How We Score Information Gain

Why we measure it: Pages that introduce vocabulary the top result does not use signal a different angle on the topic. We compare the top 12 unigrams and top 8 bigrams (after stop-word removal) and credit terms unique to your draft.

Limits: Term frequency is not topic relevance. We do not weight by IDF, do not embed semantically, and do not check whether your unique terms are actually on-topic. Treat lexical novelty as a 'different angle' indicator, not a quality score.

Why we measure it: We count four classes of fact-shaped tokens: 4-digit years, currency amounts, percentages, and time durations. Pages with more concrete quantities are more useful to extractive answer engines and easier for fact-checkers to verify.

Limits: We do not validate that the numbers are correct. A page full of fabricated statistics will score the same as a page citing primary sources. Always cite the source next to every number.

Why we measure it: We scan for 17 first-person research phrases ('we tested', 'our 2026 sample', 'internal data', 'we measured', 'over the past N'). When LLMs and Google's helpful-content systems compare two pages on the same topic, the one with provenance language is far more likely to be cited as a source.

Limits: Phrase matching does not prove the data is original. A page can claim 'we tested' and present nothing. Pair the markers with an actual table, chart, or methodology link.

Why we measure it: All else equal, deeper pages cover more facets of a query. We use sentence count rather than word count because sentence-level density correlates with how many discrete claims a page makes.

Limits: Length without substance is filler. We do not detect repetition, padding, or off-topic sections. A 3,000-word page about the wrong subtopic will out-score a 1,000-word page that is laser-targeted.

Why we measure it: LLMs preferentially extract from pages that already pose-and-answer the user's question. We count question marks and explicit heading patterns (#, ALL-CAPS lines) as a proxy for FAQ density.

Limits: We can only see structure in the text you paste. If your page has rich heading semantics in HTML but the pasted version flattened them, the score will be low. Use a markdown export when possible.