# Judging and de-duplication (/docs/prism/discover-judging) See how keywords match, how each post is scored against your question and how duplicates are removed A Discover row is kept when it is inside your date window, passes the relevance check and is not a copy of a row already kept. This page explains each step, so you can tell why a post was kept or dropped. Dropped rows carry a `drop_reason`. Every other request field is on [Discover request fields](/docs/prism/discover.md). ## Is a keyword an exact match? No, and not a wildcard either. Each keyword is sent to every platform's own search, which matches words the way that platform does: most return posts containing the words or close variants, ranked by the platform. With `expand=1` to `3`, discover also plans extra searches from your keywords and `context`. Then every row is read by the judge against your keywords, or against `context` when you send one, so rows about something else are dropped under `relevance=filter`. `literal_match` on a row says whether a keyword appears in its text as written, and `exclude_terms` and `contains` filter on the text exactly. ## Score each post against your question `context` says what kept rows must be about: a persona or a question goes there, never in the keywords. Each row is scored from 0 to 1 against it, in `judgments.relevance.p`. When `context` is the same as the keywords, the score is worked out by code and costs nothing. A `context` that differs is read by a model and billed, and a dry run shows that cost in `by_stage.judge`. `relevance` decides what the score does. `filter`, the default, drops rows below `relevance_threshold`, which you can set anywhere from 0 to 1. `score` keeps every row with its score, so you can sort and cut yourself. `off` skips the check. ## Add labels `label` adds judgments to each row, and each one is billed: `mention` (it needs `brands`), `intent`, `niche`, `sponsored_model` or `stance`. `intent` also filters: `intent=complaint,question` keeps only rows with those intents. What each label means, and which labels add no credits, is on [Labels](/docs/labels.md). ## Ask your own questions `questions` takes up to five questions of your own, as JSON keyed by the answer's id. They are judged on every account, and on every row with `view=posts`. A question about a sensitive trait of a person is refused as a free `400`. ## Remove duplicates De-duplication is done by code and is free. Rows with the same canonical URL are one row, however many searches or groups found them: the first one seen stays and gains the other copies' `matched_keywords`, and the other groups go into `also_in_groups`. A row whose text is nearly the same as a kept row is marked `state: "duplicate"` with `drop_reason: "near_duplicate"` and `duplicate_of`, so a story syndicated on five sites keeps one row. Two texts that differ only in their own numbers, such as the same notice for two dates, are not treated as copies. A row already shown to you is never taken back. ## Next steps - [Discover overview](/docs/engines/discover.md): Price a run for free and read its rows. - [Labels](/docs/labels.md): What each label means and what it costs. - [Runs, results and cost](/docs/prism/discover-runs.md): Follow a run and read why rows were dropped.