--- type: "Article" title: "latent semantic indexing | topical authority content" description: "What latent semantic indexing actually is, why Google says LSI keywords do not exist, and what to do instead for topical coverage." resource: "https://www.laurelinlabs.com/search-systems/latent-semantic-indexing" tags: ["latent-semantic-indexing", "topical-authority-content", "lsi-keywords"] generated: { by: "laurelin-site-build/2026-09-02", at: "2026-09-02T12:00:00+00:00" } status: "stable" stale_after: "2027-03-02T00:00:00+00:00" --- # Latent semantic indexing (LSI) Latent semantic indexing is the information retrieval name for latent semantic analysis, a technique that uses singular value decomposition on a term and document matrix to group words and documents by shared concepts. It dates from the late 1980s. Google has never said it uses LSI, and John Mueller has said on record that there is no such thing as LSI keywords. For sites, lists of so called LSI keywords are at best harmless and at worst keyword stuffing; covering the topic properly is what matters. ## What is latent semantic indexing? Latent semantic indexing (LSI) is the name used in information retrieval for LSA, a technique that analyses which terms occur in which documents and derives a smaller set of underlying concepts. Wikipedia records that an information retrieval technique using latent semantic structure was [patented in 1988](https://en.wikipedia.org/wiki/Latentsemanticanalysis) by Scott Deerwester, Susan Dumais, George Furnas, Richard Harshman, Thomas Landauer, Karen Lochbaum and Lynn Streeter. The technique predates Google Search and is not a Google invention. ## How does LSI work? LSI builds a matrix of terms by documents, usually weighted with TF-IDF, then applies SVD to keep only the strongest dimensions. Words that occur in similar contexts end up close together in that reduced space, so a query about cars can match a document about automobiles even without the shared word. Documents are compared with cosine similarity. It was an early answer to synonymy and polysemy, but it scales poorly to a corpus the size of the web and has been overtaken by learned embeddings, which do the same job with far more data and nuance. ## What evidence is there that Google uses LSI? This page is labelled **General theory**. There is no Google confirmation that LSI is used in Search, and Google staff have repeatedly denied that LSI keywords are a thing. John Mueller said so in July 2019 and repeated it in January 2023, answering a question about whether to put LSI keywords in headings or body text with the reply that both have no effect ([Search Engine Roundtable, September 2023](https://www.seroundtable.com/google-lsi-keywords-have-no-effect-34668.html)). Google's published explanations of meaning in Search point to RankBrain and neural matching, which are machine learning systems, not to LSI ([Google ranking systems guide](https://developers.google.com/search/docs/appearance/ranking-systems-guide)). ## Are LSI keywords a myth? Yes, as a Google ranking concept. "LSI keywords" in SEO tools are usually related searches, co-occurring terms or synonyms scraped from results pages. Some of those words may be worth using because readers expect them, but not because Google runs LSI. The risk is how they get used. Pasting a list of related terms into a page, or forcing synonyms into every paragraph, matches the pattern Google's spam policies describe as keyword stuffing. - Myth: Google ranks pages using LSI. No public evidence; Google staff deny it. - Myth: adding LSI keywords boosts rankings. Mueller: both placements have no effect. - Fact: covering the concepts a reader expects helps, because it makes the page more complete. ## Which Laurelin audit checks test for LSI? There is no LSI system to test, so these checks look for the side effects of the myth. The new check [Related term lists inserted as "LSI keywords"](https://www.laurelinlabs.com/audit/topical-content/380-related-term-lists-inserted-as-lsi-keywords) flags keyword blocks and forced synonyms. Related checks: [Keyword stuffing and over-optimisation](https://www.laurelinlabs.com/audit/topical-content/167-keyword-stuffing-over-optimisation) is the policy risk. [Keyword stuffed hidden meta, alt or ARIA text](https://www.laurelinlabs.com/audit/guardrails/306-keyword-stuffed-hidden-meta-alt-aria) catches term lists hidden in attributes. [Low information gain versus competitors](https://www.laurelinlabs.com/audit/topical-content/161-low-information-gain-vs-competitors) is what real topical coverage fixes. [Content not matching search intent](https://www.laurelinlabs.com/audit/topical-content/168-content-not-matching-search-intent) matters more than vocabulary lists. ## What are the key dates for latent semantic indexing? - 1988: Information retrieval technique using latent semantic structure patented by Deerwester and colleagues ([source](https://en.wikipedia.org/wiki/Latent_semantic_analysis)) - 2019-07-30: John Mueller: there's no such thing as LSI keywords ([source](https://www.seroundtable.com/google-lsi-keywords-have-no-effect-34668.html)) - 2023-01-02: John Mueller: LSI keywords in headings or body text both have no effect ([source](https://www.seroundtable.com/google-lsi-keywords-have-no-effect-34668.html)) ## FAQ ### Does Google use latent semantic indexing? There is no public evidence that it does. Google explains meaning matching through systems such as RankBrain and neural matching, and John Mueller has said LSI keywords do not exist. ### Should I remove LSI keywords from my pages? Remove any that were inserted as lists or forced into sentences. Keep related words that a reader would naturally expect in a complete treatment of the topic. ### How old is latent semantic indexing? The technique dates from the late 1980s; Wikipedia records a patent in 1988 by Deerwester, Dumais and colleagues. Hub: [Topical Content](/topics/topical-content.md)