What does 'distance' mean here?

It is 1 minus the cosine similarity between the sentence embedding and the centroid. Higher distance = more off-topic relative to the rest of the passage.

Are all flagged sentences actually wrong?

No. An outlier badge means the sentence is semantically distant from the dominant topic. Sometimes that is intentional — a contrast or hook. Use editorial judgment after reviewing.

Why is Semantic assist required?

There is no meaningful heuristic fallback for centroid distance. Keyword overlap cannot capture conceptual proximity across varied vocabulary.

How many sentences can I analyse at once?

Up to 30 sentences per run. For longer pieces, split into sections and run each separately.

How is the outlier threshold set?

The bottom quartile by centroid distance — 25% of sentences with the highest distance score are flagged. You cannot change this threshold directly; split the passage to control which section is analysed.

Content & Writing

New

Live

Outlier Sentence Detector

Flag off-topic or padding sentences by distance from the passage centroid.

Text input

Paste prose to detect off-topic or padding sentences. Semantic assist required — enable the checkbox below to compute centroid-distance rankings via /api/embeddings.

Enable Semantic assist — embeds all sentences, computes a topic centroid, and flags the bottom quartile by cosine-distance as potential outliers. Up to 30 sentences per run.

Enable Semantic assist above to run the outlier analysis. This tool requires embeddings and cannot produce results in local-only mode.

X / Twitter LinkedIn Facebook

Start here · What is a centroid-distance outlier?

Every sentence in a passage has an embedding — a vector that captures its meaning. The centroid is the average of all those vectors, representing the passage's central topic.

Sentences far from the centroid drift away from that central theme. They may be padding, tangential anecdotes, or copy-pasted from an unrelated source.

This tool requires Semantic assist to be enabled. It embeds all sentences, computes the centroid, ranks by distance, and flags the bottom quartile as outlier candidates.

When to use this tool

You have a real task today
Use Outlier Sentence Detector when you need to flag off-topic or padding sentences by distance from the passage centroid. That might be a launch checklist, a client question, or a ticket that says "quick check before we publish."
You are learning what the numbers or text mean
Run Examples below with a tiny input. Run one real draft through the tool before you rewrite a whole content library. Watching one clean input change is the fastest way to build intuition.
You are scaling across many URLs or writers
Prove the workflow once here, then repeat the same pattern in a spreadsheet, Notion database, or your SEO platform. This page is the sandbox; your tracker is the production line.

Examples

Walk through these with the form above — they are practice scenarios, not live data.

Minimal first run (recommended)

Try this

Give the smallest honest input the labels allow — one short paragraph, one keyword, one staging URL, etc. Avoid pasting an entire site on the first try.

What to look for

Notice the shape of the result: lists, scores, generated copy, or checks. Outlier Sentence Detector makes sense once you have seen one full input→output cycle.

Real page or real draft

Try this

Paste a draft, outline, keyword, or page text from one URL. Use something you would actually ship or review, not placeholder text.

What to look for

Look for the recommendations that make the page clearer, more complete, or easier to scan. If the result surprises you, change only one variable and run again.

Short tutorial

Follow in order the first time you use the tool; later you can skip to the step you need.

Step 1 — Read labels and placeholders
The Outlier Sentence Detector form tells you whether to fetch a https:// URL, paste raw text, or drop HTML. Placeholders show the format the tool expects.
Step 2 — Enter the smallest useful input
Start with the minimum that still describes the page or problem. Add more only if the output is empty, thin, or obviously wrong.
Step 3 — Run and scan the output
Use the main action (Analyze, Fetch, Generate, Score, etc.). Read headings and numbers before you copy anything — you are checking for sanity, not chasing perfection on the first pass.
Step 4 — Copy, file a ticket, or iterate
Paste the snippets you need into your CMS or doc. If something is off, change one field at a time so you learn what moved the result.

More detail

Padding sentences silently hurt content quality scores, increase time-on-page in the wrong ways, and dilute topical relevance signals. Manually spotting them in a 2,000-word draft is hard — they look plausible in context even when they stray. Centroid-distance ranking gives you an objective, reproducible order: the sentence farthest from your passage's topic centroid is your first revision candidate.

FAQ

What does 'distance' mean here?: It is 1 minus the cosine similarity between the sentence embedding and the centroid. Higher distance = more off-topic relative to the rest of the passage.
Are all flagged sentences actually wrong?: No. An outlier badge means the sentence is semantically distant from the dominant topic. Sometimes that is intentional — a contrast or hook. Use editorial judgment after reviewing.
Why is Semantic assist required?: There is no meaningful heuristic fallback for centroid distance. Keyword overlap cannot capture conceptual proximity across varied vocabulary.
How many sentences can I analyse at once?: Up to 30 sentences per run. For longer pieces, split into sections and run each separately.
How is the outlier threshold set?: The bottom quartile by centroid distance — 25% of sentences with the highest distance score are flagged. You cannot change this threshold directly; split the passage to control which section is analysed.