Send to fewer people, for reasons you wrote down.

With the announcement, not the database. Which subject it belongs to, which countries can act on it, which medium suits it — a funding round is a business-desk story, a product launch is a trade-title story, and they are rarely the same people.

Two filters carry most of the weight, and they are deliberately different things. A subject says what an outlet covers; prominence says how much reach it has. A 5,000-reader fintech trade title beats a 5,000,000-reader general portal on a fintech launch, which is why one never substitutes for the other.

Because a list of 4,000 that resolves to 812 recipients is telling you something, and discovering it after the send is discovering it too late. Audience resolution is subtractive and every subtraction is reported: suppressed, unsubscribed, bounced, already contacted, pattern-guessed, academic journal.

The same arithmetic runs at planning time and at send time, through the same function, so a number quoted while planning is comparable with the number on the send screen. Two code paths producing two counts is how a product loses an operator’s trust in one afternoon.

Because a rule that lives in a feature can be worked around and a rule that lives in the data model cannot. Every recipient row carries the token behind its own one-click unsubscribe. The global do-not-send list records what mailboxes told us — bounce, complaint, opt-out — and is kept separate from the verification pipeline’s flags, because they answer different questions. The sender checks both.

An operator opt-out always beats a later automated verdict. Somebody who asked to be left alone is not re-added because a vendor re-scored their address as deliverable.

Evidence that an outlet has covered this kind of subject, or that a journalist’s byline has appeared under it. Not a prediction that they will write about you. Records vary in freshness, and a name that matters should be checked against the outlet’s own page before a pitch is built on it.

The product refuses to hide that unevenness. A wrong classification is invisible until somebody builds an audience on it, so unclassified stays unclassified rather than being guessed into a category.

They did not ask to be in a database, and the design assumes that. Contact data is gated, unpublished and unindexed; the planning model is handed counts and vocabulary, never people; an objection is a stop rather than a preference.

The discipline that makes a press list defensible is the one that makes the database defensible: know where each fact came from, and be willing to withdraw it.