Information Gain (IG)

Information gain is a measure of how much new, meaningful information something provides beyond what is already known. The term has two related uses. In machine learning, it measures how much a feature reduces uncertainty, and it is used to build decision trees. In SEO, it refers to a score, described in a Google patent, for how much extra information a document offers a user who has already seen other pages on the topic.

How it works

In machine learning, information gain is the drop in entropy (uncertainty) after splitting data on a feature. A decision tree predicting whether a visitor converts picks the question that best separates converters from non-converters first, because that split gives the highest information gain.

In search, Google’s patent “Contextual estimation of link information gain” (US11354342B2, granted 2022) describes:

• Representing each document as a semantic feature vector (an embedding) based on its entities, phrases and structure.
• Comparing it with documents the user has already viewed.
• Using a machine learning model to output a score for how much new information the document adds, and potentially ranking or ordering results partly by that score.

As MLforSEO explains, this makes the score personal and contextual: two people with different reading histories could get different scores for the same page.

How it’s used in SEO and marketing

• Content planning: compare competitor coverage and find entities, attributes, data or angles nobody has covered.
• Content refresh: prioritise pages that mostly repeat what already ranks.
• Keyword research: treat low-gain informational clusters as opportunities for original research, examples or tools.
• AI search: answer engines that synthesise many sources have little reason to cite a page that adds nothing new, so unique facts and first-hand evidence matter even more.

Example

Ten ranking articles on “how to choose running shoes” all cover cushioning, fit and budget. Your page covers those too, but adds results from testing 20 shoes on the same route and a table of how weight changed with size. For a reader who has already read two of the existing articles, your page is the one with new information, which is exactly what an information gain score is meant to capture.

Limitations and things to watch

• A patent shows what Google has designed; it does not prove how the method is used in ranking today.
• Novelty is not the same as quality; unverified or irrelevant additions do not help users.
• Any score you compute yourself (for example entity or embedding overlap with competitors) is an approximation of the idea; Google’s own score is not visible.
• Google’s people-first content guidance asks whether content provides original information, reporting, research or analysis, which points in the same direction without naming the metric.

Learn it on MLforSEO

• What Is Information Gain in SEO? How Google Measures Content Uniqueness
• What Is the EAV Model in SEO?
• How to Turn Semantic Keyword Research Into an Actionable SEO Workflow
• Semantic ML-enabled Keyword Research (course)