LSI keywords are an SEO hallucination: fluent, confident, repeated everywhere, and without ground.

Latent semantic indexing is a retrieval method that uses singular value decomposition to find relationships between terms and documents in a collection. It is the application of Latent semantic analysis (LSA) to search. It was developed at Bell Communications Research in the late 1980s.

Singular value decomposition factorises the term-document matrix, and keeping only its largest values produces a low-rank approximation. Documents and queries are compared in that reduced space. The whole method rests on this one calculation over one fixed matrix.

A term-document matrix records how often each term occurs in each document of a collection. Latent semantic indexing approximates it at lower rank. The matrix has to exist in full before the method can start.

Latent semantic indexing comes from information retrieval, the academic field that studies how documents are found. It was built to help people find documents that use different words for the same thing. Its home is the library and the database, long before web search.

LSI keywords is a term from SEO marketing for related words that are supposed to help a page rank because Google uses latent semantic indexing. Google does not use it for ranking. The term borrows the name of a method to sell a list of related words.

John Mueller of Google wrote on 30 July 2019: “There's no such thing as LSI keywords -- anyone who's telling you otherwise is mistaken, sorry.” The myth continued anyway.

Google describes neural systems for understanding language in search: RankBrain from 2015, neural matching from 2018 and BERT from 2019. These models learn meaning from very large amounts of text. None of them is a singular value decomposition of a term-document matrix.

Semantic search is search that works with the meaning of queries and documents beyond exact word matches. Latent semantic indexing was one early attempt at it. Semantic search today rests on neural models, and meaning is still learned from words.

Word co-occurrence is the pattern of words appearing in the same documents or contexts. Latent semantic indexing derives its associations from these patterns. Related words matter because they occur together in real writing, whatever tool a marketer names.

A low-rank approximation represents a matrix with fewer dimensions while staying as close as possible to the original. It is what makes latent semantic indexing compare concepts beyond shared words. It also discards detail, including rare terms.

Latent semantic analysis is the broader mathematical procedure, and latent semantic indexing is its use in retrieval. The two names are often used interchangeably. The science behind both is set out in detail on the page on Latent semantic analysis (LSA).

Bell Communications Research, known as Bellcore, was the research arm of the regional Bell telephone companies. Scott Deerwester, Susan Dumais, George Furnas, Thomas Landauer and colleagues developed latent semantic indexing there. The United States patent was granted in 1989.

On-page SEO practice took up the term to argue for adding topically related words to a page. The advice to cover a topic with its natural vocabulary has merit. Attaching it to latent semantic indexing gave it a false technical authority.

Everything that shows why it cannot rank the web. The usual treatment names singular value decomposition and related terms. It leaves out that the method needs a fixed collection, that new documents require folding in or recomputation, that the reduction discards rare terms, and that search moved to neural retrieval long ago. Without these, LSI keywords sound like engineering. They are an SEO hallucination: fluent, confident, repeated everywhere, and without ground.

Truncated singular value decomposition keeps only the largest singular values and discards the rest. It is the step that turns word counts into latent concepts. What it discards includes rare, specialised terms, often the ones that make a page distinctive.

Folding-in projects a new document into an existing latent space without recomputing the decomposition, as Michael Berry, Susan Dumais and Gavin O'Brien described in 1995. It is fast and grows less accurate as the collection changes. The web changes every second.

Cosine similarity measures the angle between a query vector and a document vector in the reduced space. A small angle means high similarity. It compares directions of meaning, whatever the exact words.

The Frobenius norm measures how far the approximated matrix lies from the original. It states the information lost by the reduction as a number. Every reduction loses something, and the number says how much.

Latent semantic indexing gives each word a single position, so the different meanings of one word are averaged together. It handles words with several meanings poorly. Context-sensitive neural models were built to do better.

Latent semantic indexing maps different words that occur in similar contexts to nearby positions. That is its strength and the reason it was invented. A list of related terms pasted into a page does not reproduce it.

If LSI gained strength on large collections, its margin on TREC-3 would be wide. Dumais put it at about 5 % and called it smaller than on other test collections.

Matrix reconstruction error is the difference between the original matrix and its low-rank reconstruction. It shows that the compressed representation is lossy. The latent space is a summary, and summaries omit.

Probabilistic latent semantic analysis, introduced by Thomas Hofmann in 1999, recast the method as a statistical model. Latent Dirichlet allocation followed in 2003. The field moved on from latent semantic indexing within a decade.

A dense vector space represents words and documents as vectors in which most values are non-zero. Latent semantic indexing produced one of the first. Every modern embedding model works in such a space.

The eigenvalue spectrum, or the sequence of singular values, shows how much each latent dimension contributes. Where it is cut decides how many concepts survive. The choice is a judgement, and it changes the results.

Non-negative matrix factorization decomposes a matrix into factors without negative values, which makes the factors easier to read as topics. It is one of several alternatives to singular value decomposition. Latent semantic indexing is one technique among many.

The continuous bag-of-words model, part of word2vec from Tomas Mikolov and colleagues at Google in 2013, predicts a word from its surrounding words. It learns word vectors from local contexts. It replaced matrix factorisation in many applications.

The skip-gram model, the other half of word2vec, predicts the surrounding words from a given word. It captures fine semantic relations in its vectors. Neural embeddings of this kind stand between latent semantic indexing and today's language models.

The transformer architecture, introduced by Ashish Vaswani and colleagues at Google in 2017, builds representations with attention over all words of a text. BERT and today's large language models are transformers. Modern search understanding is built on it.

Dense passage retrieval encodes questions and passages as vectors and retrieves passages by vector similarity, as Vladimir Karpukhin and colleagues showed in 2020. It works on passages, beyond whole documents. It is what semantic retrieval looks like today.

Cross-encoder re-ranking reads a query and a candidate document together through one neural model to score their relevance, usually on a short list of candidates. Rodrigo Nogueira and Kyunghyun Cho applied BERT this way in 2019. It is the precise second stage of modern retrieval.

A bi-encoder encodes queries and documents separately into vectors, so that document vectors can be computed in advance. Sentence-BERT by Nils Reimers and Iryna Gurevych made the approach widely used. It is the fast first stage.

Approximate nearest neighbour search finds the most similar vectors in a huge collection without comparing against every one. It trades a little accuracy for great speed. Vector search over billions of documents depends on it.

Hierarchical Navigable Small World, HNSW, is a graph-based index for approximate nearest neighbour search, developed by Yury Malkov and Dmitry Yashunin. Many vector databases use it. Graph indexes of this kind have replaced term matrices in large-scale semantic retrieval.

Document vector normalisation adjusts vectors for document length so that long documents are not favoured merely for containing more words. Amit Singhal, Chris Buckley and Mandar Mitra refined it as pivoted normalisation in 1996. Length alone changes similarity scores.

The static corpus assumption is the premise that the document collection is fixed when the decomposition is computed. Latent semantic indexing depends on it. The web never holds still long enough to satisfy it.

If Google ranked the web with LSI, the collections LSI was computed on would come close to the size of its index. The published ones reach about 3.2 million documents.

The vocabulary mismatch problem is the fact that searchers and authors use different words for the same thing. George Furnas and colleagues measured it in 1987. Latent semantic indexing was invented to solve exactly this problem, and it remains real.

Distributional semantics is the theory that words occurring in similar contexts have similar meanings, going back to Zellig Harris and to J. R. Firth's remark that you shall know a word by the company it keeps. Latent semantic indexing and modern embeddings both rest on it. The theory is sound, and the myth misuses it.

Semantic drift is the gradual shift of a word's meaning over time or of a topic across a collection. A fixed latent space cannot follow it. Adding loosely related terms to a page can also drift its topic away from the question.

Inference latency is the time a system needs to compute an answer for a query. Web search answers in fractions of a second over a vast index. Recomputing a decomposition of the web for each change is not how that happens.

If LSI cost stayed flat with collection size, the times below would stay close together. They climb from minutes to most of a day, and the full TREC collection was decomposed only as a sample.

A hallucination in a language model is a fluent, confident statement without grounding in fact. LSI keywords have the same structure: a technical-sounding claim, repeated across countless pages, contradicted by the company it describes. The repetition made it look like consensus, and consensus is where such claims hide best.

Use the real vocabulary of its field: the terms its studies and books use, their synonyms and the words that belong around them. That helps, because meaning is carried by words and every semantic model learns from them. It helps without any connection to latent semantic indexing, and it needs no list from a keyword tool.