Advanced feature
Cosine similarity tests that put a number on page relevance
When a rewrite is debated on opinion, measurement helps. Cosine similarity tests compare a page with a target query using embeddings and show how semantically close they are, along with shared concepts and term gaps.
What it is for
The question it answers
“How close is this page to the query, measurably?”
Embeddings turn text into numbers that capture meaning. Cosine similarity measures the angle between two of them: the closer the score, the more the page and the query talk about the same thing. It is one of the techniques retrieval systems use to decide what content is relevant.
In AI Search Console the test is practical. Alongside the score you see shared concepts, major term gaps and suggestions to better align the page. Run it before and after a rewrite to check whether the change moved the page closer to the intent you care about.
Inputs, workflow and outputs
How it works in practice
You bring
- A target query, written the way a customer would ask it
- A page on the client website to measure
- Comfort interpreting a relative score rather than a pass mark
Pick query and page
Choose the question and the page that should answer it best.
Run the test
The page and query are embedded and their cosine similarity calculated.
Read the detail
Review the content-term comparison, shared concepts and term gaps.
Rewrite and re-test
Apply suggestions that are accurate for your business, then test again and compare.
You get
- A similarity score for the page and query pair
- Shared concepts and major term gaps
- Practical suggestions to better align the page
- A before and after comparison for rewrites
Who benefits
Who gets the most from it
- Developers
- Get a reproducible measurement instead of subjective review comments.
- Technical SEOs
- Validate whether a rewrite moved closer to search intent.
- Content leads
- Settle debates between two drafts with evidence.
Illustrative example
An insurance broker rewriting a landing page
Illustrative example, not customer data. Scores below are invented to show the workflow.
“public liability insurance for mobile hairdressers”
The generic small business page scores 0.61, with gaps for “mobile” and “salon equipment”.
A section on mobile hairdressers covers travel between clients and equipment cover.
The re-test scores 0.78. A signal to investigate further, not a ranking prediction.
Limitations
What it does not do
Honest limits help you set the right expectations with stakeholders. See the general FAQ for more.
- A higher score is a signal to investigate, not a ranking or citation promise.
- Scores are relative. Compare versions of the same page against the same query rather than chasing a fixed number.
- Adding terms to raise a score without adding genuine information can make a page worse for readers.
- Different embedding models give different absolute values.
Questions
Cosine similarity tests: frequently asked questions
What is a good cosine similarity score?
There is no universal threshold. Use the score comparatively: the same page before and after a change, or two drafts against the same query.
Do I need to code to use it?
No coding is required to run a test, but it is labelled Advanced because interpreting the results needs some understanding of embeddings and search intent.
Is this what ChatGPT uses?
Similarity measures like this are a common part of retrieval systems, but no provider publishes exactly how it chooses sources. Treat the score as a useful proxy, not a model of any specific AI.
Related features
Start with what is true today
Measure your next rewrite before you publish it
No ranking guarantee. AI Search Console provides observations and recommendations only. It cannot guarantee rankings in any search engine or mentions in any AI-generated answer, which are determined by third-party providers. Product examples on this site are illustrative and do not show real customer data.