Are we judging ads with the wrong model?

Directionally right. Materially smaller than it looks. Not the cause.

Four facts

Window for the per-ad figures: last ~4 weeks (May 22 – Jun 18). CAC figures are Northbeam clicks-only first-time CAC for Facebook Ads.

What the two models actually mean

Clicks only gives an ad credit for a sale only when the customer clicked it before buying. It's the strict, conservative read — "this ad did the work of getting the click that led to the order."

Clicks + Deterministic Views gives credit for clicks plus sales from people who saw the ad (logged-in, identity-matched) and bought later without clicking. It's a superset — it always counts more orders, so the CPO it reports is always lower.

The gap widens automatically when clicks dry up

You noticed the two models agreed in July 2025 and April, then diverged hard after the break. That's not a coincidence — it's arithmetic. Three forces stacked up at the same time:

• CTR fell → the click count went toward zero. When an ad gets 1 click-buyer instead of 5, clicks-only CPO doesn't just rise, it goes unstable — divide spend by a number near zero and you get $600. Clicks+Views still has a healthy denominator, so it stays near $100. The ratio explodes.
• CPMs rose ~40% and spend concentrated → frequency rose. You're paying to show the same warm, logged-in audience more often. More impressions to identifiable people = more view-matches — even though they aren't clicking. View-credit inflates exactly when clicks fall.
• The audience drifted older / more passive (already confirmed). That's the textbook "scrolls past, logged in, buys anyway" profile: high views, near-zero clicks. It widens the gap further.

Facebook Ads, week by week through the break

The red dashed line marks May 10, when both the account broke and the team switched the scoring model. Read these as small multiples — each panel is one signal over time.

Read it straight: the gap is real and it grew — view-dependency climbed ~40% relative (panel ①) and the two scoreboards pulled apart (panel ②). But the move at the account level is from 15% to 21%, not the 6× you saw on individual ads. That extreme lives at the single-ad level — next section — and, importantly, in Northbeam's native view model, which is more aggressive than the Meta proxy I could pull here (see Method).

The per-ad scoreboard (last 4 weeks)

Every scaled Facebook ad (spend > $2k) with its CPO under each model and the gap between them. This is where your $100-vs-$600 lives — in specific creative, not the average. Note the pattern in the high-gap rows: retargeting/MOF ads and static "problem-agitation" creative (coffee jitters, acidity, gut). Those are demand-harvesters — they get seen by warm audiences and credited on views.

"Gap" = clicks-only CPO ÷ clicks+views CPO. Amber ≥ 1.4×, red ≥ 2×. One ad (Retail/Target traffic) had 0 click-orders and 1 view-order — an infinite gap — excluded from the gap sort as noise.

The trap you're walking toward

Deterministic views are still views. Even when perfectly matched, a view-through is the least incremental signal you have — it credits an ad for a purchase that often would have happened anyway (loyal intenders, demand you already created). When you make that model the judge of what to scale, you systematically:

• Reward ads that pile up frequency on warm viewers — harvesting — because they accumulate view-credit cheaply.
• Kill or starve cold-prospecting ads that clicks-only would show are still doing real acquisition work.

There's a quieter cost too: switching the model mid-May broke your own time-series. The team is comparing May-after to April-before on two different rulers — so the account's self-reported "recovery" can't be trusted across the break.

Recommendations

• Revert intraday scale/hold/kill to Clicks Only. It's your locked standard and the more incremental, more conservative signal. Use Clicks+Views as a secondary lens — never the deciding vote.
• Make the gap itself a KPI. Track clicks-only-vs-clicks+views per scaled ad. A widening gap = the ad has stopped prospecting and is now harvesting → that's a cut-or-refresh flag, not a scale flag. This turns your discovery into a standing instrument.
• Settle the spend-pullback call with an incrementality test, not an attribution argument. Your CFO wants to cut spend; the honest arbiter of "are these customers incremental to LTV" is a Meta conversion-lift study or a geo holdout. No attribution model answers "would this have happened anyway" — worth doing before a multi-million-dollar plan cut.
• Don't let the model fix become false comfort. Even perfectly reverted, real volume stays down until the audience/creative/CVR problems resolve. This improves decision quality, not demand. Keep it out of the "we found the bug" column for the board.

How this was built — and what it can't see

• What I could pull: Northbeam's API (this connection) is locked to the "Clicks only" model and won't toggle to "Clicks + Deterministic Views." So I used the closest available equivalent: Meta's own pixel data, split 7-day-click vs 7-day-click+1-day-view — the same click-vs-view comparison, fully reproducible.
• The big caveat: Meta's view window here is 1 day. Northbeam's "deterministic views" use a longer, cross-device window — so Northbeam's native gap is larger than this proxy. In other words, the true effect in your Northbeam UI is bigger than what's charted here. The direction is confirmed; treat the magnitude as a floor.
• Granularity: weekly buckets built from per-week Northbeam exports (the export collapses a date range to totals; weekly granularity in one call times out). Each bucket is Mon–Sun, aggregated across all Facebook Ads campaigns.
• Ground-truth CAC = Northbeam clicks-only first-time CAC for Facebook Ads; runs higher than blended CAC (which is spend ÷ Shopify new customers) but the trajectory is what matters.

Next step if you want the exact Northbeam native deltas: a 2-click export from the Northbeam UI with both models selected would let me chart the real gap per ad over time. Say the word.