question that mixes a picture and a sentence can be read as a single request. And it is multilingual at a new scale, trained across roughly seventy-five languages and meant to carry meaning between them, so knowledge written in one language can help answer a question asked in another. Google illustrated the ambition with a deliberately complicated example: someone who has hiked one mountain and wants to prepare for a harder one, a question that today would take many separate searches to answer and that MUM is meant to handle as one. The crucial caveat, and the reason this note is measured rather than breathless, is that MUM is not live. Google was clear that it remains in testing, that it will come to products at some unspecified future point with no date attached, and that it wants to evaluate the model carefully — including for bias — before letting it influence what people see. So nothing in your rankings changed this month, and nothing you do this month can optimise for MUM, because there is nothing deployed to optimise against. What arrived was not a ranking update but a statement of direction, and the direction is the part worth reading closely.
02 The direction of travel
To see why an unreleased model is worth paying attention to, put it on the line it belongs to. For a long time, a search engine was fundamentally a matching machine: it took the words in your query and looked for pages containing those words, with refinements, and the craft of ranking well was substantially the craft of getting the right words onto the page in the right amounts. That world rewarded a particular kind of work — keyword research as token-counting, density as a dial, exact-match phrases placed for the crawler — because an engine that matched words could be served, and gamed, by supplying more of the right words. BERT in 2019 was a visible turn away from that. It let the engine read a query in context, understanding how words relate and what small connecting words do to meaning, which is why it helped most with conversational, natural-language queries that a pure matcher handled badly. MUM, as announced, is the same turn taken further: more capability, more formats, more languages, all in the service of understanding what a query and a page mean rather than which words they share. You do not need MUM to be live to read the trajectory, because the trajectory has been visible for years and MUM simply extends it and declares Google’s intent to keep going. The engine is getting better, step by step, at doing what a knowledgeable human reader does — grasping meaning, not tallying terms — and that single direction, more than any one model, is what content strategy should be built around. Design for where the engine is clearly going, not for the matching machine it is steadily leaving behind.
03 What that asks of your content
If the engine is moving from matching to understanding, the content that benefits is the content an understanding reader would judge well, and the work splits cleanly into what to drop and what to deepen. What to drop is the lexical manipulation that only ever worked on a matcher: writing to a keyword-density figure, repeating an exact-match phrase past the point a human would tolerate, contorting sentences so a target string appears verbatim, producing thin pages whose value is the presence of words rather than the presence of answers. None of that survives contact with a model reading for meaning, because such a model is not counting your phrase — it is assessing whether you actually addressed the question. What to deepen is everything that a knowledgeable reader rewards. Cover a topic with real completeness, answering not just the question asked but the obvious next one, because a model built for complex, multi-step questions is built to value content that resolves a whole need rather than a fragment of it. Show genuine expertise in the specifics — the precise detail, the non-obvious qualification, the thing only someone who truly knows the subject would include — because specificity is exactly what cannot be manufactured by an engine optimiser and exactly what an understanding engine can increasingly recognise. Write in the natural language a person would use to ask and answer, since that is the language the engine is learning to read. And treat your subject as an entity to be explained thoroughly and accurately rather than a keyword to be hit a quota of times. The shorthand is simple: stop writing for the crawler and write for the knowledgeable reader the crawler is turning into.
04 Why the brief barely changes
Here is the part that should be reassuring rather than disorienting: almost none of this is new advice. Write with depth, demonstrate real expertise, use natural language, cover the subject honestly and completely — this is what good writers and serious practitioners have recommended for as long as there has been content worth reading, long before any particular model. The MUM announcement does not hand you a new brief; it raises your confidence in the old one. Every step the engine takes toward understanding makes genuine substance more rewarded and lexical tricks less, which means the advice that was always right on the merits is becoming right on the mechanics too — the gap between writing well and ranking well keeps narrowing. That is why the correct response to an announcement like this is not to reorganise anything today. MUM is not live; there is no migration to run, no setting to change, no MUM-readiness checklist worth buying. The correct response is quieter and more durable: keep investing in the substance and expertise that an understanding engine rewards and a matching engine merely tolerated, and stop investing in the tricks that a matching engine rewarded and an understanding engine will ignore. You are not chasing a model that has not shipped; you are aligning with a direction that has been clear for years and just got a thousand-times-more-capable confirmation. Building content for the reader the engine is becoming, steadily and ahead of the curve, is the patient, substance-first work the AC Group has done for 27 years.
Why an unreleased model still matters
It is reasonable to ask why a model that is not deployed deserves a word of your attention, and the answer is that announcements like this are less a product than a compass. Google rarely tells you in advance exactly how it will rank; here, in describing what it is building, it told you plainly which way it is heading — toward understanding meaning across formats and languages, and away from matching words. That is strategic information even when the model behind it is months or years from your results, because the worst position in this work is to keep investing in tactics the engine is visibly leaving behind. A compass does not move your feet, but it stops you walking confidently in the wrong direction.
It also guards against the opposite error, which the hype around any big announcement encourages: treating it as an emergency that demands immediate, dramatic change. MUM is not live, there is no MUM-readiness task, and anyone urging you to overhaul your site for it this month is selling motion, not value. The measured response holds both truths at once — the model changes nothing today, and the direction it confirms should shape what you invest in from today. Reading announcements for their direction rather than their drama, and steering content toward where the engine is genuinely going, is the kind of calm, substance-first judgement the AC Group has applied for 27 years.
What to do with this
Do not optimise for MUM, because there is nothing to optimise against; instead, lean further into the work an understanding engine rewards. Drop the lexical habits that only ever worked on a matcher: stop writing to a keyword-density figure, stop repeating exact-match phrases past what a reader would bear, and stop producing thin pages whose only asset is the presence of the right words. Then deepen the things substance is made of: cover your topics with real completeness, answer the next question as well as the asked one, and show the specific, precise expertise that only someone who knows the subject could write.
Write in the natural language a person would actually use, and treat each subject as a whole topic to explain well rather than a keyword to hit a number of times. None of this is a reaction to a model that has not shipped; it is an alignment with a direction that has been clear since BERT and just got a far more capable confirmation. The reassurance is that the brief barely changes — the content that serves a reader well is the content an understanding engine increasingly rewards, so the right work was always the same work. Building substance and expertise for the reader the engine is becoming, ahead of the curve rather than in a panic after it, is the patient, substance-first work the AC Group has done for 27 years.
MUM and meaning, plainly: quick answers
What is MUM, and should I optimise for it now?
MUM — the Multitask Unified Model — is a language model Google announced at its I/O event this month, presented by its head of Search as the successor to BERT and described as around a thousand times more capable. It is built on the same Transformer architecture as BERT, but it is multimodal, meaning it is designed to understand text, images, and video together, and multilingual, trained across roughly seventy-five languages so it can carry meaning between them. The point of all that capability is to handle complex, multi-step questions that today force a searcher to run many separate searches. So that is what it is. As for optimising for it now: you cannot, and you should not try. MUM is not deployed. Google described it as still in testing and coming to its products at some unspecified future point, with no launch date given, and it was explicit that it wants to evaluate the model carefully — including for bias — before letting it shape results. There is therefore nothing live to optimise against, no MUM ranking factor to chase, and anyone selling MUM optimisation today is selling a guess. The value of the announcement is not a task; it is a direction. It tells you, with unusual clarity, which way the engine is travelling — toward understanding meaning rather than matching words — and that is worth designing for even though the model itself is not yet here.
How is MUM different from BERT?
BERT, which Google rolled into Search in 2019, was a breakthrough in understanding natural language: it let the engine read a query in context, grasping how the words relate to each other rather than treating them as a bag of independent terms, which is why it improved results for conversational and prepositional queries where word order and small words change the meaning. BERT, though, worked on text, and its multilingual reach was limited at launch. MUM is presented as the next step along the same line rather than a different idea. It uses the same underlying Transformer architecture, but Google claims far greater capability, and it adds two dimensions BERT lacked. The first is multimodality: where BERT read text, MUM is built to understand text, images, and video together, so a question that combines a photo and a sentence can be understood as one. The second is language: trained across roughly seventy-five languages, MUM is meant to carry meaning across them, so knowledge written in one language can help answer a question asked in another. The throughline from BERT to MUM is the important part for anyone writing content. Both are steps in the same journey, away from matching the literal words on a page toward understanding what a page and a query actually mean. MUM simply pushes that journey further, into more formats and more languages, and signals that Google intends to keep going in that direction.
Does this mean keyword SEO is dead?
No, but the part of it that was always a trick is worth retiring, and the announcement is a good prompt to do so. Keywords themselves are not going anywhere as a concept: the words on your page still tell the engine what your content is about, and the words in a query still express what someone wants, so matching topic and intent matters as much as ever. What loses value, as the engine gets better at understanding meaning, is the manipulation of words as tokens rather than carriers of meaning — stuffing a phrase to hit a density target, repeating an exact-match string unnaturally, writing for the crawler in a way no human would read. Those tactics worked, to the extent they ever did, because an engine that matched words could be gamed by supplying more of the right words. An engine that understands meaning is harder to fool that way, because it is reading for what you actually say, not counting how often you say a phrase. So the honest distinction is between keywords as a signal of what you genuinely cover, which remains useful, and keyword optimisation as a substitute for substance, which the direction of travel steadily devalues. Write the words a real reader needs, in the natural language they would expect, and you keep the part of keyword work that helps while shedding the part that an understanding engine increasingly ignores.
So what should I actually do?
Write for a reader the way you would if a knowledgeable person were going to read it closely, because that is what an understanding engine is increasingly approximating. In practice that means a few concrete things. Cover the topic with real depth and completeness rather than thin coverage padded with phrases, because a model built to handle complex, multi-step questions rewards content that genuinely answers the whole question and the obvious next one. Demonstrate actual expertise — the specific, the precise, the things only someone who knows the subject would write — because that is exactly what cannot be faked by an engine reading for meaning. Write in natural language, the way the question is really asked, instead of contorting sentences around exact-match strings. And treat the subject as a whole entity to be explained well, not a keyword to be hit a certain number of times. None of this is new advice; it is the advice good writers and honest practitioners have always given. What the MUM announcement does is raise the confidence that this is the durable bet, because every step the engine takes toward understanding makes substance more rewarded and tricks less. Designing content for an engine that reads like a person, well before it fully does, is the patient, substance-first work the AC Group has done for 27 years.
A note on sources and timing
This is written in May 2021, just after Google’s I/O event. We have described MUM as Google announced it there — the Multitask Unified Model, presented by the head of Search as the successor to BERT, said to be roughly a thousand times more capable, built on the same Transformer architecture, multimodal across text, images, and video, and trained across about seventy-five languages, with the mountain-preparation example used to show its handling of complex questions. We have been careful to note what Google was careful to note: that MUM is not deployed, has no launch date, and is being evaluated — including for bias — before it shapes results. We have therefore drawn no conclusions about a live ranking factor, only about the direction the announcement confirms: from matching words to understanding meaning, the line BERT began in 2019. The durable point survives whatever MUM eventually becomes: write for understanding, not matching — the substance-first work the AC Group has done for 27 years.