Entity salience is a score that shows how central a given entity, such as a business, person, place or product, is to a piece of text relative to the other entities it mentions. The term is best known from Google’s Cloud Natural Language API, which returns a salience value between 0 and 1 for each entity it finds in a document.
How entity salience works
A natural language processing model reads the text, identifies named things in it and works out which ones the text is mostly about. Signals include where an entity first appears, how often it is mentioned, whether it is the subject of sentences or a passing reference, and how other words relate to it.
Take a page from an Edinburgh wedding photographer. If the text keeps talking about specific venues, the weather and the bride’s dress, those may score higher than the photographer’s own service. Rewrite it so the photographer, their style and what clients receive lead each section, and the service becomes the most salient entity, with the venues in support.
Salience is not the same as frequency. An entity mentioned once in the opening sentence as the subject can score higher than one mentioned five times in passing.
Why it matters
Google has not said that Search uses the same salience model as its Cloud API, and nobody outside Google knows how its ranking systems weigh the subject of a page. So salience is a diagnostic, not a ranking factor you can measure.
It is still a useful one. Search engines and AI systems need to work out what a page is mainly about, and the same writing habits that confuse a public NLP tool are likely to blur the signal for them too. Pages that clearly centre on one subject tend to be easier to match to the right queries, and easier to summarise accurately. This connects closely with semantic search, where meaning rather than exact wording decides relevance.
It is especially helpful for spotting drift. Service pages that spend half their length on company history, and blog posts that start on one topic and finish on another, often show it clearly in a salience report.
Common mistakes
- Chasing the number. Rewriting copy until the target term scores 0.8 usually produces awkward, repetitive text that reads worse for people.
- Testing whole pages including navigation. Menus, footers and cookie text add noise. Test the main content only.
- Overusing pronouns. Long passages built on “it”, “they” and “this” give the model less to attach meaning to.
- Treating one tool as the truth. Different NLP models disagree, and none of them is Google’s ranking system.
- Ignoring intent. A page can have perfect salience for the wrong subject if it does not match what searchers want.
How to act on it
For important pages, paste the main body text into the Google Cloud Natural Language demo and look at the entity list. If the top entity is not what the page should be about, reread the copy. Usually the fix is structural: put the main subject in the title, the H1 and the first sentence, open each section by naming it, and trim tangents into their own pages or cut them.
Then judge the page as a reader would. If the rewrite makes the subject obvious to a person, you have done the job; the score is a side effect. I use this kind of check during on-page SEO reviews, mainly to catch pages that have lost focus over years of edits.