LSI Keyword Research

LSI keywords are words and phrases that are conceptually related to your main topic — the vocabulary a search engine (or a reader) expects to see around a subject. For "espresso machine", that's terms like "portafilter", "milk frother", "bar pressure", or "grind size". They round out an article so it reads as genuinely comprehensive rather than repeating one phrase.

Yes — the term comes from Latent Semantic Indexing, a 1980s information-retrieval technique that maps documents and terms into a shared vector space to detect relationships between words. In SEO, "LSI keyword" has become informal shorthand for "semantically related term," even though modern search engines don't use the original LSI algorithm itself.

Not literally — Google has said its systems don't use classic LSI. What it does use is far more advanced semantic and contextual understanding (embeddings, BERT/MUM-style models) that achieves a similar practical effect: pages that cover a topic's full vocabulary tend to be understood as more relevant. "LSI keywords" survives as a useful SEO shorthand for that broader idea, not a literal ranking factor.

Synonyms mean nearly the same thing ("buy" / "purchase"). LSI keywords are broader — they include synonyms, but also related concepts, attributes, and subtopics that aren't interchangeable with your main term at all (for "espresso machine": "tamper", "crema", "descaling"). LSI is about topical completeness, not just word substitution.

Long-tail keywords are about search intent and specificity — longer, more precise phrases with lower volume and less competition ("best espresso machine under $300"). LSI keywords are about topical coverage. In practice they overlap a lot: many long-tail phrases are also good LSI candidates, which is why this tool surfaces both from the same Autocomplete expansion.

They help you write content that's naturally thorough instead of thin or repetitive. Practically, that means better coverage of what readers actually search for around your topic, fewer missed subtopics, and content that doesn't rely on stuffing one exact-match phrase to signal relevance.

In seed mode, it expands your phrase through Google Autocomplete using dozens of question words, commercial modifiers ("price", "best", "vs"), and alphabet sweeps, in the language you select. In domain mode, it first crawls the target site and extracts its real keyphrases (via YAKE), then runs those through the same Autocomplete expansion.

Seed mode starts from a phrase you type in — useful when you already know your target topic. Domain mode starts from a competitor's (or your own) website and reverse-engineers likely topics from its actual page content, which is useful for competitive research or auditing your own site's coverage.

Every candidate keyword is converted into a vector embedding with a multilingual sentence-transformer model, then grouped with agglomerative clustering based on cosine similarity. Keywords that are close in meaning end up in the same cluster, even if they don't share any words — that's what separates this from a simple word-overlap grouping.

A multilingual sentence-transformer (paraphrase-multilingual-MiniLM-L12-v2), running locally on the server rather than through a paid API. It supports all the languages in the dropdown, so a query and its suggestions in Ukrainian, German, or Spanish get clustered just as reliably as English.

There's no fixed number — it depends on the topic's breadth. A better approach is to skim the clusters this tool returns and pick a handful of the most relevant ones per section of your outline, rather than trying to cram in every candidate. Quality of coverage matters more than raw count.

Naturally, wherever they fit the sentence — headings, intro, body copy, image alt text, FAQ sections. There's no special weighting for putting them in any one spot; the goal is that the finished page genuinely discusses those subtopics, not that the terms appear in a particular position.

Yes. The same clusters that make good subheadings for content also make good ad group themes or negative/positive keyword ideas for search campaigns, since they represent how people actually phrase related searches.

No — stuffing unrelated or awkwardly-forced terms into a page doesn't help and can hurt readability and rankings alike. LSI keywords work as a checklist for topical completeness, not as a quota to hit. Use the ones that genuinely fit what you're writing.

Yes. Every finished job has a "download CSV" link with cluster labels and keywords, plus a JSON view. You can also skip the UI entirely and call POST /api/research with a seed or domain, then poll GET /api/jobs/{id} for the clustered result.