Tuning Your Ranking

Every article gets a score, and the feed is that score sorted. The scoring is not a secret — the ranking code is open source — and everything below describes what it actually does.

What moves an article up or down

Each article starts at a baseline and is adjusted by a handful of signals.

Recency decays logarithmically. That is the difference between "hours old" and "days old" mattering a lot, and "nine days" versus "ten days" mattering almost not at all. It never goes negative, so an old article is not punished, only less boosted.

Your interests push matching articles up. A topic you asked for is checked against the title, summary and category.

Your muted topics push matching articles down, and this is the heaviest lever in the system — several times stronger than an interest of the same weight. That is deliberate: "less of X" should be felt immediately, and usually is, whereas "more of Y" competes with everything else already scoring well.

Sources and bylines you prefer get a boost. A byline counts for more than a publication, because naming a writer is a narrower statement than naming an outlet and should not be outvoted by one.

Diversity stops one source or one writer taking over the page. There are caps per source and per author within a single view, so a publication having a busy day cannot crowd out everything else.

Similarity to what you read — the ranking compares articles against what you have saved and liked, so it can surface something relevant that matches none of your stated words.

Every article carries its reasons. If something is at the top and you cannot see why, the reasons say which of the above put it there.

Four ways to change it

Say it to the assistant. The fastest. Open AI Assistant and write plainly:

more AI infrastructure and chip supply, much less crypto and celebrity gossip

It proposes the change, you press Apply. Nothing is written until you do.

Set it by hand. Algorithm & Agents exposes the same things as lists and sliders — interests, muted topics, languages, and the weights themselves. Anyone who wants to move the knobs directly can; the assistant is a convenience, not a gate.

Let your agent do it. A connected assistant like ChatGPT or Claude can tune the ranking on your behalf when you ask it to. See connecting an assistant.

Read. Saving and liking feed source and topic preferences back into the ranking without you writing anything.

All four write the same thing. There is no "advanced mode" holding anything back.

Interests versus muted topics

Muted topics are the stronger tool, and usually 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. "Nothing about crypto" removes a category of noise in one line, and because the penalty is heavy, you feel it on the next refresh.

One caution on wording: topics match as substrings. A short entry like ai also fires inside "said", "detail" and "again", which makes it worse than useless. Prefer specific multi-word topics — ai infrastructure, celebrity gossip — over single short words. Bylines are the opposite: they match exactly, so spell the name the way the publication does.

Every change is a version

A ranking change never overwrites the previous one. It creates a new version, and the old one stays.

That means you can experiment without care. Turn recency all the way down for a week of depth over news; if the feed goes strange, roll back and you are exactly where you were. Nothing is lost and nothing needs remembering.

It also means an agent can tune your ranking without you having to trust it in advance. If a connected assistant makes a change you dislike, the previous version is one action away — which is why the skill instructions tell agents to act rather than ask permission for each weight. The safety comes from reversibility, not from asking.

What the ranking never does

No engagement optimisation. Nothing is boosted for being provocative or for keeping you scrolling. The scoring above is the whole of it; there is no term for how likely you are to react.

No advertiser anywhere in it. Nobody pays to rank higher. There is no inventory to sell.

Your reading is not sold, and it is not an ad profile. It is used to rank your own feed, which is the only thing it is for.

No trending tab. What is popular elsewhere does not enter your ranking. Trending Radar is different and deliberately narrow: it watches your sources for a story several of them picked up at once.

If the feed does not look right

I changed something and nothing moved — weights apply to the next draw. Refresh Recommended.

A muted topic is still showing — check the wording matches how articles phrase it. Muting crypto will not catch a piece that only ever says "digital assets".

Everything is from one source — diversity caps are per view, so this usually means few sources rather than a ranking fault. Add more; the caps do the rest.

It got worse and I do not know which change did it — roll back to a version you remember liking, in Algorithm & Agents. Working forwards from a good state beats diagnosing a bad one.