Recommendation Settings, Control by Control

Tuning your ranking explains how the scoring works and gives you four ways to change it. This page is one of those four, gone through control by control: Algorithm & Agents → Recommendation Settings.

Nothing here is hidden from the assistant, and nothing the assistant does is unavailable here. Both write the same object. Use whichever suits the moment — talking is faster, this shows you every knob at once.

Changes are saved per section, with a Save button on each. They apply to the active persona, which is worth checking before you start.

Topic Preferences

Two lists: topics you want more of, and topics you want less of.

The muted list is the stronger of the two by a wide margin, and it is also the easier one to write — people find it hard to list what they want and easy to name what they never want to see again.

Wording matters more than you would expect. Topics match as substrings, so ai also fires inside "said", "detail" and "again", which makes a short entry actively harmful. Prefer ai infrastructure over ai. Bylines under Sources I prefer are the opposite: they match exactly, so spell the name the way the publication does.

Languages I prefer

Which languages the shared pool may draw from. Sources you subscribe to are never filtered by language — this only affects what Recommended tops up with.

The default is English. If you read in another language and never set this, sources exist that you will simply never be shown, and nothing about the screen suggests anything is wrong.

Recommendation style

Six sliders, each 1 to 10. They are the weights from the scoring, with the maths hidden.

Slider	What moving it up does
Prefer newer articles	Recency counts for more against relevance
Follow my interests more strongly	Your stated topics outrank everything else
Avoid things I dislike more strongly	Muted topics push down harder
Prefer sources I like	Favourited sources and preferred bylines rise
Learn from saved and liked articles	Your reading history counts for more
Learn from hidden or skipped articles	Negative signals count for more

The two "learn from" sliders are the ones worth understanding. They do not decide whether Feedpik learns from your reading — they decide how loudly. Turn the first one down and saving something still registers, just quietly.

A useful pair of extremes: Prefer newer high with Follow my interests low gives you a news wire. The reverse gives you a research feed where a good three-week-old piece outranks today's noise.

Quality filters

Four switches, all about what to leave out rather than how to order:

• Hide muted feeds — sources you silenced without unsubscribing
• Hide articles I already read
• Hide skipped articles
• Hide articles I gave negative feedback

These change what appears at all, so they are blunter than the sliders. Hide articles I already read is the one to think about: with it on, Recommended never shows you something twice, which is usually what you want and occasionally means a story you half-read at breakfast is gone by lunchtime.

Diversity and explanations

Avoid too many articles from the same author or source. On by default and worth leaving on. It caps how much of one view a single publication can occupy, so an outlet having a busy day cannot take over your feed. With few sources it does little; the caps only bind when there is something to promote instead.

Use semantic similarity to find related articles. Lets the ranking match on meaning rather than words, so a piece about "digital assets" can surface for someone interested in crypto without either word matching. Turning it off makes the feed more literal and more predictable.

Show explanations for recommendations. Adds the reason chips that say why something is in front of you. Costs 1 credit each time one is generated.

Learned from your reading

Read-only. The sources you engage with most, as the ranking currently sees them.

This is the page's honesty check: if the list does not look like what you think you read, the ranking is working from a different picture of you than you are, and that explains more feed complaints than any slider does.

Agent boundaries

Two switches governing what a connected agent may do:

Allow agent override previews — whether an agent may propose a temporary reordering at all.

Allow temporary rerank — whether it may apply one. A temporary rerank affects a view without touching your saved settings, so it is the least invasive thing an agent can do to your feed and the first thing to turn off if you want an agent that only reads.

Version history and the advanced section

Version history lists every version of your ranking, and any of them can be restored. This is what makes experimenting safe — turn recency to 1 for a week of depth over news, and if the feed goes strange, roll back to exactly where you were.

Advanced Profile Data exports the raw personalisation object as JSON, and can reset it. Most people should edit Topic Preferences instead; this exists for export, migration and getting out of a state you cannot otherwise explain.

Advanced AI explanation prompt replaces the wording used when the AI explains a recommendation. Note that this travels: a shared persona can carry the author's prompt, and it is stripped on install unless the installer opts in.

Raw draft shows the saved policy object as it is stored. Nothing to edit — useful when reporting something odd.

If something looks wrong

I saved and nothing moved — weights apply to the next draw; refresh Recommended. Also check you saved the right section: each has its own Save.

I changed settings for the wrong persona — likely, if you have more than one. The page always edits the active one. See personas.

A muted topic still shows — check the wording matches how articles actually phrase it. Muting crypto catches nothing that only ever says "digital assets", which is exactly when semantic similarity earns its place.

Everything is from one publication — diversity caps are per view. With few sources there is nothing to promote in its place; add more before blaming the setting.

I made it worse and cannot tell which change did it — roll back to a version you remember liking rather than diagnosing the current one. Working forwards from a good state is faster every time.