Check this page with an assistantOpens a chat asking it to summarise this article and name the evidence behind each claim.
Claude opens with the prompt on your clipboard: Anthropic does not support prefilled prompts on the web, and we would rather copy it than ship a button that drops it.
The Evidence Is in Your Own Data
Open Search Console, pick any page, and read the queries it receives. A substantial share will be phrasings the page does not contain anywhere. That is not an anomaly to be optimised away; it is the system working as designed, and it disproves the model of search most advice still assumes.
String matching would make that impossible. A document either contains a phrase or it does not. The only way a page ranks for words it never uses is if the system is representing what the page is about and what the query is asking, and comparing those representations rather than the text itself.
This is also the honest explanation for why a long list of tactics quietly stopped working. Keyword density, exact-match repetition, and stuffing synonyms into a paragraph were all optimisations against string matching. They did not become forbidden so much as irrelevant, which is a more durable reason to stop doing them.
Three Ideas That Explain Most Behaviour
You do not need the mathematics. Three concepts account for most of what people find surprising about modern results: meaning representation, entities, and passage-level retrieval.
| Concept | What it means practically | What it explains |
|---|---|---|
| Meaning representation | Text is represented by what it is about, not the words used | Ranking for phrasings your page never contains |
| Entities | Specific things, not strings, with relationships between them | Why identity clarity matters; see entity SEO |
| Passage retrieval | Parts of a page can be selected independently | A long page cited for one paragraph on a niche question |
| Query interpretation | The question is expanded and decomposed before retrieval | Results answering things you did not literally ask; see query fan-out |
Passage retrieval is the one with the clearest practical consequence. If parts of a page can be selected on their own merits, then a page is better understood as a collection of potential answers than as a single document competing for a single query. That reframing changes how you structure writing more than any keyword decision does.
What This Actually Changes About Writing
Less than the vocabulary suggests, and in a specific direction. Semantic retrieval closed most of the gap between writing well for a person and being understood by a machine, so the surviving advice is mostly just clarity with a structural requirement attached.
- Use your readers' vocabulary, not a density target.
The words people use are worth using because that is how you communicate, and because they signal what you are about. That is a different instruction from repeating a phrase a set number of times, and it produces different sentences.
- Cover the subject, not the phrase.
A page that genuinely addresses a topic will naturally match many phrasings of questions about it. A page assembled around one keyword and its variants covers a phrase, which is a smaller thing than a subject.
- Be unambiguous about what you mean.
Name the entity rather than relying on pronouns and implication. A passage lifted out of context has to stand alone, and a paragraph full of "it" and "they" does not.
- Structure for extraction.
Headings that name questions, direct answers underneath, and sections that make sense in isolation. This is the one genuinely structural requirement, and it is the same one that serves snippets and generated answers.
- Stop trying to trick a representation you cannot see.
You cannot inspect an embedding, tune against it, or verify a change to it. Any service claiming to optimise for one is describing something unobservable, which is a useful tell.
What This Retires
A number of habits survive in templates and plugin scoring long after the mechanism that justified them disappeared. They are not dangerous so much as pointless, and they make writing worse in exchange for nothing.
Keyword density targets, exact-match repetition in headings that reads awkwardly, stuffing every synonym into an opening paragraph, and building separate near-identical pages for phrasings of the same question. The last is the most expensive, because it produces competing pages that split signals, which is the mechanism described in keyword cannibalization.
Plugin scoring is where most of these persist. A content analyser telling you to use a phrase three more times is applying a rule from an era of string matching, and following it makes the page slightly worse for the reader in pursuit of a number nobody outside the plugin can see. That is the argument made at greater length in WordPress SEO.
The Honest Limits of This Explanation
This is a working model built from what search engines have published and what practitioners observe. Nobody outside these companies knows how the representations are built, weighted, or combined with everything else that goes into a result.
Be sceptical of anyone offering more precision than that. Confident accounts of how vectors are compared, what the exact retrieval pipeline is, or how to write for a specific model architecture are inference presented as engineering, and the inference is frequently built on a paper that describes something adjacent to what is actually deployed.
The reason the model is still worth holding is that it explains observable behaviour and gives correct advice. Whatever the internals are, pages that cover subjects properly, name things clearly, and can be read in sections do better than pages built around phrase repetition. That has been stable for years and does not depend on the mechanism being described exactly right, which is the property to look for in any explanation of this kind. The stage this fits into is in how Google Search works.
Questions People Ask About Semantic Search
- What is semantic search?
Retrieval based on meaning rather than string matching. Instead of finding documents containing your exact words, the system represents both the query and the content in a way that captures what they are about, and matches on that. It is why a page can rank for a phrase it never contains, and why repeating a keyword stopped helping many years ago.
- Do keywords still matter at all?
As evidence of what a page covers and as the vocabulary your readers actually use, yes. As a density target to hit, no, and they have not for a long time. Use the words your audience uses because that is how you write clearly for them, not because a counter needs satisfying.
- What is an entity in search?
A specific thing rather than a string: this company, this person, this product, distinct from others with similar names. Search systems reason about entities and their relationships, which is why disambiguating your own identity matters and why two pages using the same words about different things are not treated as equivalent.
- Does this mean I should write for machines?
The opposite. Semantic retrieval closed most of the gap between writing clearly for a person and being understood by a system, which is why the tactics that survived are the ones that also make content better to read. The residual difference is structural: clear headings, direct answers, and self-contained sections help both.
- How do I optimise for embeddings?
You do not, and anyone offering to is selling something. You cannot see the representation, tune against it, or verify a change. What you can do is cover a subject thoroughly, be unambiguous about what you are talking about, and structure the page so specific answers are extractable. That is the whole actionable surface.
