Editorial and methodology
Purpose
JERP exists so an agent can be strict about homepages: real offer vs noise, fit to a named category, relationship between two sites, or a question you specify — each with a confidence score. We would rather return “unsure” than a fluent wrong list. This page is how we produce those numbers, not a claim that they replace a crawl of the live web.
Inputs
Cards you send on keep, fit, compare, rerank, and judge: domain, title, summary, optional categories. We do not fetch the live HTML of those URLs on the free judgment path.
Index search on find and discover: query or seed against our homepage index. Matches are ranked inside that index, not as Google positions. firstIndexed (and legacy wentLive) is crawler first-seen or recrawl — not founding year. Use businessFoundedYear when present if you care about company age.
Pipeline
Work is cascaded so most noise dies before an expensive model runs: local heuristics (known aggregators, parked patterns, platform shells), then lexical/semantic triage against the seed or category, then model-based viability and relationship labels. Internal cost target is under $0.0008 per site on the hot path; you still pay your chat provider for the short tool round trip. Free-tier JERP calls are $0 within caps.
Every judgment field can carry confidence. Thin titles and empty summaries should score low confidence. Do not treat a high score on a one-line card as a completed investigation.
What the tools mean
keep — viable commercial homepage vs noise (directories, parked, filler), niche-agnostic: shops, wholesale, SaaS, and services can all be viable.
fit — match to the category string you provided, not a global taxonomy we invented.
compare — relationship between left and right: same offer, lookalike, platform/aggregator, or unrelated — not “which brand is better.” Catalog and audience overlap are the competition question; geography alone is not. Score table: docs.
rerank — order the cards you sent against a query you sent.
judge — your yes/no, choice, or score questions on those same cards.
find / discover — index retrieval and grouped discovery from a seed; still not a SERP and not citation share in ChatGPT.
Corrections and versions
Where the API includes policyVersion or similar, that is the rule set that produced the row. Models and heuristics change; two calls a month apart can disagree. If a score is systematically wrong for a class of sites, tell us via contact with example domains — not with production keys.
Third-party index tools
When we mention Ahrefs or Semrush, we are contrasting a crawl/index product with JERP’s judgment layer. Their methodology is theirs, for example Ahrefs crawler overview and Semrush data documentation. Token prices in the homepage comparison table are list prices from provider docs as of the date on that table, not a billing quote from JERP.
Independence
Scores are not paid placements. We do not sell “guaranteed keep” for a domain. Editorial independence here means the model stack and heuristics are the same for every caller on the free tier, within rate limits — not that the models are infallible.