Week-over-week comparison

Two baselines, always

Versus last week answers what changed. Versus a trailing multi-week average — four prior weeks is a reasonable window — answers whether it matters.

Showing only the first is how teams end up rewriting a campaign every Monday in response to noise. Most single-week movement on a reply rate is sampling variation, and it reverses the following week without anyone touching anything.

When the two disagree — down against last week, up against the trailing average — the trailing comparison is usually the true reading, and the disagreement itself is worth a line.

Get the direction right

Up is good for acceptance, opens, replies and positive replies. Down is good for bounce.

An arrow coloured the wrong way on a bounce improvement is worse than showing no arrow, because it converts good news into an alarm. Encode the direction per metric rather than assuming higher is better.

Rates move in percentage points, counts move in units. Don't render a five-point rise in a reply rate as "+5%" — that's a different, smaller claim, and mixing the two makes a report untrustworthy in a way that's hard to pin down.

The first run

With no history there is no delta. Say "baseline week — trends start once there's more than one week of history" rather than rendering zeros, which read as "no change" and imply a comparison that didn't happen.

Two weeks gives a week-over-week comparison. The trailing average needs more; until then, show what exists and label it.

Re-running mid-week

A weekly review gets re-run — Monday, then again Thursday.

A refresh may only add to the week's counts, never reduce them. Keep the identifiers of what has already been counted and take the union with the fresh pull. Otherwise a reply that was counted on Monday and archived on Wednesday vanishes from Thursday's number, and the week's total falls while the team is doing more work, not less.

Counts reset at the week boundary, which is expected and should be visible — a Monday morning number near zero is correct, not a bug.

What to trend

Trend the counts that represent outcomes — positive replies above all — and the rates that diagnose the funnel.

Resist trending everything. A comparison against two baselines on twelve metrics produces twenty-four numbers, which is not a weekly review, it's a spreadsheet. Pick the handful that would change a decision.