How to get Plex recommendations based on watch history

Short answer: Plex records everyone’s watch history, but it doesn’t use it to build the rows on your server. Recommended, Home and every pinned collection are library-wide — identical for every account. Getting rows built from a person’s own viewing means reading that history yourself and creating collections from it.

This page covers what Plex genuinely does, the two manual approaches people are usually pointed at, and where each one runs out.

What Plex does with watch history today

Plex tracks watch state per account, and it does use it — just not where you want it.

Continue Watching and Up Next are personal. These are the only genuinely per-user rows on your server, and they only ever contain things you’ve already started. They answer “where was I”, not “what next”.

Discover recommendations are personal, but off-server. If you’ve enabled it, Plex builds suggestions from your account activity and shows them under the Discover source. They’re mostly titles on streaming services, not things in your library, and the setting that controls it lives in your Plex account profile rather than your server.

Library rows are not personal at all. Recommended, Home shelves, genre rows, “Recently Added”, and anything an admin has published — every account with access sees the same rows with the same titles. Plex has no per-account library personalisation, and no setting anywhere turns it on.

That last point is the whole problem. On a server with any number of users, the person who has watched every sci-fi film you own and the person who has watched nothing but comedies are looking at an identical home screen.

Manual approach 1: smart collections

This is the advice you’ll get most often, and it’s worth understanding exactly what it buys you.

A smart collection is a saved filter. In Plex Web: open a library → Collections → create a collection → set it to Smart, then pick criteria — unwatched, genre, rating above some threshold, added in the last year, and so on. Plex keeps it up to date as the library changes, and you can promote it to the Home or Recommended shelf.

Unplayed = true
AND Genre = Thriller
AND Audience Rating >= 7.5
AND Year >= 2015

What it gets you: structure. “Highly rated thrillers you haven’t seen” is a genuinely better starting point than an alphabetical wall of posters, and it costs nothing.

Where it stops: the filter is evaluated against the library, not against a person. Unplayed means unplayed by the admin account — smart collection criteria don’t follow the viewer. So every user sees the same thriller row, including the ones who’ve already watched everything in it and the ones who hate thrillers. You’re organising the library, not personalising it.

That’s a real improvement and it is where most people should start. It just isn’t recommendations.

Manual approach 2: read the history and build collections yourself

The data you need does exist and is reachable.

Each account’s watch history lives on the Plex Media Server and can be read per user, which is what every tool in this space is doing under the hood. There are two routes:

• The PMS API, using each share’s own token. When someone accepts a share, that share carries a token scoped to them, and history read with it is genuinely that person’s. This is the accurate route, and it’s the one Shortlist uses.
• Tautulli. Tautulli has watched your server for as long as it’s been installed and exposes per-user history over its API. Widely used and easy to query. Its identifiers are display names rather than stable account IDs, which matters if anyone on your server has ever renamed themselves or shares a name with someone else.

From there the shape of the job is: take what a person watched, find similar titles that are already in your library, drop anything they’ve seen, and put the result in a collection. Similarity usually comes from TMDB — shared genres, keywords, cast, crew — or from a recommendations API like Trakt’s, optionally with an LLM ranking the shortlist at the end.

Where it stops: the collection you just built is visible to everyone with access to that library. You’ve made a personal row and published it to the whole server, which is both a privacy problem and a clutter problem — twenty users means twenty rows on everyone’s home screen. Fixing that is a separate mechanism, covered in per-user collections.

What “good” looks like

If you’re evaluating approaches — your own script or someone else’s tool — these are the things that actually distinguish a usable result from a demo:

Picks must exist in your library. Any approach that asks a language model “what should this person watch?” and trusts the answer will produce titles you don’t own, titles that don’t exist, and titles under slightly wrong names. Generate candidates from your library and use the model to rank them, never to invent them.

History has to be per person, not per server. Reading the admin’s history and calling it everyone’s is the most common shortcut, and it produces one recommendation set wearing several names.

Rows need to refresh, and to change when they do. A row rebuilt nightly from an unchanged history should still shuffle its picks, or people stop looking at it after a week.

Say why. “Because you watched Arrival” is the difference between a row people trust and a row that looks arbitrary. It’s also how you debug a bad pick.

Handle the person who’s watched nothing. New users have no history. Falling back to library-popular or recently-added is fine; producing an empty row is not.

The automated version

Shortlist is a self-hosted container that does this for every user on your server, on a schedule. It reads each person’s own watch history through their share’s token, finds similar titles verified to exist in your library, and builds them a “Picked for You” collection that only they can see.

Every pick carries its reason. AI is optional — the built-in picker runs entirely in code with no keys and no cloud — and if you do enable a provider (Claude, GPT, Gemini, or a local model via Ollama, llama.cpp, LM Studio, vLLM or LocalAI) it only ever ranks and explains candidates drawn from your library.

docker run -d --name shortlist -p 5959:5959 \
  -v /path/to/config:/config \
  stevezzau/shortlist:latest

Set -e SHORTLIST_DRY_RUN=1 to see every change it would make without writing one.

Related

• How to improve Plex recommendations — the settings to change first
• AI recommendations for Plex — where a model helps, and where it invents films
• Per-user collections — making a row only one person can see
• A different home screen per user — the surfaces rows appear on
• Plex recommendation tools compared — the other projects in this space
• What goes in a row — tuning sources and seeds in Shortlist