A writer gets a page to produce for “affordable project management software for small teams.” To play it safe, the exact phrase goes into the H1, the opening sentence, three subheadings, and the meta description, alongside “cheap project management tool,” “budget PM software,” and “low-cost task manager.” The page reads like a stack of search phrases wearing a paragraph. It ranks for none of them. A thinner competitor, one that never repeats the exact phrase and instead walks a five-person team through choosing a tool, sits at the top of the results. The distance between those two pages is the distance between writing for a string of characters and writing for what the person meant.
That distance is the work of a set of language systems Google has built into search over the past decade. You do not need to reverse-engineer them. You do need to understand what they changed, because they quietly rewrote the rules that a lot of on-page advice still assumes.
What “understanding language” refers to #
Google has named several of the systems involved, and the public timeline is worth knowing because each one moved search a step away from literal matching.
- RankBrain (announced 2015). Google described it as a machine-learning system that helps interpret queries, especially ones it has never seen before, by relating them to queries it does understand. Google has said a large share of daily searches are genuinely new, so a system that guesses intent from unfamiliar phrasing matters more than it sounds.
- Neural matching (described around 2018). Google framed this as connecting the words in a query to the concepts behind them, so a search that never uses your exact terms can still be matched to your page.
- BERT (rolled into search in 2019). Google presented BERT as a model that reads a query in context, paying attention to how words relate to each other rather than treating them as a bag of terms. In Google’s own example, the query “2019 brazil traveler to usa need a visa” hinges on the word “to,” which sets the direction of travel. Older systems tended to drop small connecting words like that. BERT was built to keep them.
- MUM (introduced 2021). Google described MUM as multimodal and multilingual, and claimed it was far more capable than BERT. Google has been careful about where and how much it uses MUM, so the honest reading is that it points at a direction (understanding meaning across formats and languages) rather than a switch that flipped overnight.
The exact weight of any single system, and how much of search it touches today, is not something Google publishes in a way anyone outside the company can verify. Treat the specific mechanics as background. The pattern across all of them is the part that changes your writing: search moved from matching words to interpreting meaning.
Why phrase-matching lost its edge #
For years, the safe move was to identify a target phrase and place it in the important spots, then repeat close variants through the body. It worked because the ranking systems were closer to a sophisticated find-and-count. When the systems started reading context, that tactic stopped buying anything, and past a certain density it started signaling low quality instead.
Consider a search like “do estheticians stand a lot at work,” one of the queries Google used to explain BERT. The word “stand” here means physical endurance, not a display rack or a legal position. A page optimized to repeat “esthetician standing” would miss the point. A page that plainly answers whether the job is physically demanding, and for how long, matches the meaning even if it phrases things differently. The searcher was never looking for a word. They were looking for an answer to a question, and the systems now try to read the question the way a person would.
Write the answer, not the target phrase #
The practical shift is small to describe and hard to unlearn. Instead of asking “did I include the keyword enough times,” the more useful question is “does this page answer the question behind the keyword, completely, before the reader gives up and clicks back.” Those two habits pull content in different directions.
| Keyword-first habit | Meaning-first habit |
|---|---|
| Repeat the exact phrase in every section | State the phrase naturally once or twice, then answer the underlying question |
| Pad the page to hit a word count | Cover what the question requires, then stop |
| Write for a phrase that has no reader behind it | Write for the specific person who typed it and what they will do next |
| Treat synonyms and related terms as risky dilution | Use the natural vocabulary of the topic, because the systems connect concepts, not strings alone |
None of this means the target phrase disappears. It still tells both readers and search systems what the page is about, and it still reflects how real people phrase the thing they want. The change is that the phrase became a starting point rather than the finish line.
Cover the neighborhood of the question #
Most searches carry a cluster of unasked follow-ups. Someone searching “how to read a P and L statement” almost certainly wonders next what a healthy margin looks like, how it differs from a cash flow statement, and which line to distrust. A page that answers only the literal question and ignores the neighborhood feels thin, and the reader leaves to find the rest somewhere else. That return-and-click-again behavior is exactly the signal a page does not want to generate.
A reliable way to find the neighborhood is to write down the three or four questions a reader asks immediately after the headline question, then make sure the page answers each one in its own right. This is where the language systems and simple reader empathy point at the same target: the page that resolves the whole cluster tends to read as the more complete answer, and completeness is something the systems can approximate by looking at which concepts and related entities a page covers.
Build passages that stand on their own #
Google has said it can rank an individual passage from a page, not only the page as a whole, which means a single well-built section can surface even when the rest of the page is about something broader. The same structure that helps a passage rank also helps it get pulled into a featured snippet or quoted in an AI answer, because all three want a self-contained chunk that resolves one question cleanly.
In practice that means each section carries its own context instead of leaning on the paragraph above it. A subheading phrased as the actual question, followed by a direct answer in the first sentence or two, gives the systems a clean unit to lift. It also serves the impatient reader who scrolled straight to the part they came for. One structural habit, two audiences.
What this does not mean #
The correction here is easy to overshoot. Reading that Google understands meaning, some writers swing to the opposite failure and write loose, keyword-free prose that never names the thing it is about, then wonder why it does not rank. Meaning-first is not vocabulary-free. The words a topic naturally uses are still evidence of what the page covers, and dropping them entirely removes signal.
There are real limits worth stating plainly. Understanding language does not fix a page that answers the wrong intent, and it does not reward padding disguised as thoroughness, since covering the neighborhood is different from stuffing in every loosely related term you can find. Chasing semantic completeness too hard produces its own kind of bloat, where a focused answer drowns in tangents. The discipline is to answer the question and its close relatives well, not to prove how many related words you can fit.
If a page hits its target phrase and still underperforms, the most useful next step is usually not more optimization of the phrase. Re-read the page as the person who typed the query, ask whether it answers what they meant and what they would wonder next, and fix the first place the answer runs thin. The systems have spent a decade learning to read for meaning. The pages that do well are the ones written the same way.