Semantic Match

Why does the obviously-better page sometimes lose to a worse one?

A brand can have the strongest page in its category and still be passed over — because the engine was never matching the words on it.

For twenty years, being found meant containing the words someone searched. Pack the right keywords onto a page, and it surfaced. That reflex is now close to useless: Princeton’s GEO study found that repeating keywords did nothing for how often AI quoted a page, and in places performed slightly worse than leaving it alone.

What replaced it is a match on meaning. An AI engine reads the question for what it is asking, not which words it uses, and looks for the passage that answers that — even if the page and the prompt share almost no vocabulary. The better page loses when it answers a slightly different question than the one that was asked. We call the work of closing that gap Semantic Match: being the page that answers the question behind the prompt.

What does an engine actually do with a prompt?

It turns both the question and your page into math, then measures distance.

As Anthropic describes in its work on contextual retrieval, a page is split into passages and each is converted into a vector — a long string of numbers that stands in for its meaning. The prompt is converted the same way. To answer, the engine pulls back the passages whose vectors sit closest to the prompt’s, then quotes from those.

Two things follow from that. First, vocabulary overlap barely matters; a passage can win on meaning while sharing few words with the question. Second, a passage that only makes sense in the context of the whole page tends to lose, because once it is lifted out it no longer sits close to anything. The unit that gets retrieved is the self-contained passage, not the page.

What is the question behind the question?

Every prompt carries an intent the words only partly state — and matching that intent is the whole game.

“Best CRM for a small agency” is not really a request for a list of CRMs. It is a request for the one that fits a small agency’s budget, team size, and workflow — stated by someone who does not want to read ten reviews to find out. A page that lists every CRM answers the words. A page that says which one suits a small agency, and why, answers the intent.

Same category, different visitors — each prompt is a different intent, and needs its own answer

The same category, then, is not one question but many — comparisons, use-cases, budget cases, the “is it worth it” cases. Each is a different question being asked. A page wins one only when its meaning is sized to the exact question in front of it.

Why can’t one page answer every prompt?

Because a passage sized to one intent sits far, in meaning, from a different intent — so a single page cannot be near all of them at once.

This is the quiet limit behind the “one big page for everything” instinct. A sprawling page that gestures at comparisons, pricing, and use-cases all at once has no passage that sits cleanly closest to any one of them. It is impressive from a distance and useless to the buyer asking one precise thing. The brands that win the long tail of prompts tend to have a distinct, answer-first page per real intent — not more pages for their own sake, but the right page for the right question.

Finding those intents is reporting, not guessing. Ask the engines the questions buyers actually type, note which prompts return a competitor and which return no one, and build the page where the gap is widest. That is where Semantic Match starts: with the real prompts, not the ones a brand wishes it were asked.

How do you build for Semantic Match without losing the reader?

Lead each page with a self-contained answer to its specific question, then earn it with the voice and proof a person reads for.

The mechanics are the same ones that help any careful reader. Princeton’s study found that adding citable statistics lifted visibility about 41% and direct quotations about 28% — the specific, attributable substance that also makes a passage easy to lift and verify. State the answer plainly in the first line of the section; develop it underneath. That is the discipline we call the Dual Legibility Tension, applied at the level of a single page.

Semantic Match is one of the five signals of Narrative Architecture — the one that carries the intent load. It does not stand alone: a page that perfectly matches a prompt still needs a true claim (Authority) and independent confirmation (Corroboration) behind it. The broader practice of organising all of this to win the answer is GEO. SOLEDAD, our engine, surfaces the prompts where a brand is absent and builds the pages that close them — with a human approving every one. With roughly two-thirds of searches ending without a click, the page that matches the question is increasingly the only one a buyer ever reaches.

Frequently asked questions