Information Gain

It is what readers learn from a piece that the other results for their question don't offer: an original measurement, a comparison from practice, a well-argued limit of a method, an example under real conditions. A piece that restates the standard definition in different words has no information gain, even if it is long, current and accurate. The yardstick sits outside the text. It is set by what already exists on the topic.

In SEO circles the term is usually traced to a Google patent, "Contextual estimation of link information gain" (US 11,354,342 B2, filed October 18, 2018, granted June 7, 2022). A patent describes an invention. It does not prove the invention is in use, and Google has not publicly confirmed using the method for ranking (as of November 2026). What is documented is what Google asks in its guidance: whether content provides original information, reporting, research or analysis, and whether it offers substantial value compared with other pages in search results.

For executives, the term matters because language models now produce standard copy in any quantity. One more article on the same topic costs almost nothing, and it is worth about as much. What stays scarce is knowledge that only exists inside the company: data, cases, experience and reasoned judgment. Anyone who commissions or signs off content should ask what this piece knows that other pieces don't.

How information gain works.

Information gain is a comparison. On one side is everything a reader already finds for their question, such as the top search results or an AI answer. On the other side is the new piece. Only what it adds beyond that baseline counts.

Added value mostly comes from six sources: original data or measurements with a disclosed method, hands-on experience, a comparison under the same conditions, edge cases and counterexamples, a reasoned judgment, and a new state of affairs when rules or technology have changed.

Length does not create value. Neither do extra keywords, rewording, or condensing the top ten results into an even longer eleventh. That technique, known as skyscraper content, by definition delivers nothing that does not already exist.

The baseline depends on the question. For a basic question like "What is a heat pump?" it is high, because hundreds of good explanations exist. For a narrow question from practice it is often low, and one well-documented case can add a lot. When planning topics, check where the baseline sits first.

What the Google patent describes.

Patent US 11,354,342 B2 lists Victor Carbune and Pedro Gonnet Anders as inventors, with Google LLC as assignee. It describes this situation: a person has already seen some documents on a topic. For further documents, a system calculates an information gain score, a value for how much new information a document contains compared with what the person has already seen. A machine learning model processes data from the documents to do this. Documents with a higher score can be shown first or, for example, read aloud by an automated assistant.

In this method the score depends on what the individual person has already seen. It is not a fixed property of a page.

Many patented methods never reach a product, so the patent does not show that Google scores pages this way today. The direction is clear anyway, because it also appears in Google's public guidance. There Google asks whether content provides "insightful analysis or interesting information that is beyond the obvious." In May 2025, Google's Search Central blog advised site owners to focus on "unique, non-commodity content."

Why generative AI makes the term more important.

Language models write a decent piece on almost any topic in seconds. That piece reflects what appeared often in the training data. It is pretty much the baseline that content has to beat.

Google has drawn two documented conclusions. Its spam policies list, as an example of scaled content abuse, using generative AI to generate many pages without adding value for users. Its guide to generative AI features in Google Search advises against recycling what others on the internet have already said or what a generative model could easily produce. The same guide says these AI features are rooted in Search's core ranking and quality systems.

A practical point follows. If an AI answer already summarizes the common knowledge on a question, readers have little reason to open a page that says the same thing again. Whether AI search systems cite such pages more often is not established and can only be measured case by case.

AI can help with the writing. The added value has to come from the company.

How to spot information gain in a piece.

Checking whether a piece adds anything takes four steps and no special tools.

First, establish the baseline: read the top search results and an AI answer for the question and note their core points. Second, name the addition in one sentence. If you can't, there isn't one yet. Third, check the evidence: where does the addition come from, is the method stated, are time frame and conditions clear? Fourth, ask whether the addition helps the reader decide something or just sounds new.

In the text itself, strong pieces show specific conditions instead of generalities, numbers with source and date, counterexamples and a clearly stated limit. A reasoned disagreement with common advice is another good sign.

Tools that output an information gain score usually compare texts with other texts. They can measure how much a piece differs. They cannot check whether the difference is correct or whether it helps anyone. That remains a judgment call for whoever approves the piece.

Common misconceptions.

The most common misconception is that a longer piece has more information gain. Length is not the measure. A short section with one documented case can add more than a long guide that summarizes what is already everywhere. New wording doesn't change that either: a rewritten piece still makes the old point.

Equally common is "new equals good." A surprising claim without evidence hurts credibility. This is where E-E-A-T comes in: new claims need a source, a method or visible experience before readers will trust them.

Information gain is also often treated like a metric you can optimize. There is no confirmed score that Google calculates for pages and discloses. Optimizing against a homemade score means optimizing for the tool's logic rather than for what readers need.

An original contribution does not require an original study. Often a precisely described decision, a before-and-after comparison or a limit that only practitioners know is enough.

Information gain in machine learning.

In information theory and machine learning, information gain is a quantity you calculate. It measures how much uncertainty about a target variable drops once a feature is known. Uncertainty is measured as entropy, a concept from Claude Shannon's information theory.

The formula is IG(T, a) = H(T) − H(T | a). H(T) is the entropy of data set T before the split, H(T | a) the expected entropy after splitting the data by feature a. The larger the difference, the better the feature separates the classes.

The quantity is best known from decision trees. The ID3 algorithm, which J. R. Quinlan described in the journal Machine Learning in 1986, picks the feature with the highest information gain at each split. Because the measure favors features with many distinct values, such as a customer ID, Quinlan proposed the gain ratio as a correction, dividing the value by the information in the split itself. The Gini index from the CART method is a different measure of impurity and often produces similar trees in practice.

It shares only the basic idea with the editorial meaning: how much does something add?

Where the term stops working

Information gain describes a comparison with existing answers and guarantees neither rankings nor citations in AI answers. It differs from E-E-A-T, which asks whether content and source are trustworthy, while information gain asks whether a piece adds something new; good content needs both. The editorial term shares only the basic idea with the machine learning quantity of the same name.