Google’s NLP systems evaluate relevance by modeling the contextual relationships between words in a query and a page, rather than matching individual keywords in isolation, which lets the system interpret nuance, word order, and connecting terms like prepositions that change a query’s actual meaning. BERT, introduced in 2019, was specifically about improving that kind of contextual understanding within queries. MUM, announced at Google I/O in May 2021, extended this further as a multimodal, multilingual model Google discussed in the context of handling complex, multi-step informational needs. Both represent a real shift away from literal keyword matching toward contextual language understanding, but it’s important to be precise about what Google has and hasn’t disclosed here: Google has never published a checklist of specific “content characteristics” that correlate with a stronger semantic score. That part of the question is answerable only as informed inference, not confirmed Google guidance.
What BERT and MUM actually changed, per Google’s own announcements
BERT (Bidirectional Encoder Representations from Transformers) was announced by Google as a significant update to how Search understands the intent behind queries, particularly ones where small words carry outsized meaning. Google’s own examples at the time centered on prepositions and connecting words, phrases where a word like “for” or “to” changes which entity the query is actually about, cases where older, more literal keyword-matching approaches struggled because they weighted individual terms without fully modeling how the words relate to each other in context. BERT’s contribution was processing language bidirectionally, considering the words before and after a given term simultaneously, which allows the system to better resolve what a query is actually asking rather than treating it as a bag of independent keywords.
MUM, announced roughly two years later, was described by Google as multimodal (able to work across formats like text and images) and multilingual (able to draw on information across languages without requiring a direct translation step), and positioned around handling more complex queries, including ones that might previously have required several separate searches to answer. Google’s own framing at announcement emphasized MUM’s potential for tasks requiring the synthesis of information across multiple sources and formats, representing a further step beyond BERT’s within-query contextual understanding toward broader cross-format, cross-language comprehension.
Successors to both systems have continued in the same general direction Google has publicly described, deeper contextual and semantic understanding rather than literal term matching, but the specific technical details of any newer systems and exactly how they weight content signals haven’t been disclosed with the same level of specificity as the original BERT and MUM announcements.
Why “content characteristics that correlate with semantic scores” isn’t a confirmed list
It’s tempting to translate “Google understands context now” into a tidy checklist of writing characteristics to target, but that leap isn’t something Google has actually confirmed. Google hasn’t disclosed a “semantic score” as an exposed, nameable metric, and hasn’t published which specific content characteristics correlate with performing better under these systems. What can honestly be said is inference, grounded in how the underlying technology works rather than in a Google disclosure: language models built to understand contextual relationships between words are, by the nature of that architecture, better equipped to parse content that’s clearly and unambiguously written, with coherent sentence structure and logical relationships between ideas, than content that’s disjointed, keyword-stuffed, or relies on ambiguous phrasing to cover multiple search terms at once.
This is a reasonable technical inference, clear language is easier for a context-modeling system to parse accurately, but it should be presented as inference, not as a Google-confirmed content checklist. There’s a meaningful difference between “Google’s models are architecturally better at parsing clear, well-structured language” (a defensible technical inference) and “Google scores content on these five characteristics” (an unconfirmed claim dressed up as a fact).
A worked example of the distinction BERT was built to address
Google’s own public examples at the time of the BERT announcement centered on queries where a preposition changes the entity being asked about. One of Google’s cited examples involved the query “2019 brazil traveler to usa need a visa,” where the word “to” is the key signal distinguishing this from a query about a US traveler needing a visa for Brazil. A literal keyword-matching approach might weight “brazil,” “traveler,” “usa,” and “visa” as roughly independent terms and struggle to distinguish which country’s traveler needs a visa for which destination, since all four keywords appear in both possible interpretations of the query. BERT’s contribution was modeling the relationship between “to” and the surrounding words well enough to correctly identify which direction of travel, and therefore which visa requirement, the query was actually asking about.
This example is useful precisely because it illustrates what changed at a mechanical level, understanding relationships between words, not because it implies anything about how to write content differently. A page about Brazilian citizens’ US visa requirements doesn’t need to be restructured to be “BERT-friendly,” it already contains the information a correctly-functioning contextual model would extract. The shift was in Google’s query interpretation and matching capability, not in a new set of content production requirements imposed on publishers.
What about newer systems beyond BERT and MUM
Google has continued to reference ongoing improvements to language understanding in its systems since MUM’s 2021 announcement, generally framed as continued progress in the same direction, better contextual and multimodal understanding, but without the same level of named, standalone technical disclosure that accompanied BERT and MUM specifically. It would be inaccurate to name specific successor systems and attribute detailed, confirmed mechanics to them beyond what Google has actually published, since the pace and specificity of public disclosure has not matched the original BERT and MUM announcements. The honest position for anything beyond those two named systems is that Google has described a general trajectory (continued investment in language understanding, multimodal capability, and handling complex informational needs) without a comparable level of confirmed technical detail to draw specific new conclusions from.
Does this mean keyword usage no longer matters at all
Not quite, and it’s worth being precise about this rather than overcorrecting. Contextual language understanding means Google’s systems are less reliant on exact keyword matches to determine relevance than older, more literal matching approaches were, which is why content that naturally uses synonyms, related concepts, and varied phrasing can still be correctly matched to a query even without repeating the query’s exact words. But this doesn’t mean keywords became irrelevant or that a page should avoid ever using the terms searchers actually type. The practical implication is closer to: writing naturally and clearly for the topic, including the terms a knowledgeable person would actually use, tends to serve contextual matching well, whereas artificially avoiding natural terminology in favor of only synonyms, or conversely stuffing a page with exact-match keyword variants at the expense of clear writing, both work against what these systems are built to parse well.
What this means practically
Given that Google hasn’t disclosed a scoring rubric, the reasonable response isn’t to chase a specific technical checklist but to write in a way that’s genuinely unambiguous about what a page is about and what question it answers, since that aligns with what contextual language models are actually built to parse well regardless of the exact underlying architecture. This means resolving pronouns and references clearly, structuring content so the relationship between a claim and its supporting detail is explicit rather than implied, and avoiding phrasing that’s deliberately vague or keyword-dense purely to capture multiple search variants. This isn’t a guaranteed ranking lever, since Google hasn’t confirmed that a clarity checklist directly produces measurable gains, but it’s consistent with the technical direction Google has publicly described across BERT, MUM, and the general trajectory of its language understanding systems since.
