Cadence, Volume and the Slop Problem

The case against publishing more because it became cheaper is not a moral one. It is arithmetic: the cost of a send fell to nearly nothing while the cost of reading did not, unsubscribing is one-way, and additional sends add reach that saturates while adding churn that does not. All of that is worked below.

The claim

Language models made producing a plausible newsletter issue nearly free. The obvious response is to send more. The argument here is that for most publications this is a losing trade, that it is losing for reasons you can calculate rather than reasons you have to believe, and that the exceptions are identifiable in advance.

The underlying asymmetry: what a newsletter provides is not text. It is selection — somebody deciding that these seven things are worth your attention and the four hundred others are not. Generation got cheap. Selection did not, because selection is judgement and judgement is the thing a model cannot supply about your subject and your readers.

The cost curve moved on one side only

Two costs govern this market. The cost of producing an issue, borne by the sender, and the cost of reading one, borne by every subscriber. The first collapsed. The second is unchanged and is measured in the only genuinely fixed resource in the system.

When one side of a market gets much cheaper and the other does not, the scarce side becomes the one that sets value. Attention is now the binding constraint, and the operation that conserves it is refusal. A publication that increases output has increased its supply of the abundant thing while consuming more of the scarce one, which is the shape of a bad trade regardless of how good each individual issue is.

Cadence multiplies churn linearly

Unsubscribing is monotonic. A subscriber can leave and cannot leave twice, and almost nobody resubscribes. That makes list size a stock depleted by a flow, and the flow is proportional to how often you send.

Assumptions, all substitutable:
  unsubscribe rate per send   u = 0.3%   (a normal, healthy figure)
  sends per month             s

Monthly retention = (1 − u)^s
Annual retention  = (1 − u)^(12s)

  s = 1  →  monthly 99.70%   annual (0.997)^12  = 96.5%
  s = 2  →  monthly 99.40%   annual (0.997)^24  = 93.0%
  s = 4  →  monthly 98.81%   annual (0.997)^48  = 86.6%
  s = 8  →  monthly 97.62%   annual (0.997)^96  = 74.9%

Annual loss, no new subscribers:
  monthly   3.5%
  weekly   13.4%
  twice a week 25.1%

On a list of 10,000, standing still requires gross additions of:
  monthly       ~350 a year
  twice a week  ~2,510 a year

The last two lines are the whole treadmill. Going from monthly to twice weekly means finding roughly seven times as many new subscribers every year simply to stay the same size — and acquisition is the expensive activity, which means the cadence decision quietly became an acquisition-budget decision.

Two honest caveats. Unsubscribe rate is not constant across cadence: a publication that raises frequency without raising quality usually seesu rise as well, which makes the table optimistic. And a small number of very engaged subscribers will tolerate a much higher frequency, which is exactly why the exception in a later section is real.

Reach saturates

The case for volume is reach: send more often and more of your list sees you at least once. That is true, and it has sharply diminishing returns while the churn above has none.

Assume a subscriber opens any given send with probability o = 0.35,
independently. Probability of seeing at least one send in a month:

  s = 1  →  1 − 0.65^1 = 35.0%
  s = 2  →  1 − 0.65^2 = 57.8%
  s = 4  →  1 − 0.65^4 = 82.1%
  s = 8  →  1 − 0.65^8 = 96.8%

Marginal reach per doubling, against annual list loss:

  1 → 2   +22.8 points of reach    3.5% → 7.0% annual loss
  2 → 4   +24.4 points of reach    7.0% → 13.4% annual loss
  4 → 8   +14.7 points of reach   13.4% → 25.1% annual loss

Reach is bounded above by 100%. Churn is not bounded below by anything.

The independence assumption is generous and the direction of the error matters. Opens are strongly correlated within a person — some subscribers open almost everything and some open almost nothing — so real incremental reach is lower than the binomial suggests. The calculation is therefore the best case for volume, and the best case still turns over somewhere around the fourth send.

What a newsletter actually sells

Consider what changes for a reader when a publication doubles its output with the same editorial effort. Each issue contains, on average, half as much of the thing they subscribed for. The rejection rate — the proportion of candidate material that did not make it — halved, and the rejection rate is the product.

This is why the machine helps least exactly where publishers most want it to help. It can draft the write-up of an item you selected. It cannot tell you that the item everyone else is covering is not worth covering, because that judgement depends on knowing what your readers already know, and the most probable continuation is by construction the consensus one. That is the same division that runs through every writing stage: safe wherever the context already contains the answer, unsafe wherever something has to be supplied.

There is a second-order effect worth naming. If drafting is cheap, the marginal item costs almost nothing to include, so the threshold for inclusion drifts down without anybody deciding to lower it. Nobody ever resolves to publish worse things; they just stop having a reason to say no.

The same asymmetry shows up one level down, in the subject line. There is no way to select a subject line by experiment at ordinary list sizes, because an adequately powered test needs thousands of impressions per variant. So even the smallest editorial decision in a newsletter is judgement exercised without evidence, which is a reason to have fewer of them rather than more.

The feedback loop that punishes twice

Beyond the reader there is an intermediary, and it makes the trade worse in a way that is not obvious from the arithmetic.

Mailbox providers decide inbox placement partly from engagement signals — opens, replies, deletions without reading, complaints — aggregated per sender. The exact weightings are not published by any provider and change; the general mechanism is documented widely enough in sender guidance to plan around, and it is the mechanism rather than the weights that matters here.

The loop runs like this. More sends at the same effort means lower average engagement per send. Lower engagement worsens placement for subsequent sends. Worse placement lowers engagement further. Volume therefore does not just cost subscribers directly; it degrades the delivery of the issues you still send to the people who still want them.

The observable consequence for a publisher is a lag that makes the problem hard to see in time. Readers disengage well before they unsubscribe, so open rate falls first and list size looks healthy for months. By the time list size moves, the cause is a quarter old.

Where volume genuinely works

The argument above is about a publication whose value is selection. Not every newsletter is one, and the exceptions are specific.

• Where more items means more selection. A curated link roundup that goes from ten to twenty items per week has, if the candidate pool grew accordingly, raised its rejection rate rather than lowered it. The test is whether the pool grew with the output.
• Where the content is a record rather than a judgement. Changelogs, status notices, transactional summaries, market opens. Nobody subscribed for the editing, and completeness is the value.
• Where every issue is requested. Per-topic subscriptions and digest controls turn cadence into the reader’s decision, which removes the trade entirely.
• Where the subject genuinely produces more. A beat in a period of real activity supports more issues, because the material is there. This is the only exception that expires, and it expires quietly.

The common structure: volume is safe when it tracks the supply of things worth saying, and unsafe when it tracks the supply of capacity to say things. Machine assistance moved the second and left the first alone, which is why it makes this mistake so much easier to make.

Four signals you have crossed the line

• Nothing was rejected this month. If everything considered was published, no selection occurred, and selection was the product.
• Open rate is falling while list size is flat. The leading indicator, and typically a quarter ahead of the lagging one.
• You could not say what any recent issue argued. Issues that covered a subject rather than claiming something are issues a reader has no reason to remember.
• The reply volume dropped faster than the open rate. Replies are the strongest engagement signal available and the first to go when readers move from reading to skimming.

The remedy is unpopular and simple: publish less and reject more. Machine assistance is entirely compatible with that — use it for the write-up, the transcript, the structure, the checks — as long as it is not used to lower the bar for what gets in. What that bar looks like in a single piece is the subject of the next page.