Neural Matching and MUM: How Google Understands Searches, and What MUM Doesn’t Do

Neural matching and MUM are two of the least understood names in Google’s ranking guide. One quietly shapes a large share of searches. The other gets blamed for things it does not do.

Both are worth understanding for the same reason: they change which searches your pages can answer, and knowing that stops you chasing penalties that do not exist.

What neural matching does

Google describes neural matching as an AI system that understands representations of concepts in searches and pages, and matches them to each other. When it was introduced, Google’s Danny Sullivan called it a kind of super synonym system.

His example has become the standard one. Someone searching for why their television looks strange may never have heard of the soap opera effect, the term for motion smoothing that makes films look like video. Neural matching lets Google connect that plain-language search to pages explaining the effect, even though the searcher never typed the words those pages use.

When it arrived

• Mid 2018: Google begins using neural matching in search.
• 24 September 2018: Google reveals it publicly around its 20th anniversary, saying it affected about 30 percent of searches.
• 2019: Google says neural matching had nothing to do with that year’s March core update, after site owners linked the two.
• December 2025: it remains listed in Google’s official guide to ranking systems.

Neural matching and RankBrain

The two sound alike, and Google has drawn the line between them: RankBrain helps relate pages to concepts, while neural matching helps relate words in a search to the concepts behind them.

In practice both widen the set of searches a good page can answer. RankBrain also works on the order of the top results, as covered in RankBrain explained, and the earlier history is in Hummingbird, RankBrain and BERT.

A simple way to keep them apart: neural matching works from the search side, working out what the person means; RankBrain works from the page side, working out what each page is about. Both then feed into ranking alongside the quality signals that decide which matching pages deserve to win.

BERT in one paragraph

BERT, added to Google Search in 2019, completes the picture. Where neural matching connects concepts, BERT reads context: how word order and small words such as “to” and “for” change what a search means. Google’s own example was a search about a Brazilian traveller needing a visa for the United States. Before BERT, results could ignore the word “to” and return pages for Americans travelling to Brazil. After BERT, the direction of travel was understood. The lesson for writers is the same as for neural matching: write naturally and precisely, and the systems will follow.

What the 2023 testimony added

During the United States antitrust trial, Google’s Pandu Nayak described several deep learning models used in ranking, including RankBrain, DeepRank and one called RankEmbed, and confirmed they are trained partly on click and query data.

RankEmbed was described as also helping to find relevant documents, not only to order them. Some analysts link it to neural matching, but Google has not said publicly that they are the same thing, so treat that as informed speculation. The click data question is covered in do clicks affect Google rankings.

What MUM is

MUM, the Multitask Unified Model, was announced by Google in 2021 as a model able to understand and generate language and to work across text and images.

Google’s ranking guide is unusually direct about it. MUM is not currently used for general ranking. It is used for specific applications, such as improving searches for vaccine information and improving the callouts shown in featured snippets.

Myths about MUM

• “MUM demoted my site.” It is not used for general ranking, so it cannot be the cause of a ranking drop.
• “There was a MUM update.” Google has never announced one. Updates are announced for core, spam and similar systems.
• “MUM powers AI Overviews.” Google has said its AI features in search are built on customised Gemini models.
• “You need to optimise for MUM.” There is nothing to optimise for in a system that does not rank general results.

When a name like MUM is used to explain a drop, it is usually a sign the real cause has not been found yet. How AI features in search actually work is covered in AI Mode vs AI Overviews.

What this means for your content

• Write for the concept, not the keyword. A page that genuinely explains an idea can match searches it never names.
• Include how people describe the problem. Use both the plain-language symptom and the technical term, so readers recognise their situation.
• Do not stuff synonyms. Systems that understand concepts make lists of variations pointless.
• Keep one page per concept. Several pages chasing variations of the same idea compete with each other.

Finding out how people describe the problem

• Read your support inbox for the words customers use before they know the technical term.
• Look at long queries in Search Console, which often show exactly how people phrase their situation.
• Read forum threads in your field, where people describe symptoms in their own words.
• Note the questions shown under People also ask for your main topics.

Put that language near the top of the page, in the description of the problem, and the expert explanation underneath. Readers recognise their situation, and Google sees the full span of the concept.

A worked example

Imagine a page about turning off motion smoothing on televisions.

• A weak version repeats “motion smoothing settings” throughout and lists twenty brand names in a block.
• A strong version opens by describing what people notice, films that look oddly smooth or cheap, names the soap opera effect, explains why it happens, and walks through the setting on each major brand.

The strong version can match searches about a television looking strange, films looking like a soap opera and motion settings alike, because it covers the concept from the searcher’s side as well as the expert’s. That is exactly the gap neural matching was built to close.

Why this matters more for small sites

Large sites win many broad, short searches on authority alone. Long, specific and oddly phrased searches are different. They depend far more on whether a page genuinely understands the problem.

That is where neural matching helps a small expert site most. A page that describes a niche problem the way people experience it can be matched to hundreds of variations no keyword tool would list, and many of those searches have little competition from larger sites.

Language systems and traffic drops

Neural matching, RankBrain, BERT and MUM are all about understanding language. None of them judges whether your site deserves to rank, so none of them can penalise it.

• They can change which searches send you traffic, as Google gets better at matching pages to the way people actually ask.
• They rarely explain a sudden large drop. That pattern points to a core or spam update, or to a technical problem.
• Check the queries, not just the total. A page that lost one phrase but gained dozens of related ones may have been re-matched rather than demoted.

Reading that kind of shift is covered in rankings dropped but traffic is the same. Where these systems sit among Google’s others is in Google’s ranking systems explained, and other common misreadings are collected in Google SEO myths.