Semantic search explained: how an engine turns words into meaning

Semantic search gets one sentence in most articles: the engine understands meaning, not just words. That sentence is true and useless, because it never says what understanding meaning is built from, so it cannot tell you what to write differently. This post takes it apart into five mechanics, each followed by the writing decision it forces, then names what those mechanics quietly make pointless.

From tokens to vectors

A model does not read a sentence the way you do. It breaks the text into tokens, then converts each token, and the passage as a whole, into a long list of numbers called a vector. That vector is a position in a space built so that similar meanings sit close together. This is the literal design goal, not a metaphor: Mikolov and colleagues showed in 2013 that words used in similar contexts end up as neighbouring points, which is why car, automobile and vehicle land near each other without anyone declaring them synonyms. Transformer models such as BERT extended this from single words to whole phrases by reading a passage in both directions at once.

Take the query "cushioned shoes for knee pain when running" and three candidate phrases. Phrase A: trainers with extra shock absorption for joint-friendly jogging. Phrase B: how to build a twelve-week marathon training plan. Phrase C: running shoes fifty percent off this weekend. A meaning-based system places A closest, despite sharing almost no exact words, because it describes the same underlying concept. B shares the topic and even the word training, but answers a different need. C shares the most literal words of the three and is about a discount. That ordering is what the mechanism is designed to produce, described in plain language rather than captured from a live model.

The tactic. You do not need to repeat a target phrase to be found for it. Write the concept in the vocabulary someone actually experiencing it would use: symptoms, comparisons, situations. When you refer to the same thing again, use a different accurate descriptor rather than the exact phrase. Pull that vocabulary from real reviews, support questions and forum threads, not a keyword tool's synonym list, because those are the phrasings a model has genuinely learned to place together.

Similarity is not the same as relevance

Two texts can sit close together in meaning space and still be the wrong answer, because closeness measures topic overlap rather than fitness for what the searcher is trying to do. A hospital's paediatric emergency page and a symptom checklist for worried parents are both, topically, about dehydration in toddlers. Only one of them is what a parent typing at two in the morning wants.

The tactic. Classify the intent behind the query before writing, then pick the page type that intent demands. If a page already ranks for the right topic but is the wrong type for the intent behind it, the fix is a different template and structure, not more paragraphs bolted onto the old one.

Query understanding and fan-out

Google documents that AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources, then assembling the results into one response and surfacing a wider set of supporting pages than a single query would. One question becomes several sub-queries running in parallel, and each can be answered by a different page. For a query like best time to visit Japan with a toddler, the fan-out would plausibly split into season and weather, travel logistics with a small child, family-friendly regions, and typical cost. That is a reconstruction of the documented pattern applied to a sample query, not a captured system log.

The tactic. List every sub-question your topic raises and give each one its own heading with a direct answer in the first sentence beneath it, ordered the way a person would ask them. A page that answers only the headline question is a candidate for one sub-query. A page that answers all of them is a candidate for several.

Passages, not pages

Google names a passage ranking system in its own documentation, describing an AI system that identifies individual sections of a page for relevance separately from the page as a whole. The unit that gets matched and surfaced is often a single paragraph buried on the page. A paragraph that only makes sense after the three above it is invisible to that mechanism.

The tactic. Write each major answer as a stand-alone unit. Restate the question in your own words in the first sentence, answer it immediately, and put the nuance after rather than before. Remove as mentioned, like we said, and unresolved pronouns from any paragraph you want lifted out on its own.

Entities, or things rather than strings

In 2012 Google described the shift behind its Knowledge Graph as things, not strings: matching a query against a real-world entity rather than a sequence of characters. An ambiguous name becomes a retrieval problem under that model, because the system must decide which thing a mention refers to before it can attach authority or facts to it. If a brand's site says one name, its social profile says a second and a directory says a third, with nothing tying them together, reviews and authority scatter across three names instead of consolidating behind one.

The tactic. Use the identical name, spelled the same way, everywhere the business appears. State plainly near the top of your key pages what the entity is, in one declarative sentence, rather than assuming the reader or the system already knows.

What these mechanics make pointless

How we apply this

Once Be Found reads a page the way a retrieval system does rather than the way a person skims it: whether its meaning sits near the queries it should answer, whether its passages stand alone, and whether its entities are named consistently across the web. Where a page ranks for the right topic but is the wrong type for the intent, we change the template and the passage structure before adding any copy. The result is usually fewer pages, each matching more of the sub-questions a topic actually produces.