LSI keywords are not a thing.

What LSI actually is

Latent semantic indexing is a real information retrieval technique, patented in the late 1980s. It works by building a matrix of terms and documents and reducing its dimensions, so that words appearing in similar contexts end up near each other. It was designed for small, static collections of documents, of the order of thousands, at a time when that was an ambitious corpus.

It does not scale to a live index of the whole web, and nothing about modern search resembles it. Google has said publicly and repeatedly that there is no such thing as LSI keywords and that it does not use LSI. The systems doing this job now are neural and vastly more capable: they work with meaning, entities and the relationships between concepts, not with a term-document matrix rebuilt on a corpus that stopped being small around 1995.

The right instinct underneath

Here is why this refuses to die: pages that cover a subject properly do use the surrounding vocabulary. Write seriously about mortgages and you will inevitably mention interest rates, deposits, terms, affordability and repayment, because you cannot explain the subject without them. A page that manages to discuss mortgages without ever touching those concepts is almost certainly shallow.

So the correlation is real, exactly as it is with word count. Thorough pages contain related terms. The error is reversing it: adding related terms to a page does not make it thorough, any more than adding words makes it comprehensive. The vocabulary is evidence of understanding, not a substitute for it.

Why sprinkling makes it worse

Treat the list as a checklist and you get the familiar damage. Terms appear in sentences constructed to hold them rather than to say anything. Paragraphs get bolted on to accommodate a phrase that did not belong. Sub-headings name concepts the page then does not actually cover, which is a promise broken twice, once to the reader and once to whatever is trying to extract an answer from it.

This is keyword density wearing a better suit. Same shape: a computable target replaces an editorial judgement, the cheapest way to satisfy it is mechanical insertion, and the page gets worse at the only thing that decides its fate.

What to do instead

Use the term lists as a gap check, not a fill list. Read them after drafting and ask whether anything on the list points to something you genuinely should have covered and did not. Sometimes it will, and that is a real finding worth acting on: not "insert this phrase" but "this section is missing".

Otherwise, the reliable method is unchanged and unglamorous. Understand the subject well enough to write about it without a reference list, answer the question the searcher actually has, include the specifics only someone with real experience would know, and let the vocabulary follow from that. That is the substance of good copywriting, and the structural half is on-page work that has nothing to do with word lists.

The general lesson

Be suspicious of any SEO rule that arrives with an impressive-sounding mechanism attached, particularly a named one. The name is doing persuasive work the evidence is not. Ask where it came from, whether the party naming it sells the tool that measures it, and whether the underlying claim would still stand if you stripped the terminology out. With LSI, what survives that test is "cover your subject properly", which was always the advice and never needed a patent number to justify it.