Determining Relevance: How Similarity Is Scored – Moz

Search engines are basically professional matchmakers. You show up with a vague “best running shoes,”
and they have to figure out whether you meant “best shoes for running,” “best shoes that can run,” or
“best shoes to run away from my responsibilities.” The whole game is relevanceand under the hood,
relevance is often expressed as a similarity score.

Moz popularized this topic for SEOs by translating classic information-retrieval ideas into plain English:
break a query and a page into signals, score their similarity, then combine that with other signals
(quality, authority, freshness, location, etc.) to decide what ranks. This article updates and expands that
idea with modern exampleswhile keeping the spirit: relevance isn’t magic, it’s math plus judgment.

Table of Contents

• What “Relevance” Actually Means
• Lexical Similarity: Matching Words Like a Spreadsheet With Opinions
• TF-IDF: The “Rare Words Matter More” Rule
• BM25: TF-IDF’s More Responsible Older Sibling
• Beyond Bag-of-Words: Proximity, Fields, Entities, and Structure
• Semantic Similarity: Matching Meaning With Vectors
• Hybrid Ranking: When Lexical and Semantic Shake Hands
• SEO Playbook: How to Align With Similarity Scoring
• Common Mistakes (AKA “How to Lose Relevance in 3 Easy Steps”)
• Wrap-Up
• Field Notes: of Real-World Relevance Experiences

What “Relevance” Actually Means

In information retrieval, relevance is the degree to which a document satisfies a user’s information need.
That’s important because the user’s need is not always identical to the words they typed. Someone searching
“apple care cost” might want pricing, plan differences, eligibility rules, or cancellation terms.
The query is the clue; the need is the case.

So similarity scoring usually starts with a simple question: How well does this page match the query?
Then ranking layers on other considerations (credibility, usefulness, freshness, location, speed, and more).
But if you fail the relevance test, the rest doesn’t matterlike showing up to a pizza party with a salad
and insisting “it’s basically pizza if you believe in yourself.”

Lexical Similarity: Matching Words Like a Spreadsheet With Opinions

The earliest (and still very influential) approach to similarity is lexical matching:
compare the terms in the query to the terms in the document. You tokenize, normalize (case-folding,
stemming/lemmatization sometimes), and then compute a score based on overlap.

Why lexical signals still matter in 2026

Even with modern AI systems, search still needs reliable, fast, interpretable matchingespecially for:
navigational queries (“IRS refund tracker”), product queries (“noise canceling headphones”), and exact
names (“Form 1040 Schedule C”). Lexical scoring is also great at precision: if the query includes “2026,”
pages that actually contain “2026” often deserve a boost.

A tiny example: matching “best car insurance”

Suppose your query is “best car insurance.” A lexical model will reward pages that mention those terms,
especially in important locations (title, headings) and in close proximity (the words appear near each other
rather than scattered across a 4,000-word epic poem about vehicles).

TF-IDF: The “Rare Words Matter More” Rule

TF-IDF stands for term frequency (how often a word appears in a document) and
inverse document frequency (how rare that word is across the whole collection).
The intuition is wonderfully human: common words (“the,” “and,” “website”) don’t tell you much,
but distinctive words (“thoracic,” “Medicare,” “deductible”) carry meaning.

What TF-IDF gets right

• Specific terms are informative. If “deductible” appears, you’re probably in insurance territory.
• Repetition helps… to a point. Mentioning “Medicare representative” once is a hint; five times might be relevant; fifty times is a cry for help.
• It’s fast. Search engines can compute and rank at web scale.

What TF-IDF gets wrong (or doesn’t even try to do)

• It can be tricked by spammy repetition (hello, keyword stuffing).
• It doesn’t understand synonyms (“attorney” vs. “lawyer”).
• It ignores intent unless you engineer additional signals.

BM25: TF-IDF’s More Responsible Older Sibling

If TF-IDF is a talented intern, BM25 is the manager who learned boundaries.
BM25 is a probabilistic scoring model that improves on TF-IDF by handling two big issues better:
term saturation (the 50th occurrence of a word shouldn’t be worth the same as the 1st)
and document length normalization (a 10,000-word page naturally contains more wordsdon’t reward it just for being long).

BM25 in plain English

BM25 still likes frequent query terms and still rewards rare terms, but it tempers the score so you can’t
win purely by repeating the keyword or inflating length. In many modern search stacks (including popular
open-source engines), BM25 is the default baseline for “text relevance.”

A practical SEO takeaway

If your content only “wins” because you repeated a phrase 27 times, BM25-style scoring tends to shrug.
What helps more is: clear topical focus, the right supporting terms, and smart structure
(so the engine can confidently map your page to the query’s concept).

Beyond Bag-of-Words: Proximity, Fields, Entities, and Structure

Similarity scoring rarely treats the whole page as one blob of text. Most retrieval systems score across
fields (title, headings, body, anchor text) and may apply different weights. Why?
Because “best running shoes” in a title is a strong signal; buried once in the footer is… less romantic.

Common similarity boosters (that aren’t “stuff more keywords”)

• Term proximity: “running shoes for flat feet” close together often beats the same words scattered.
• Field weighting: titles and H1s often matter more than paragraph #37.
• Phrase matching: exact sequences can be rewarded for certain queries.
• Anchor text context: how other pages describe yours can reinforce topic alignment.
• Structured hints: clear headings, descriptive subheads, and consistent terminology reduce ambiguity.

This is one reason “clean SEO” works: not because headings are magical, but because they make your topical
signals legible to both humans and retrieval models.

Semantic Similarity: Matching Meaning With Vectors

Lexical matching asks: “Do these words overlap?” Semantic matching asks: “Do these meanings align?”
Modern systems often represent text as vectors (embeddings) and compute similarity using metrics like
cosine similarity or dot product.

Why vectors changed the game

With embeddings, “how to choose a health plan” can match a page about “selecting coverage options”
even if the wording differs. That’s huge for long-tail queries, conversational phrasing, and
“I can’t remember the exact term but I know what I mean” searcheswhich is most of humanity, honestly.

Cosine similarity vs. dot product (no, you don’t need a PhD)

Cosine similarity measures the angle between vectors (direction), ignoring magnitude. Dot product considers
both alignment and magnitude. Different systems pick different metrics depending on how embeddings are trained
and how they want scores to behave. The practical SEO point is not “optimize for cosine,” but:
write so your page’s meaning is unambiguous and richly connected to the topic.

Hybrid Ranking: When Lexical and Semantic Shake Hands

A common modern approach is hybrid search: combine BM25-like lexical relevance with
vector-based semantic similarity. Lexical helps with precision and exact terms; semantic helps with intent,
paraphrases, and concept-level matching. Then a re-ranker (sometimes machine-learned) can blend signals into
the final ordering.

Translation for content creators: you can’t “trick” relevance anymore with a single tactic.
You have to match both the language people use and the meaning they intend.

SEO Playbook: How to Align With Similarity Scoring

Let’s turn theory into action. If you want to rank, your job is to send strong, consistent signals that your
page is the best match for a query’s intent. Here’s how to do that without sounding like a robot reading a
keyword list at gunpoint.

1) Start with intent, then choose the “core phrasing”

Pick one primary query theme (your “head term”) and define the intent behind it:
informational, navigational, commercial investigation, or transactional.
Then support it with natural variations (LSI-style related terms) that humans use.

2) Use a relevance-friendly structure

• Title: include the core topic in a natural way.
• H1: mirror the promise of the title, not a random poetic remix.
• H2/H3s: map to sub-questions people actually have (cost, steps, pros/cons, examples).
• First 200 words: clarify what the page is and who it’s for (helps meaning matching).

3) Add “supporting vocabulary” like a pro

Similarity scoring improves when the page contains concept neighborsterms that typically appear in
relevant documents. For relevance scoring, this works like corroborating evidence. For humans, it reads like
you know what you’re talking about. Everybody wins.

Example: For “similarity scoring,” supporting vocabulary might include TF-IDF, BM25, term frequency, inverse document frequency,
query intent, vector embeddings, cosine similarity, proximity, and field weighting.

4) Demonstrate usefulness (because ranking is not only similarity)

Relevance gets you invited to the party. Usefulness gets you asked to stay.
Add concrete examples, step-by-step explanations, FAQs, visuals (where appropriate), and clear definitions.
Many ranking systems also consider overall quality signalsso thin content that’s “technically on-topic”
but unhelpful may not stick.

5) Avoid accidental ambiguity

Ambiguity is a relevance tax. If your page title is “Similarity Scoring Explained,” but the content is
actually about “social media engagement metrics,” search systems will see mixed signals. Humans will too.
Mixed signals lead to mixed rankings, like mixed metaphors lead to mixed… never mind.

Common Mistakes (AKA “How to Lose Relevance in 3 Easy Steps”)

Stuffing keywords and hoping BM25 doesn’t notice

Modern scoring models handle repetition with diminishing returns. If your strategy is “repeat the phrase until
it becomes true,” you’re basically doing wishful indexing.

Writing one page to rank for fifteen different intents

A single page can cover a topic comprehensively, but it can’t be everything to everyone. If you try to target:
“what is BM25,” “BM25 formula,” “Elasticsearch tuning,” “Google ranking secrets,” and “best pizza near me,”
the similarity signals blur. Choose a lane, then cover it deeply.

Ignoring the SERP reality check

If the top results for a query are calculators, comparison tables, or official documentation, and you publish
a 2,000-word narrative essay about your feelings, similarity scoring won’t save you.
The engine has learned what satisfies that query.

Wrap-Up

Moz’s core lesson holds up: search engines need a way to score “how well this matches,” and that usually starts
with similaritylexical signals like TF-IDF/BM25 and, increasingly, semantic signals from embeddings.
The best SEO strategy is not to chase one scoring trick, but to build pages that are
clearly about the thing, useful for the intent, and structured so machines can verify it fast.

Sources synthesized (no links, just receipts)

• Moz (Similarity scoring & relevance concept popularization for SEOs)
• Google Search Central (ranking systems, how Search works)
• Microsoft Bing (Webmaster Guidelines, Search Quality insights)
• Stanford University (Information Retrieval textbook & lectures)
• Apache Lucene (Similarity and BM25Similarity scoring components)
• Microsoft Learn (TF-IDF/BM25-style relevance scoring in search services)
• AWS OpenSearch Service docs (BM25 baseline & learning-to-rank)
• OpenSearch documentation (BM25, similarity configuration, hybrid search)
• NIST TREC (relevance judgments and IR evaluation foundations)
• Pinecone (vector similarity metrics)
• Redis (vector similarity metrics and practical considerations)
• Algolia (ranking criteria like proximity and attribute weighting)
• Elastic (BM25 explainability and scoring behavior)

Field Notes: of Real-World Relevance Experiences

The most useful “relevance lessons” rarely come from reading a formula (although formulas are fun if you’re into
that sort of thing). They come from watching what happens when content meets the real worldmessy queries, weird
wording, and users who type like they’re texting while riding a roller coaster.

One pattern that shows up again and again in published SEO case studies and practitioner experiments is the
“title-body mismatch penalty.” A page might have a beautifully optimized titlesomething like
“How Similarity Scoring Works (TF-IDF, BM25, and Beyond)”but the body spends half its time on unrelated
tangents (company history, product plugs, or a motivational speech about hustle). In lexical scoring terms,
the title may win a little, but BM25-like systems and modern re-rankers quickly notice the body doesn’t carry
the same topical density. Users notice too. The fix is boring but effective: outline first, make every section
earn its spot, and keep the page’s “about-ness” consistent.

Another common experience: relevance improves dramatically when writers stop thinking in single keywords and
start thinking in supporting term clusters. For example, content targeting “BM25” that also
explains term frequency, inverse document frequency, length normalization, saturation, and field boosts
tends to perform better than content that repeats “BM25” like it’s trying to summon it. This is not mystical.
It’s what similarity scoring rewards: the presence of concepts that co-occur in truly relevant documents.
And it’s what humans reward: you answered the question thoroughly.

A third lesson: “semantic SEO” gets misunderstood. People hear “vectors” and assume the move is to jam synonyms
into a page like you’re stuffing a suitcase. In reality, semantic matching tends to improve when you add
clarifying context, not random variations. A page about “Medicare representative appointment”
becomes easier to match semantically when it includes context like authorized representative, SSA forms,
Medicare claims, appeals, and beneficiary permissionsthe real-world concepts connected to that intent.
The page’s meaning becomes sharper, not fuzzier.

Finally, a practical “relevance debugging” habit from the trenches: run a simple before/after test using your
own search feature (if you have one) or a controlled content change. Pick a handful of queries, record current
rankings/clicks, adjust only one thing (e.g., reorganize headings to mirror user sub-questions, tighten the
introduction so intent is explicit, add a missing definitions section), then measure again. Even when you can’t
see Google’s internal scoring, you can see the outcomes: improved engagement, fewer pogo-sticks back to the SERP,
and better alignment with query intent. Relevance isn’t just a scoreit’s observable behavior.

The punchline is reassuring: you don’t need to reverse-engineer every weighting knob. If your page reads like
the best answer and is structured like the best answer, similarity scoring usually follows. And if you’re ever
tempted to keyword-stuff, just remember: BM25 has boundaries, and so should you.