LSI Keywords Are Dead — Here's What Actually Works Now

"Add LSI keywords to your article" is advice that circulates in SEO communities constantly. It sounds sophisticated. It refers to a real technique — Latent Semantic Indexing — that was developed in 1988. The problem is that Google has stated directly that it doesn't use LSI, and never did. The keyword lists that "LSI keyword tools" generate are synonyms and related terms, which is useful, but the rationale behind them is wrong.

Here's what's actually happening when Google evaluates whether your content covers a topic, and what that means for how you write.

What LSI actually is

Latent Semantic Indexing is a document-retrieval technique developed in the 1980s. It uses singular value decomposition — a matrix math operation — to identify co-occurring terms in a corpus of documents. In a simplified example: if the words "bat," "pitcher," "innings," and "outfield" frequently appear together in documents, LSI infers they're related to "baseball."

This was genuinely useful for document retrieval systems in the pre-web era. Libraries used it. Early search engines experimented with it. It addressed the problem of keyword mismatch — a document about "automobiles" could be retrieved for a search for "cars" if the system had learned the terms co-occurred.

The problem is computational. LSI doesn't scale. Processing a corpus the size of the web with SVD decomposition is computationally prohibitive in a way that made it unusable for Google from the beginning.

Google's statement

In 2019, Google's John Mueller addressed LSI keywords directly on Twitter:

"There's no such thing as LSI keywords — anyone who's telling you otherwise is mistaken, sorry. (...) We look at the words on a page, but there's no magic list of words we need to see on a page."

This isn't ambiguous. LSI keywords, as a concept applied to Google's algorithm, are a myth. The tools that generate "LSI keyword" lists are generating related terms — which is useful — but the reason they give for why those terms matter is wrong.

What Google actually does

Google's understanding of text has gone through several distinct phases:

Pre-2013 — Keyword matching. Google primarily matched query keywords to document keywords. Synonyms were handled crudely. This is the era that produced keyword stuffing as a tactic.

2013 — Hummingbird. Google's first major step toward semantic understanding. The update improved handling of conversational queries and concept-level matching, moving beyond exact keyword matching.

2015 — RankBrain. Google began using machine learning to handle novel queries it hadn't seen before. RankBrain embeds queries and documents in a vector space and measures similarity. This is the beginning of embedding-based relevance.

2019 — BERT. Bidirectional Encoder Representations from Transformers. BERT understands context and nuance in natural language far better than any previous approach. It reads sentences in both directions, capturing how words modify each other. "How to treat a patient nurse" and "how to nurse a patient back to health" have completely different meanings; BERT handles both correctly.

2021+ — MUM and beyond. Google's models became multimodal (text, images, video) and increasingly capable of understanding topic relationships at a knowledge-graph level.

The practical result: Google doesn't match keywords. It matches meaning. A page about electric vehicles doesn't need to say "EV" to be relevant for the query "EV range anxiety" — if the page covers the topic thoroughly, Google's model will identify it as relevant.

What "semantic SEO" actually means

The phrase "semantic keywords" in modern SEO refers to terms and concepts that signal topic depth to Google's embedding-based models. These are not magic words you need to include. They're the natural vocabulary of a topic that a thorough, expert treatment would naturally use.

If you write a complete article about content marketing, you'll naturally mention "editorial calendar," "content distribution," "target audience," "content audit," "blog posts," and "conversion rate" — not because you're trying to include semantic keywords, but because those are concepts that belong in a complete treatment of the topic.

The difference from LSI: you're not looking for statistically co-occurring terms in a fixed corpus. You're looking for the conceptual vocabulary of a topic as understood by practitioners.

The practical workflow for semantic coverage

Step 1 — Read the top 10 SERP results for your target keyword. Open each page and skim for subheadings (H2s, H3s) and frequently repeated terms. These are the concepts Google has determined belong in a comprehensive treatment of the topic.

Step 2 — Note concepts and subtopics that appear across multiple top results. If 6 of the 10 top results address a specific subtopic, and your draft doesn't, that's a coverage gap. Not because of keyword density, but because the topic is incomplete without that concept.

Step 3 — Write the missing concepts into your content. Don't insert keywords — write the concept. A paragraph that explains the idea, gives an example, and answers the implicit question it raises is more valuable than a sentence that mentions the term.

Step 4 — Check your draft for unnatural repetition. Google's models are good at distinguishing natural topical coverage from keyword stuffing. Write naturally; optimize for meaning, not frequency.

What "LSI keyword" tools are actually generating

The tools that market themselves as "LSI keyword generators" are generating related terms using TF-IDF analysis, co-occurrence in Google's top results, or synonym databases. The output is often useful. The problem is the framing.

When a tool tells you to add "feline" to your article about "cats" because it's an "LSI keyword," it's giving you a related synonym. Using it won't hurt your content. But the reason it tells you this — that Google uses LSI — is wrong, and believing that wrong reason leads to bad decisions.

The bad decision pattern: treating keyword insertion as the goal rather than topic coverage. You can hit every word on an "LSI keyword list" and still write a thin, low-value piece of content that doesn't rank. Conversely, you can write a thorough, expert article that naturally uses all the relevant terminology without ever consulting a keyword list, and it will rank.

The keyword density question

Because Google doesn't use keyword frequency thresholds as a direct ranking signal, there's no magic keyword density to hit. What matters:

• The primary keyword should appear in the title, the first 100 words, and naturally throughout the content.
• Related terms and concepts should appear as naturally covered topics, not inserted phrases.
• Over-optimization (cramming a keyword into every other sentence) creates unnatural writing that both readers and Google's models penalize.

RankCrab's keyword density analyzer identifies your current keyword frequency and flags over-optimization risks. Use it as a sanity check — not as a target to hit.

What to do instead of LSI keyword research

The reliable process for semantic coverage:

1. Identify your target keyword and open the SERP.
2. Read the top 10 results' H2 subheadings — these are the subtopics Google expects.
3. Use People Also Ask to find related questions your content should address.
4. Write complete coverage of the topic, addressing each major subtopic with genuine depth.
5. After drafting, check density to confirm you haven't over- or under-optimized the primary keyword.

This produces content that ranks on the basis of actual quality and topical completeness — which is what Google is trying to reward. For the full keyword research framework, including intent analysis and page mapping, see the keyword research hub.