Semantic Relevance Isn’t One Metric — It Has Three Jobs

Welcome back to our GEO series!

In the previous article, we explored candidate page selection and how AI systems determine which pages are eligible for retrieval.

Semantic relevance can help inform that decision, but its role extends beyond candidate selection. It can serve different purposes throughout retrieval, which makes the advice to simply “improve semantic relevance” less straightforward than it sounds.

In this article, we’ll look at those different roles and why measuring semantic relevance in one place can give an incomplete picture of AI search visibility.

Only got 30 seconds? Here are the key takeaways…

Semantic relevance isn’t a single score

Semantic relevance is often treated as though it can be reduced to a single embedding, a single cosine similarity score or a single judgment of quality.

That simplification may be useful for illustrating the concept, but it doesn’t reflect how relevance appears to be applied throughout a production retrieval system. Where relevance applies matters just as much as how much semantic similarity exists.

Semantic relevance for candidate page selection

At the candidate page selection level, semantic relevance can help answer a relatively broad question:

“Is this page about the query?”

Here, the focus is less on identifying the single best answer and more on establishing whether a page is clearly about the topic being searched for. 

In information retrieval, this is sometimes described as “aboutness”. A page can become eligible for retrieval without containing the strongest answer, as long as it can demonstrate clear topical relevance.

That’s why pages may become weaker candidates when they:

• Cover multiple unrelated subjects.
• Use lengthy introductions that delay the main topic.
• Rely on vague titles and headings that obscure the page’s primary focus.

Semantic relevance for chunk retrieval

Once a page becomes eligible for retrieval, the focus shifts from the page as a whole to the individual passages within it. The question now becomes:

“Which passage best answers this question?”

Here, semantic relevance may become much more fine-grained, with AI systems comparing individual passages rather than the page as a whole. As a result, precision and specificity matter far more than they did during candidate page selection.

This is where details such as entities, units, exact phrasing and clearly defined steps start to matter. They help AI systems identify the passage that most directly answers the user’s query. 

Without that level of precision, a page may pass candidate selection but still fail to surface a relevant passage during chunk retrieval.

Semantic relevance for citation

When it comes to citation, relevance alone may not be enough. A useful passage also needs to provide suitable evidence for the generated response. This means considering:

“Is this chunk safe and precise enough to cite?”

A passage may be semantically similar to the query and contain the right answer, yet still not be used in the final response.

One reason is that the passage can’t stand on its own. Ambiguous wording, missing units, references that depend on surrounding paragraphs, or instructions spread across multiple sections can all make a chunk less suitable for citation.

For a passage to be cited, it needs to be complete enough to be understood in isolation, without relying on additional context elsewhere on the page.

Why a single similarity score can be misleading

Some GEO tools summarize semantic relevance using a single page-level similarity score. While that can be useful as a high-level indicator, it can blur the boundaries between different stages of retrieval.

A page may demonstrate strong topical relevance during candidate selection, yet fail to retrieve the most relevant passage, or retrieve a strong passage that isn’t suitable for citation. A single score makes it difficult to see where in that process content succeeds or falls short.

Reducing those stages to a single score makes it easy to optimize the wrong thing. This helps explain why some pages appear highly relevant yet rarely contribute to AI-generated responses.

A better way to think about semantic relevance

The three roles can be summarized as follows:

Role	What semantic relevance is evaluating	What it optimizes
Candidate Selection	Whether the page is about the query	Topical clarity
Chunk Retrieval	Which passage best answers the question	Answer match
Citation	Whether the passage can stand on its own	Self-contained facts

Practical implications for GEO

These differences have practical implications for how content is structured for GEO. 

Start by making the page’s primary topic clear. Within it, keep sections narrowly focused so individual passages can closely match specific queries.

Those passages also need to make sense independently, without relying on surrounding content to carry essential context.

Each of these serves a different purpose: 

• Topical clarity supports candidate selection.
• Focused answers support chunk retrieval.
• Self-contained passages support citation.

What comes next

Semantic relevance can play different roles depending on whether we’re looking at candidate pages, individual passages, or their suitability for citation. The next article looks more closely at how those passages are structured for use in AI-generated responses.

We’ll explore chunkability, what makes a passage suitable for citation, and how content structure can influence whether information is ultimately used in a generated response.