The Substack Notes algorithm: what we know

There is no published Substack Notes ranking model. Nobody outside Substack has seen one, and anyone selling you "the 7 signals the algorithm rewards" is describing their own last good week. This page separates three things that usually get mixed together: what Substack has stated publicly, what is mechanically observable in the product, and what is folk knowledge that we can only test rather than confirm.

Where we do not know, we say we do not know — and we give you the test to run on your own account.

1. What Substack has said publicly

Substack has positioned Notes, repeatedly and in public, as a feed built to send readers toward subscriptions and writers, not to maximise time in the app. Their stated business reason is straightforward and worth taking seriously: Substack makes money when readers pay writers, not when readers scroll. A feed tuned purely for dwell time would be misaligned with that.

The practical reading: signals that suggest "this reader would subscribe to this writer" plausibly matter more than signals that suggest "this reader will keep scrolling". A Note that produces a profile visit and a subscription is worth more to the platform than a Note that produces ten seconds of attention.

That is a direction of travel, not a formula. Treat it as the frame, not the answer.

2. What is mechanically true

These are not inferences. They are how the product works, and you can watch each one happen:

• The feed mixes your network with strangers. You see Notes from people you follow and subscribe to, plus Notes from outside that graph. This is why an account with 200 subscribers can reach thousands of impressions on a single Note — and why impressions alone tell you little about who saw it.
• A restack republishes your Note to someone else's followers. This is the amplification primitive. Everything that spreads on Notes spreads because somebody restacked it into a new audience.
• A reply is itself a Note. When you reply to someone, that reply is visible content in its own right, attached to a conversation that is already getting attention. This is why replies are the cheapest distribution available to a small account — you are borrowing an audience that already gathered.
• Followers and subscribers are different populations. Someone can follow you on Notes without subscribing to your publication, and vice versa. Growth on Notes is follower growth first; the conversion to subscriber is a separate step that happens on your profile and your posts.
• Notes are not emailed. No inbox, no open rate, no unsubscribe. The only distribution a Note gets is feed distribution — which is exactly why the first hour matters more than it does for a post.
• Stats are per-Note and public to you. Substack shows impressions, likes, replies, restacks, and where views came from. Everything in section 4 is measurable from data you already have.

3. Folk knowledge — and how to test it

These circulate as facts in every Substack growth thread. We have not been able to confirm any of them, and neither has anyone else outside Substack. Each comes with a test you can actually run.

"Links kill your reach." The claim: Notes containing an external link get suppressed. Plausible mechanism: a link sends readers off-platform, which is not what a subscription-oriented feed wants to encourage. Status: unconfirmed. Test: post twenty Notes over four weeks, ten with a link and ten without, matched roughly for topic and time of day, and compare median impressions — not the mean, which one viral Note will wreck. If the gap is smaller than the spread within each group, you have learned nothing, which is itself a useful result.

"Post at [specific hour] for maximum reach." Status: unconfirmed, and probably not a global constant. Your audience has a timezone distribution; the feed's average has nothing to do with it. Test: use your own Notes stats, group by hour posted, and look at median impressions per hour bucket over at least thirty Notes. Fewer than thirty and you are reading noise.

"The first hour decides everything." Status: directionally plausible, unconfirmed as a hard window. Every feed we know of weights early engagement, and Notes has no email fallback, so a Note that gets nothing early has nowhere else to go. Test: reply to every reply within the hour for two weeks and compare median replies per Note against the two weeks before.

"Long Notes are penalised." Status: unconfirmed, and confounded. Long Notes tend to be essays with no gap left for a reader, which suppresses replies for reasons that have nothing to do with ranking. Test the reply rate, not the length.

"Restacking others gets you restacked." Status: unconfirmed as an algorithmic effect, obviously true as a social one. People notice who amplifies them. That is enough of a reason without a ranking theory.

The honest summary: on the ranking questions, the confident answers you read online are guesses with better formatting.

4. What you can control

Ranking is not controllable. These are:

• Replies per Note. The one metric that is almost entirely a function of how you write. A Note that closes its own argument leaves nothing to answer; a Note that leaves a specific gap gets filled. This is the highest-leverage change available, and it is a writing change, not a posting-frequency change. The mechanics are in how to write Notes that get replies.
• Replies you write on other people's Notes. Free distribution into an assembled audience, with no ranking theory required.
• Whether the Note is restackable. A Note that is only interesting if you already follow the author cannot travel. A Note that carries its own context can.
• Your profile. Every impression that becomes a profile visit lands on a page you fully control. If that page does not say what someone gets by subscribing, the feed did its job and you did not.
• Consistency. Not because "the algorithm rewards consistency" — because you cannot learn anything from six Notes.

5. Running your own test properly

If you take one thing from this page, take the method:

1. One variable at a time. Changing hook style and posting time in the same fortnight teaches you nothing.
2. Thirty Notes minimum per condition. Note performance is heavily skewed; small samples are dominated by one outlier.
3. Median, never mean. One Note at 40 000 impressions will make any group look good.
4. Pick the metric before you start. Impressions measure the feed. Replies, profile visits and new followers measure whether it worked. If you are optimising for impressions you are optimising the wrong end of the funnel — see the numbers that matter.
5. Write the result down even when it is null. "No detectable difference" is the most common honest outcome and it saves you repeating the test in six weeks.

This is exactly the analysis Subshack does locally on your own Notes stats — medians by format, by hour, by presence of a link, computed on your machine rather than on someone else's average. It will not tell you the algorithm. It will tell you what your audience does.

6. What to do on Monday

Skip the ranking theory entirely for a month:

• Write one original Note a day with a gap in it — a missing item, an unexplained number, a position you could be talked out of. Templates to start from: 40 Substack Notes templates.
• Write five substantive replies a day under Notes from writers slightly bigger than you.
• Answer every reply on your own Notes within the hour, with a question rather than a thank-you.
• Track median replies per Note, weekly profile visits, and new followers. Ignore impressions.

If that sounds unrelated to the algorithm, that is the point. The reachable levers are upstream of ranking, and they work whatever the model turns out to be. The same logic applies to the shortcut everybody eventually tries — see growing without subscribe-for-subscribe.

The most controllable lever of all is the page the feed sends people to. The free Substack profile audit scores yours out of 100 and lists what to fix, in weight order — nothing algorithmic about it, which is exactly why it is worth doing first.

What would change this page

We will update it if Substack publishes ranking documentation, if a large enough public dataset appears, or if our own tests reach a sample size where a median difference holds up. Until then, sections 1 and 2 are what we know, section 3 is what nobody knows, and the rest is what to do about it.