What is LSI Keywords?

What latent semantic indexing really is

LSI is an information retrieval method developed in the late 1980s by Scott Deerwester, Susan Dumais, George Furnas, Thomas Landauer and Richard Harshman, most of them researchers at Bellcore. It was first presented in a 1988 conference paper, and its most cited description appeared in a 1990 journal article. The problem it tackled was simple: a search for "automobile" would miss a document that only said "car", because plain word matching cannot connect the two.

The method builds a large term–document matrix recording how often each word appears in each document of a collection, then applies singular value decomposition (SVD). Thousands of words are compressed into a much smaller space of typically a hundred to a few hundred "latent" dimensions. Words that tend to occur together end up close to each other in that space, and queries and documents are compared there, bridging synonyms.

LSI was designed for fixed, relatively small collections. Any change to the collection means recomputing the decomposition, which makes it impractical for a web index of billions of constantly changing pages.

Where "LSI keywords" came from

From the 2000s onwards, some SEO tools and blogs began calling lists of topic-related words "LSI keywords". Those lists are not produced by LSI. They are typically assembled from autocomplete, related searches or words that appear frequently on top-ranking pages. The label borrowed the credibility of an academic method and stuck.

Google's John Mueller stated plainly on social media in 2019 that there is no such thing as LSI keywords, and Google's search documentation does not describe LSI as a ranking concept. What Google does describe is the use of neural language models such as BERT, announced in 2019 for understanding queries, along with semantic search systems in general. These models represent words and passages as vectors called embeddings. The goal is the same one LSI pursued; the machinery is entirely different.

What people actually mean

The intuition behind the term is sound: text that covers a subject thoroughly will mention the concepts that belong to it. An article on brewing filter coffee that never mentions grind size, water temperature, brew time or coffee-to-water ratio is probably only half an answer. Those concepts are not words added to fool a search engine; they are necessary parts of a good explanation. More accurate names are "semantically related terms", "topical coverage" or the completeness of the entities a page discusses.

Putting the idea to work

• Use related-term lists as a checklist for subquestions the article has not answered yet.
• Do not try to sprinkle every listed word into the text; terms dropped in without context hurt readability and drift into keyword stuffing.
• Instead of adding a term, add the information it points to: a definition, an example, a comparison.
• Lasting gains come not from the vocabulary of a single page but from content that covers a subject across all its subtopics, which is the basis of topical authority.