Rebuild the number from its inputs

Four of our recent Quebec projects turned on a number that did not hold up. A retargeting line overstating delivery by 400,000 impressions. A results model claiming double the entire campaign goal. A tracking install that was live, transmitting, and sending the wrong event. A pilot's zero that was unmeasured rather than nil.

None of those were caught by looking harder at a slide. They were caught by rebuilding the number from its inputs, and by reading the platform's own record instead of the report on top of it.

The line that never followed

For the Journées de la persévérance scolaire, the draft plan carried a retargeting line where the CPM had been raised from $5.00 to $9.00 and the impression row never followed. The result: 400,000 impressions of delivery that would never exist. The same deck had three different phase splits across three slides, and a "30 / 70" that was 28 / 72 in the tables it was drawn from.

We audited it line by line in code, so every figure was reproducible rather than argued. The whole estimate now regenerates from a single script.

The promise nobody could keep

For Dans la rue, the draft results model had digital delivering between 170% and 219% of the entire $100,000 campaign goal. Rebuilt from scratch, it showed four independent defects, and underneath them a 4.8% to 6.9% visit-to-gift rate, which is warm branded-traffic territory, when cold paid social to a donation page runs 0.5% to 2%.

The corrected line: digital contributes roughly 13% to 27% of the goal in the campaign year. One result survived every assumption. The longer scenario's monthly donor base is 2.0 times the short one's in all three conversion cases. Dollar totals collapse as you get prudent. The ratio does not move.

The zero that was not a zero

A pharmacy pilot out of Cité Mirabel was reporting zero sign-ups and was about to be called a failure. The registration event had fired exactly twice, on the same day, matching the developer's acceptance test. Close to 10% of the dataset's volume was coming from a development environment. We reported the zero as unmeasured, not nil, and put publishing the production tags ahead of every budget decision. Meanwhile the ads were working: a 1.51% link click-through rate against a 1.5% target.

The install that looked fine

For a pharmacy management summit, the client's tracking was live and transmitting. Ads Manager showed no registrations. The event table explained it: the registration event had received two events in total, both from our own test. The install was sending form views, not registrations, so Meta had nothing to learn on.

From the client's side, that looks identical to a working install. The right question is never "is it firing". It is "which event, on which dataset, last received when".

The habit

Every number in a plan has inputs: a budget, a rate, a date, an event. If you cannot regenerate the number from those inputs, you do not know it yet. That check takes minutes. A number that does not hold up can stay in a client's memory for eighteen months.