Case Study 2: The "LSI Keywords" Myth — How a Real Idea Became Folklore

A contested/debunked practice, told to teach the chapter's limits (theme 3: evidence over folklore). Built from the documented public record — a real 1990 information-retrieval technique, the SEO market that misappropriated its name, and public statements by Google representatives. Reconstructed framing is labeled; no statistics are invented.

Background

Latent Semantic Indexing (LSI) is a genuine information-retrieval technique, introduced in a 1990 research paper ("Indexing by Latent Semantic Analysis," Deerwester, Dumais, and colleagues). It uses a linear-algebra method (singular value decomposition) to find latent relationships between terms in a fixed, modest-sized collection of documents, so that a search can retrieve documents that are conceptually related even when they don't share exact words. It was a clever solution to synonymy and was designed, in the era it was born, for collections of thousands or tens of thousands of documents.

Somewhere in the SEO industry, that real term acquired a second, unofficial life. Tools and courses began selling "LSI keywords" — lists of "semantically related" words you were told to sprinkle into a page to improve its relevance. The phrase spread until it became received wisdom, repeated in countless blog posts and baked into paid tools.

The search / SEO issue

Two things were wrong at once, and untangling them is the whole lesson.

First, "LSI keywords" is not a real thing. LSI is a document-indexing technique, not a category of keyword. There is no such object as an "LSI keyword"; the phrase was a marketing coinage. The idea it points at — that covering semantically related concepts helps — is partly true, but for a different reason (co-occurrence and topical coverage, §4.3), and it does not require, involve, or resemble the 1990 LSI method.

Second, Google almost certainly does not use LSI on the web. LSI was built for small, static collections; it does not scale to a corpus of trillions of constantly-changing pages, and Google long ago moved to far more powerful, purpose-built language systems (word embeddings, neural models, and eventually BERT and its successors — Chapter 2). Google's own representatives have said so publicly and bluntly. Google Search Advocate John Mueller stated on social media (2019) that "there's no such thing as LSI keywords — anyone who's telling you otherwise is mistaken." (Public statement, widely cited.)

What it shows

  • A real concept can be mythologized. The credibility of a genuine academic technique (LSI) was borrowed to lend authority to a tactic (sprinkling "related" words) that the technique does not actually describe. The scientific-sounding name did the persuading.
  • The kernel of truth is what makes the myth durable. Covering related concepts does help — because it demonstrates genuine topical depth (co-occurrence), not because of anything LSI does. When advice is "half-right for the wrong reason," it survives, because people see it "work" and credit the wrong cause.
  • Confident, technical-sounding folklore is the industry's default failure mode. "LSI keywords," "ideal TF-IDF density," "the perfect keyword density" — each dresses a plausible half-truth in false precision, and each sells better than the honest version ("cover the subject the way an expert would").

Outcome

Despite repeated, explicit debunking by Google representatives and by rigorous practitioners, "LSI keywords" remains one of the most stubbornly repeated terms in SEO, and tools still sell "LSI keyword" features. The honest version of the underlying advice, meanwhile, is not only true but more useful: use the words and concepts a genuine expert on the topic would use, because that is how you demonstrate — to readers and to a semantic search engine — that your page really covers the thing. No word list required; understanding the subject is the work.

The lesson

This is the chapter's honesty theme made concrete. When you meet an SEO claim — especially one with an impressive acronym — run it through the evidence tiers (Chapter 2): is it confirmed by Google, a labeled correlation, professional experience, or speculation dressed as fact? "Add LSI keywords to rank higher" collapses immediately under that test: the mechanism it names isn't used, the object it names doesn't exist, and the real advice it gestures at is better stated plainly. The durable skill is not memorizing tactics; it is the reflex to ask "what actually supports this?" — and to prefer an honest "cover the topic genuinely" over a confident, sciencey fake.

Discussion questions

  1. "LSI keywords" is half-right (covering related concepts helps) for the wrong reason (LSI is not how). Why are half-right-for-the-wrong-reason claims more dangerous than claims that are simply false?
  2. A vendor's tool outputs a list of "semantically related terms to add." Is the tool useless, or can it be used well? Describe the honest way to use such a list and the misuse to avoid.
  3. Compare the "LSI keywords" myth to the "ideal keyword density" myth (Chapter 9). What do they share, and which chapter concept debunks each?
  4. Draft the two-sentence reply you would give a client who forwards you a course promising "rank higher with LSI keywords and optimal TF-IDF." Keep it honest and non-condescending.