Skip to content
Browse courses

Model Evaluation

Sentiment Score (Polarity)

In shortA number showing whether text is positive, negative or neutral. In Google's Natural Language API it runs from -1.0 (very negative) to +1.0 (very positive).

Updated 24 Oct 2026· 2 min read

A sentiment score, also called polarity, is the number a sentiment analysis model gives a piece of text to show whether its overall tone is positive, negative or neutral. In Google’s Cloud Natural Language API the score runs from -1.0 (clearly negative) to +1.0 (clearly positive), with values near 0 for neutral or mixed text. Other tools such as IBM Watson NLU report similar scores with their own scales and labels.

How it works

The model reads the text and estimates its emotional direction for the whole document, for each sentence and, with entity sentiment analysis, for each entity mentioned. The score tells you the direction. It is paired with sentiment magnitude, which tells you how much emotion there is regardless of direction:

  • score near 0, low magnitude: genuinely neutral text,
  • score near 0, high magnitude: mixed feelings, strong positives and negatives cancelling out,
  • score near +1 or -1, high magnitude: clearly and strongly positive or negative.

Google recommends setting your own thresholds for what counts as clearly positive or negative after testing on your data.

How it’s used in SEO and marketing

  • Review analysis: score customer reviews and group them by product, entity or theme.
  • SERP monitoring: score the titles and descriptions ranking for your brand queries to see how coverage leans.
  • Content checks: compare the tone of your pages with the tone of pages that rank.
  • Brand and PR tracking across news, forums and social mentions.
  • Labels for reporting: convert scores into tags such as slightly negative or very positive for dashboards.

Example

MLforSEO’s SERP analysis dashboard plots ranking titles by position on one axis and sentiment score on the other, with bubble size showing magnitude. Filtering by keyword shows where negative titles about a brand sit in the top results, which helps prioritise reputation work. In the no-code Sheets template, the same scores come with a custom label, such as “slightly negative”, that you can rename in Apps Script.

Limitations and things to watch

  • Sarcasm, irony, negation and domain jargon trip models up.
  • A document score can hide mixed opinions; look at sentence or entity level for reviews that praise one thing and criticise another.
  • Scores are not comparable across tools, so stick to one model for trend tracking.
  • Short texts such as titles carry little signal, so treat individual scores with caution and look at patterns.

Learn it on MLforSEO

Where Sentiment Score (Polarity) shows up on MLforSEO

Learn it in a course

Read about it on the blog

Try it yourself

Sources and further reading

Other terms from Model Evaluation

See all 19 Model Evaluation terms →