Embeddings and Vector Search Explained for SEOs

When SEOs hear the words “embeddings” and “vector search,” many switch off. It sounds like something for data scientists, not for people who write title tags and fix internal links. I understand that. But I think it is a mistake to ignore it.

Embeddings are the reason Google can match a page about a “dripping tap” to someone searching “leaky faucet,” and the reason ChatGPT can pull one useful paragraph from a long article. They are also a practical tool you can use today to cluster keywords, find internal link opportunities and map redirects, often without writing code.

In this guide, I will explain embeddings and vector similarity in plain words, show how Google and AI assistants use them, and then walk you through the practical SEO uses, with a worked example and the limits you should know about.

What an embedding is, without the maths

An embedding is a way of turning text into a list of numbers so a computer can compare meanings. OpenAI’s documentation gives the shortest definition I know:

“An embedding is a vector (list) of floating point numbers.”

— OpenAI, Vector embeddings guide

Let me explain with a simple picture. Imagine a map of a city. Every place has two coordinates, and places that are close on the map are close in real life. An embedding does the same for meaning, except the “map” has hundreds or thousands of directions instead of two. Texts with similar meaning end up close together. Texts about different things end up far apart.

A few facts make this more concrete:

• Each number is one “direction” of meaning. No single number means “price” or “plumbing.” Meaning is spread across all of them.
• Modern models produce long lists. OpenAI’s documentation lists 1,536 dimensions by default for text-embedding-3-small and 3,072 for text-embedding-3-large. Google’s Gemini embeddings documentation also defaults to 3,072 and lets you choose smaller sizes from 128 upward.
• The same model must be used for everything you compare. Numbers from two different models live on different “maps” and can’t be compared.
• You can embed almost anything. A keyword, a paragraph, a whole page, or with newer multimodal models, images too.

If you already think about topics and meaning rather than exact words, you are halfway there. My guide to semantic SEO covers the writing side. This guide covers the machinery underneath.

How vector similarity is measured, with a tiny example

Once text is turned into vectors, a system needs a way to say how close two of them are. The most common method is cosine similarity. In plain words, it measures whether two arrows point in the same direction. A score near 1 means very similar meaning. A score near 0 means unrelated.

Here is a toy example with only three dimensions, so you can follow the maths. Real embeddings have thousands of dimensions, but the idea is the same. (These vectors are invented for illustration.)

Text	Vector (made-up)
A: “fix a leaky faucet”	(0.8, 0.6, 0.0)
B: “repair a dripping tap”	(0.6, 0.8, 0.0)
C: “best pizza in Chicago”	(0.0, 0.2, 0.98)

For vectors with a length of about 1, cosine similarity is just the sum of the matching numbers multiplied together:

• A and B: (0.8 × 0.6) + (0.6 × 0.8) + (0.0 × 0.0) = 0.48 + 0.48 + 0 = 0.96. Very similar, even though they share almost no words.
• A and C: (0.8 × 0.0) + (0.6 × 0.2) + (0.0 × 0.98) = 0 + 0.12 + 0 = 0.12. Unrelated.

This is the whole trick. Ahrefs gives a real-world version in its guide to semantic search, where “How to fix a leaky faucet” and “repairing dripping tap” score about 0.89 despite sharing few words.

One practical warning: scores are only comparable within one model. A 0.8 from one model is not the same as a 0.8 from another. Always set your thresholds by looking at real examples from your own data.

How Google Search has used meaning-based matching since 2015

Google has been moving from word matching to meaning matching for over a decade. In a February 2022 post, Pandu Nayak, Google’s VP of Search, described the main systems:

System	Launched	What it does, per Google
RankBrain	2015	Helps match words to related concepts; used to rank results
Neural matching	2018	Looks at broader concepts in a whole query or page to retrieve relevant documents
BERT	2019	Understands how combinations of words express meaning; used in ranking and retrieval
MUM	Announced 2021	Multitask, multilingual model; at the time not used for general ranking like the others

“When we launched RankBrain in 2015, it was the first deep learning system deployed in Search.”

— Pandu Nayak, Google, How AI powers great search results

Google doesn’t publish exactly how embeddings fit into its ranking today. But its Cloud team has been more open about the technology behind its enterprise search product. A February 2024 Google Cloud post says Vertex AI Search uses the same RankBrain and neural matching processes as Google Search to create query and document embeddings. It also describes hybrid search: keyword search and semantic search run at the same time, and their results are merged and re-ranked.

The speed comes from approximate nearest neighbor search. Comparing a query against every stored vector is slow, so systems group vectors and only search the most promising groups. A 2021 Google Cloud post explains that Google’s ScaNN technology powers this kind of search and that the same backend serves Google Image Search, YouTube and Google Play.

What this means for you: Google can find your page for queries that don’t contain your exact words, but keyword signals still matter because they are part of a hybrid system. You need both clear wording and genuine topical depth.

How AI assistants pull passages with vector search

AI assistants that search the web, and most chatbots built on company documents, use a method called retrieval-augmented generation, or RAG. I explain it fully in my guide to RAG, but here is the short version as it relates to embeddings:

1. Documents are split into smaller pieces, often called chunks or passages.
2. Each piece is turned into an embedding and stored in a vector index.
3. When a question arrives, it is embedded with the same model.
4. The system finds the pieces whose vectors are closest to the question’s vector, often combined with keyword search.
5. A re-ranking step picks the best pieces, and the language model writes an answer from them.

Google’s Gemini documentation shows how specific this has become. Its older text model accepts task types such as RETRIEVAL_QUERY, RETRIEVAL_DOCUMENT, QUESTION_ANSWERING and FACT_VERIFICATION, because a question and the passage that answers it are not worded alike. The model is tuned to bring them together anyway.

Two practical lessons follow. First, passages that answer one question clearly are easier to match than paragraphs that wander across several ideas. Second, AI assistants often break one question into several searches. I cover that in my piece on ChatGPT query fan-out, and it means your page may be retrieved for a sub-question you never targeted directly.

A word of balance: Google’s 2026 guide to its AI features says you don’t need to “chunk” content specially for AI. I agree. Write clear sections for humans, and you will usually get retrievable passages as a side effect. My guide on content engineering covers structure in more depth.

Keyword matching vs semantic matching: where each still wins

Embeddings did not make keywords irrelevant. Each approach has strengths, which is why real systems combine them.

Situation	Keyword matching	Semantic (vector) matching
Synonyms (“sofa” vs “couch”)	Misses unless both words appear	Matches well
Exact product codes, model numbers, names	Strong	Can blur similar codes together
Long, conversational questions	Weak	Strong
Rare jargon or new terms	Strong if the term is on the page	Weak if the model hasn’t learned the term
Negation (“not gluten-free”)	Weak	Also often weak
Spelling variants and typos	Weak	Usually better

So keep using the words your customers use, especially for product names, places and specific services. Then make sure the page covers the topic well enough to match related questions too. My guide on entity SEO explains how named things fit into this picture.

Practical SEO uses: clustering, internal links, redirects and gaps

This is where embeddings stop being theory. Here are the uses I find most valuable for everyday SEO work.

Keyword clustering

Embed every keyword in your research list, then group keywords whose vectors are close. Each group becomes one page target. This is faster than grouping by hand and catches synonyms that word-based grouping misses. Always sanity-check against search results, though, because two keywords can mean the same thing yet need different pages if Google shows different results. My guide to keyword clustering covers the SERP-based method to use alongside this.

Internal link suggestions

Embed every page on your site. For each page, list the five pages with the highest similarity that it doesn’t already link to. Those are your strongest internal link candidates. I explain how to place and word those links in my guide to internal linking for SEO.

Redirect mapping during a migration

Embed old and new pages and match each old URL to its closest new URL. Screaming Frog publishes a tutorial for this method and is honest that results should not be trusted blindly. Review the lowest scores by hand. My website migration checklist covers the rest of the process.

Cannibalization and near-duplicates

Pages with very high similarity to each other may compete for the same queries. They are candidates to merge, differentiate or canonicalize.

Off-topic pages and content gaps

Pages far from the center of your site’s topics may be off-brand or weakly connected. And if you embed competitor pages too, topics where they have clusters and you have nothing are gaps worth reviewing.

Here is a step-by-step version using Screaming Frog’s SEO Spider, based on its semantic similarity tutorial:

1. Under Config > API Access > AI, choose a provider (OpenAI, Gemini or Ollama) and add your API key. A paid SEO Spider licence is required.
2. In the prompt configuration, add the embeddings preset from the library and connect.
3. Under Config > Spider > Extraction, enable Store HTML and Store Rendered HTML.
4. Under Config > Content > Embeddings, enable embeddings, then Semantic Similarity and Low Relevance.
5. Crawl the site and let both the crawl and API progress bars finish.
6. Run Crawl Analysis to fill the Semantically Similar and Low Relevance filters.
7. Review the Content tab and the content cluster diagram under Visualisations.

“Pages scoring above 0.95 are considered semantically similar by default.”

— Screaming Frog, How to Identify Semantically Similar Pages & Outliers

The tutorial lists a default low-relevance threshold of 0.4. Both thresholds are adjustable, and I suggest you tune them after reviewing a sample of real pairs from your site.

A worked example: choosing internal links from similarity scores

Here is a hypothetical example. A home services site has a new page, “How to choose a tankless water heater.” After embedding all 120 pages, the top matches it doesn’t already link to are:

Candidate page	Similarity (example)	Decision
Tankless water heater installation cost	0.91	Link: natural next step for the reader
Tank vs tankless water heaters	0.89	Link: and check it isn’t cannibalizing the new page
Water heater repair service page	0.82	Link: commercial page that needs internal authority
How to flush a water heater	0.78	Link from the maintenance section only
Boiler repair service page	0.64	Skip: related but not what this reader needs
Company news: new office	0.21	Skip

Two lessons from this example. First, a high score suggests a link but doesn’t decide it. The 0.89 page might be too close and compete with the new page, so it deserves a manual check. Second, the commercial page at 0.82 may be the most valuable link for the business even though it isn’t the top score. Similarity tells you what is related. You decide what is useful.

If you do this for all 120 pages and keep the top three to five links per page, you end up with a few hundred suggestions. A sensible threshold filters out the obvious mismatches first, so the person reviewing the list only checks the borderline cases.

Tools for trying embeddings without heavy coding

Option	What it involves	Good for
Screaming Frog SEO Spider with an AI provider	Paid licence plus an OpenAI, Gemini or Ollama API key; menu-based setup	Site-wide similarity, outliers, redirect mapping
OpenAI embeddings API	An API key and a short script or spreadsheet add-on	Keyword clustering, custom similarity tables
Gemini embeddings API	An API key; supports task types and smaller output sizes	Retrieval tests, clustering, classification
Ollama (local models)	Runs models on your own computer	Privacy-sensitive content; no per-call API cost

Whichever you choose, keep a simple log of which model and settings you used. If you switch models later, you must re-embed everything.

Limits and misconceptions about embeddings in SEO

Embeddings are useful, but they are often oversold. Here is what I would keep in mind:

• Your scores are not Google’s scores. You use a public model. Google uses its own systems and many other signals. A high similarity between your page and a keyword does not mean you will rank.
• Similar is not the same as helpful. A page can be semantically close to a query and still be thin, outdated or untrustworthy.
• Stuffing related terms doesn’t “raise your vector.” Adding a list of synonyms makes a page worse for readers and does little for meaning-based matching.
• Long pages get averaged. One embedding for a 5,000-word page blends many topics. For analysis, embed sections or use the main content only.
• Thresholds are judgment calls. 0.95 might be right for duplicates on one site and wrong on another.
• Retrieval is only the first step. In AI answers, being retrieved doesn’t guarantee being cited or named. I explain what happens next in my article on how AI answers flatten original ideas.

Questions SEOs ask about embeddings and vector search

Do I need to know how to code to use embeddings?

No. Crawlers like Screaming Frog let you generate embeddings through menus once you add an API key. Coding helps for custom clustering, but it isn’t required to get started.

Is semantic search the same as vector search?

Not quite. Semantic search is the goal: matching meaning. Vector search is one technique to achieve it, by comparing embeddings. Real search engines combine it with keyword matching and many other signals.

What similarity score means two pages are duplicates?

There is no universal number. Screaming Frog uses 0.95 as its default for “semantically similar.” Check a sample of pairs from your own site and adjust.

Should I rewrite my content to improve its embeddings?

No. Write clearly, answer the question directly and cover the topic properly. Those habits naturally produce content that matches well. Google’s 2026 AI guide also says you don’t need to rewrite content just for AI.

Does Google use the same embeddings as OpenAI or Gemini APIs?

Google does not publish the models behind its Search ranking. Public embedding models are a useful approximation for analysis, not a copy of Google’s systems.

Putting embeddings to work on your site this month

In brief: embeddings turn text into numbers that capture meaning, and similarity scores tell you which texts are close. Google has used meaning-based matching since RankBrain in 2015, and AI assistants rely on vector retrieval to find passages. For SEO, start with one practical use, such as internal link suggestions or a redirect map, and always review the results with human judgment.

If you would like us to run a semantic analysis of your site, from clustering and internal linking to finding cannibalization, our SEO team can help. Request a free quote and tell us what you want to fix first.

Further reading on embeddings vector search

• How to Track Where Your YouTube Videos Get Mentioned and Cited
• How ChatGPT Turns One Prompt Into Many Web Searches
• Proving AI Search Impact With Holdout and Pre/Post Tests

Sources for this embeddings and vector search guide

1. Ahrefs: Semantic Search Is the Only Search That Matters Now
2. OpenAI: Vector embeddings guide
3. Google AI for Developers: Gemini API embeddings
4. Google: How AI powers great search results (Pandu Nayak, 2022)
5. Google Cloud: RAGs powered by Google Search technology, part 1
6. Google Cloud: Find anything blazingly fast with Google’s vector search technology
7. Screaming Frog: How to identify semantically similar pages and outliers
8. Screaming Frog: How to use vector embeddings for redirect mapping
9. Google Search Central: Optimizing for generative AI features