Latent semantic indexing in SEO: the real technique, the LSI keyword claim, and what to do instead

Latent semantic indexing is a real information retrieval technique from the late 1980s, and "LSI keywords" is an SEO term that was built on top of it and does not describe anything Google does. The technique was patented as US4839853A, "Computer information retrieval using latent semantic structure", filed on 15 September 1988 and published on 13 June 1989, assigned to Bell Communications Research Inc, with Deerwester, Dumais, Furnas, Harshman and Landauer named as inventors. It works by applying a matrix decomposition to a fixed collection of documents to find statistical relationships between words. Google's Search Advocate John Mueller stated publicly in 2019 that there is no such thing as LSI keywords. The useful part of this story is that the thing people were reaching for when they said LSI is real, important and better understood under its own name, which is semantic and entity based retrieval. This page separates the three.

What latent semantic indexing actually is

Take a collection of documents and build a matrix of which terms appear in which documents. Apply singular value decomposition to reduce that matrix to a smaller number of dimensions. Words that tend to occur in the same documents end up near each other in the reduced space, so a query for "car" can return a document about "automobile" even where the exact word never appears. The patent abstract describes it as presuming "an underlying, latent semantic structure in the usage of words in the data objects".

That was a significant advance for its time and it still has uses in closed collections: a legal archive, a corporate document store, a journal index. Three properties explain why it was never the basis of web search.

• It needs the whole collection at once. The decomposition is computed over the full term document matrix, so a collection that grows constantly means recomputing constantly.
• It scales badly. The computation grows steeply with the number of documents and terms, and the web is measured in hundreds of billions of pages.
• It has no notion of an entity. It finds statistical co-occurrence, not that Paris is a city and Apple is both a company and a fruit. Modern retrieval needs the second kind of knowledge.

Where "LSI keywords" came from

Search engines plainly understood synonyms and related concepts long before anybody could explain how. The SEO industry needed a name for the practice of including related vocabulary on a page, borrowed the most academic sounding term available, and "LSI keywords" was born. Tools appeared that sold lists of them. The advice attached to it, which was to include terms related to your topic, is reasonable advice. The mechanism it claimed was not the mechanism.

This matters for one practical reason. Believing in LSI keywords leads people to sprinkle a list of related terms into a page in the hope of satisfying a statistical model. Understanding what actually happens leads people to cover the topic properly, which produces the related vocabulary naturally and also answers the reader. Same vocabulary, very different page.

What search engines actually do instead

What people meant by LSI	What the real mechanism is called	What it changes for you
The engine understands synonyms	Query understanding, including synonym expansion	Write for the reader. You do not need every variant on the page
The engine knows related concepts	Entity recognition and a knowledge graph	Name things precisely, and connect them to known identifiers
The engine reads meaning, not strings	Neural language models applied to queries and passages	Clear, complete prose beats keyword density
The engine groups related pages	Topical coverage across a site, and internal links	Build a set of pages that fully covers a subject
Related terms help ranking	Covering the sub questions a searcher has	Answer the whole question, including the parts competitors skip

What to do on the page instead of chasing LSI keywords

1. Answer the question in the first paragraph. Passage level matching means the clearest paragraph is the one that gets used, so put the answer where it can be found rather than building up to it.
2. Cover the sub questions. Take the questions people ask around the main one and answer each in its own section. This is what produces real topical depth, and the related vocabulary arrives with it.
3. Name entities precisely. Full product names, full place names, the organisation's registered name, the standard's reference number. Precision is what lets a machine resolve what you are talking about.
4. Use the words your readers use, including the informal version. You are not feeding a model, you are matching the way the question gets asked.
5. Add structured data so the main facts are machine readable as well as readable.
6. Link the related pages on your own site with descriptive anchor text. Your internal links are how you tell an engine which of your pages belong to the same subject.
7. Delete the keyword stuffing you added for LSI. Lists of near synonyms in a closing paragraph help nobody and make the page worse to read.

The tools that sell LSI keyword lists

They typically return terms that co-occur with your query in the pages currently ranking. That is a genuinely useful output under an accurate name: it is a coverage check, telling you which subtopics the current results treat as part of the answer and which ones your draft has missed. Use them that way. Read each suggested term as a question, ask whether your page answers it, and add a section where the answer is no. Ignore any instruction to include a term a given number of times.

This page covers latent semantic indexing, the LSI keyword claim, and what to do instead. For how engines model things rather than strings, including the markup that makes a business resolvable, see entity based SEO. For the three stages a page passes through before any of this applies, see how search engines work. For choosing which questions to cover, see keyword research in five steps.

latent semantic indexing in SEO: what is still unclear

Does Google use latent semantic indexing?

No. Google uses query understanding, entity recognition and neural language models. John Mueller of Google stated in 2019 that there is no such thing as LSI keywords, and the technique itself was designed for fixed collections rather than for a web that changes every second.

So are LSI keyword tools useless?

They are useful under a different name. Treat their output as a coverage checklist drawn from the pages currently ranking, and add the subtopics you have genuinely missed.

Should I still include related terms in my content?

Yes, because a page that fully covers a topic contains them naturally. Write the complete answer and the vocabulary follows. Adding terms to a page that does not cover the subject is the part that does nothing.

What replaced the idea of LSI in SEO practice?

Topical coverage, entities and passage level clarity. Practically: one page per question, the answer at the top, precise names, structured data, and internal links that group your pages by subject.

Is latent semantic indexing useful anywhere?

Yes, in closed collections that do not change much, such as an archive or a document management system, where the full matrix can be computed and kept. That is the setting it was designed for and it works well there.

Is semantic SEO the same as LSI?

No. Semantic SEO is the current practice of writing and marking up content so meaning and relationships are explicit. It is the thing LSI keywords was a mistaken explanation of.

Retrieval for an assistant is not the retrieval you tuned for

LSI was never the technique behind ranking, and the better move was to cover a topic in full. An assistant reads that fuller coverage differently again: it pulls passages, merges them and writes a fresh answer in its own words.

So the check changes. After you cover the topic, ask the question and read who gets credited. AI Knows Us asks ChatGPT, Gemini, Perplexity and Claude on your behalf, and reports which sites and brands show up in what they write.