Question 1

Why not just ask a chat model for an SEO report?

Accepted Answer

A chat model returns confident prose with a number you cannot defend and a different number on retry. Here the rubric levels are in the repo, Jev returns probabilities plus separate confidence, and changing a weight re-scores without new tokens.

Question 2

What happens when Jev is unsure about a judgment?

Accepted Answer

It is flagged in the report as needing a human check instead of being averaged in. Confidence is a separate axis from the answer, so low-confidence calls stay visible rather than becoming a tidy number.

Question 3

How do single-page and whole-site modes differ?

Accepted Answer

Single-page mode pastes a URL or draft and returns three scores with fixes. Site mode crawls one origin using robots.txt and sitemap, scores every page on the same signals, assigns leave, improve, rewrite, merge, split, or prune, and builds one ordered plan.

Question 4

Where do my keys and crawl data actually go?

Accepted Answer

It runs locally with keys on your machine or in your browser for the hosted demo. Data lives in one SQLite file via built-in node:sqlite, Postgres via DATABASE_URL for deploys, and production instances disable file-based credential writes.

Question 5

What do the three scores actually measure?

Accepted Answer

Three separate targets rather than one SEO number. SEO asks whether the page serves the intent behind the query and says something the other results do not. AEO asks whether a machine can lift a correct, complete answer out of it and show that on its own. GEO asks whether there is anything worth quoting and enough provenance to justify citing it. They are scored apart because the fixes differ and sometimes conflict.

Question 6

How much does a page cost, and what keeps it down?

Accepted Answer

Roughly two seconds and £0.0001 per page. One request answers all 25 questions, a typical page is about 4,000 input tokens at $0.042 per million, and output tokens are free. Batching is the reason, since the TypeSafe docs measure it as roughly 12 times cheaper and 10 times faster than asking one at a time. Judgments are then stored, so changing a weight re-scores instantly without spending another token.

Question 7

What is the rule behind the codebase?

Accepted Answer

Code owns the rules, the model owns the meaning, code owns the policy. Anything with a fixed threshold, a count or a computable yes or no is a rule in src/lib/extract/rules.ts, anything needing reading comprehension is a judgment in src/lib/judge/dimensions.ts, and any weight, blend or cut-off is policy in src/lib/score/compose.ts. Title length is not a question for a model; whether the writing shows first-hand experience is not something you can regex. MIT licensed.

Jevseo

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community