SEO tool

Cosine similarity between two texts

It is the measure search systems use to ask whether two documents talk about the same thing. Paste two texts, two titles or two descriptions and see the angle between them. It works well for investigating cannibalization.

When you fetch by address, the page is read by this site's server, because the browser does not allow one page to read another. The extracted text is kept for one hour, so the same page is not downloaded again, and then discarded. Text you paste into the boxes never leaves your browser.

0.000

Cosine similarity

angle 90.0 degrees

TermABWeight in the product

Limits of this measure

Vectorization here is by term frequency, with no TF-IDF and no embeddings. It shows the mechanics of the calculation.

A search engine uses dense vectors produced by a language model and arrives at a different number. The question stays the same: what is the angle between these two documents.

Why measure this

What cosine similarity answers

Two of your pages can use different words and still compete for the same search intent. Cosine similarity puts a number on that competition. It represents each text as a vector, where each dimension is a term, and measures the angle between the two vectors.

Measuring angle instead of distance has a practical consequence. A three-hundred-word text and a three-thousand-word text on the same subject end up close, because the measure ignores length and looks only at direction. Counting shared words behaves differently, always favoring the longer text.

Dois vetores saindo da mesma origem, com o ângulo entre eles marcado, e um terceiro vetor mais longo na mesma direção do primeiro θ Texto B Texto A Texto A, bem mais longo
The measure reads the angle θ between the vectors. Length follows the size of the text and stays out of the calculation: text A and a much longer version of it point the same way and give the same result against text B.

Where this comes into the work

Cannibalization

When two URLs compete for the same intent, the search engine alternates between them and the ranking becomes unstable. The measure points out the suspicious pairs before you open the queries report.

Deciding to merge or differentiate

With the number on the table, the conversation with the content team turns to how much the two pages actually overlap.

Platform migration

To match an old URL with a new one, comparing content works better than comparing paths, especially when the folder structure changes entirely.

Large catalog

Product descriptions copied from the manufacturer or generated in bulk tend to repeat. The measure finds the nearly identical blocks without you having to read everything.

Production at scale

If the operation publishes many pieces a week, measuring the similarity between them shows whether there is real variation or the same piece with swapped synonyms.

New topic

Before publishing, comparing the draft with the pages the site already has on the topic shows whether the new text answers another question or repeats one that already has an answer.

What it does not answer

The vectorization on this page is by term frequency. It knows nothing about synonyms or context, so "car" and "automobile" go in as unrelated terms, and two texts on the same subject with different vocabulary come out with low similarity.

It also does not measure quality, depth or intent. Two shallow pages on the same topic give high similarity, and the number says nothing about the quality of either one.

Verified fact
Cosine similarity is a standard measure in information retrieval, used both in the classic vector space model and in dense vector search.

Hypothesis
The number shown here does not reproduce what a search engine calculates. Production systems use dense vectors produced by language models, in addition to signals that are not in the text. Use this result to see the mechanism. It does not estimate what Google sees.