Information gain is the most cited and least operationalized idea in current SEO writing. Everyone repeats the definition from Google’s patent. Almost no one says how to tell, in advance, whether the page you are about to publish has any. This page covers both: what the term means, where it comes from, and how to estimate it against the results already ranking.
Where the term comes from
Information gain is borrowed from information theory, where it has a precise formula. The information gain of a split is the entropy of the parent set minus the weighted entropy of the parts it produces: IG = H(parent) − Σ H(children). In machine learning it is the rule a decision tree uses to pick which feature to split on — the split that reduces uncertainty the most. That is the meaning behind most of the “information gain formula,” “entropy,” and “gini index” searches, and it has nothing to do with SEO.
SEO borrowed the name, not the calculation. The SEO sense comes from a Google patent, Contextual estimation of link information gain, which applies the same intuition — value the thing that reduces what you still don’t know — to ranking a sequence of documents rather than to splitting a dataset. Same root, different object.
What the patent actually describes
The patent scores documents in context. A searcher reads the first result, then the second; by the third, the system can estimate how much each new document adds on top of what the person has already read. Results that mostly restate earlier ones score low. Results that introduce information the earlier ones lacked score high.
The unit being judged is not your page in isolation. It is your page given the set already ranking. This is the part most SEO advice skips, and it changes the job: you are not writing the best standalone article on a topic, you are writing the document that adds the most to the ones a reader has just seen. The reference set is the SERP.
Why you can’t calculate your own score
People search for an information gain calculator. There isn’t one, and there cannot be a real one. Google’s score depends on the corpus, the session, and the order a given searcher encounters results in — none of which you have access to. Any tool claiming to output a literal information gain number is estimating, not measuring.
What you can measure is the proxy the score rewards: the topics and entities present in the ranking set versus the ones present in your draft. The difference between those two sets is your estimated information gain. That is a graph operation, and it is exactly what content gap analysis performs.
How to estimate information gain before you publish
- Graph the ranking set. Take the pages currently ranking for the query and build one combined topic graph from them. This is the “already seen” baseline — the information a searcher has after reading the SERP.
- Graph your draft. Build the same kind of graph from the content you intend to publish.
- Subtract. The topics and entity relationships present in your draft but absent from the ranking set are your information gain. The topics present in both are table stakes — necessary for relevance, worth nothing for novelty.
- Decide. If the difference is empty, the page adds nothing and will struggle no matter how clean the optimization. If the difference is a coherent cluster, that cluster is the page’s reason to rank — lead with it.
In practice you often don’t need a draft at all. Build the combined graph of the ranking set and its gaps show up on their own: the topic clusters every result orbits, and the weak bridges between those clusters that none of the ranking pages develop. Those under-built bridges are the content gap for the query — the information the whole SERP collectively leaves out. The page that fills one has high information gain by construction, because nothing in the set covers it yet. Reading the gap straight off the SERP graph is how you decide what to create, before writing a word.
KeywordGraph runs this as one workflow: import the SERP, build the combined graph, and read the gaps in it directly. Overlaying your own draft afterwards is the confirmation step, not the starting point. The output is never a score; it is the specific set of topics that make a page additive rather than redundant.
Information gain and AI search
The mechanism matters more for AI search than for Google, because retrieval-augmented systems assemble an answer from several passages and have an explicit incentive not to repeat them. A passage that restates what the model already pulled adds nothing to the synthesis and tends not to get cited. A passage that supplies the missing piece earns its place in the answer.
So information gain is the same bet on two surfaces. On Google it decides whether a result is worth ranking given the set. In an AI answer it decides whether a passage is worth quoting given the other sources. How LLMs pick sources covers the retrieval side; from content gap to LLM citation covers what to do once you have found the gap.
How KeywordGraph helps
The bottleneck in this workflow is the comparison. Reading ten ranking pages, holding their combined coverage in your head, and spotting what your draft adds is the part that does not scale. KeywordGraph does that comparison as a graph operation, so the estimate takes minutes instead of an afternoon.
- Import the SERP and read the gap directly. Pull the pages ranking for your query, merge them into a single topic graph, and content gap analysis surfaces the weak bridges between its clusters — the topics the whole result set leaves out. That is the content to create, found before you write anything.
- Overlay a draft to confirm. Drop in your content or a competitor’s and the graph shows whether it actually fills the gap or just restates the set. The difference is your estimated information gain.
- Get clusters, not a vague score. The gap comes back as named topic clusters with their bridging concepts, so you know exactly what to lead with — and whether the gain is a coherent angle or just noise.
- Run it from your AI client. The KeywordGraph MCP server exposes the same SERP analysis and gap detection to Claude, ChatGPT, and Cursor, so the model checks information gain against live results instead of guessing from its training data.
The output is never a number pretending to be Google’s score. It is the specific set of topics that make your page additive — the thing the score actually rewards. Run it free on a query you are about to write for.