The report a founder actually reads: what went in, what came out, where the funnel leaks, and who to call before lunch.

The business story

Why this system exists

A solo marketing operator at a B2B SaaS can't open Google Ads, Meta, the CRM, the email tool, and web analytics every morning and still notice the thing that matters. The numbers existed; the synthesis didn't. Meanwhile the founder's question was never "what was our CTR" — it was "are we getting demos, and if not, where exactly does it break?"

What I built (the production original)

A scheduled job that pulled every channel's API each morning and sent one email at 08:00 IST, framed Input → Output → Outcome — spend and sends first, leads second, and demos booked as the number the whole report answers to. It ran on managed infrastructure (not my laptop), kept a daily funnel snapshot for honest day-over-day deltas, and had a watchdog that alerted if the report itself failed to send — a report you can't trust to arrive is a report you stop reading.

The judgment calls

• Outcome over output. Clicks and impressions appear only as context. The subject line carries demos booked, because that's the number a founder acts on.
• Refuse stale comparisons. If yesterday's snapshot is more than 36 hours old, the report shows no deltas and says why, loudly. A missing number gets investigated; a quietly wrong number gets believed.
• Silent degrade → loud truth. Every data source prints its fetch time. If a section can't be computed honestly, it says so instead of rendering something plausible.
• Every recommendation is labeled a hypothesis. "Shift budget to Google" is a bet, not a fact — the label keeps the report honest about the difference.
• Small denominators are never "leaks." Three leads converting at 0% is noise. Flagging noise trains the reader to ignore the flag that matters.

What went wrong along the way

An early version computed day-over-day deltas against whatever snapshot existed — until a dead snapshot job made a flat week look like growth. That incident produced the freshness guard this demo runs (try the stale-data scenario above). A later audit also removed a WhatsApp inbound metric entirely: the counting method was misleading, and omitting it honestly beat reporting it wrong.

What I'd change today

The baseline should be a rolling computation with seasonality awareness rather than a fixed window, and the ranked actions should track their own hit rate — did yesterday's "call this lead" hypothesis convert? — so the report learns which of its own suggestions deserve rank.