How Typographer works.

It starts with the SERP, not a keyword list

When you enter a target keyword, Typographer pulls the top ten organic Google results through SERPAPI, with the People Also Ask questions and the related searches. It then fetches up to five of those ranking pages, strips them to text, and counts their length to set a word target. The target is the average length of the pages that rank, with a sensible floor and ceiling around it.

A language model reads those pages as an analyst, not a keyword counter. From the pages that rank, it extracts the substance a competing article needs: the named entities the winners reference, the subtopics that form the topic graph, the questions readers expect answered, the authoritative sources the winners cite, and the original research that would let you beat them. Every ranking page also gets a coverage score, so you can see how complete the competition actually is.

Coverage, not keyword density

Your draft is measured against that map on three axes: entity, topic, and intent. Not how many times you repeat a phrase. Coverage is checked two ways. A literal pass looks for each item in your text. Then, after you pause typing, a semantic pass embeds your draft and compares it to each guideline item with a vector database, so a draft that says “accredited” earns credit for the entity “accreditation” even when the exact word never appears. An item counts as covered if either pass finds it, so the score never punishes you for writing like a person instead of a machine.

The semantic layer uses OpenAI embeddings and cosine similarity in Postgres with the pgvector extension, counting an item covered above a set cutoff. Item embeddings are cached per keyword, so only your draft is re-read as you write.

The score: three pillars

The overall number is weighted by design: information gain takes the largest share at forty percent, with voice and readability at thirty percent each. Information gain carries the most weight because it is what gets a page cited.

Information gain rewards what a reader cannot find on ten other pages: named sources you can check, concrete numbers and dates, specific proper nouns, real quotes. It subtracts for the opposite. Vague quantifiers like “several” and “a range of”, weasel attribution that names no source, and hedging all pull the number down. Linked and attributed sources count most, then hard data, then specificity.

Voice begins at the top of the scale and subtracts for the machine habits that also happen to be weak writing. It runs the same word lists as our editorial linter: empty vocabulary like the word “delve”, throat-clearing that delays the point, fake-profound phrasing that inflates a small thing into a milestone, the correlative reflex that denies one claim to assert a bigger one, the triad reflex, and the dash habit. It adds a small bonus for the marks of a real writer: first person, contractions, varied sentence length, a question now and then.

Readability models how the prose reads, not a grade-level shortcut. It looks at sentence length against a target near seventeen words, the variance between long and short sentences, the share of passive voice, adverb density, and how many sentences you pack into a paragraph. Flat, even rhythm is the tell it punishes hardest, because that is how machines write.

Net-new, not just covered

Coverage tells you whether you touched the topics that rank. It does not tell you whether you said anything new. So the information-gain pillar adds a measurement the older tools skip. Typographer embeds the ranking pages and your draft, then scores the share of your draft that sits far from everything the winners already said. A high net-new number means you are adding information the SERP lacks, which is the signal Google’s own information-gain work rewards. The same pass marks every required entity and topic as baseline, meaning most competitors already cover it and it is table stakes, or as a gain opening, meaning few of them do and it is yours to take. It also pulls the external sources the winners link to, so you can cite the same primary data or beat it.

Readability shows its work too. A sentence-rhythm strip charts the length of every sentence as a bar. Machine writing is a flat wall of even bars. Human writing spikes and dips, a short line after a long one. Seeing the shape makes a low readability score obvious.

The adversary, run in reverse

The voice and readability pillars are our prose adversary turned inside out. Normally that adversary gates a draft. It hunts mechanical tells with real pattern matches and refuses to pass the piece until they are gone. Typographer runs the same checks live and turns them into a score you watch move as you write. A clean score proves the absence of those tells. It does not prove your facts are true. That part stays with you.

Completions that have to pass their own test

When you ask Typographer to draft the parts you are missing, it writes in a human voice, then attacks its own output. It runs the tell detector and the readability check on what it just wrote, sends the draft back to revise up to twice, and returns only prose that clears the bar. If an average sentence runs too long, or the rhythm goes flat, it gets fixed before you ever see it.

What the score does and does not promise

Coverage and a high score make a page competitive. They do not guarantee a ranking, and they were never meant to. Google weighs inbound links, domain authority, freshness, and a hundred other signals no editor controls. What Typographer controls is the part on the page. Whether you cover what the winners cover, add something they do not, and sound like a person worth citing. Get that right and the rest has something to work with.