Fourteen campaigns became two, and the CPA finally sat still

By Caleb Nguyen, co-founder & account operations lead · Published May 14, 2026 · Updated August 2, 2026

A furniture retailer with six showrooms across two neighboring states came to us in March 2026 spending about $38,000 a month on Meta. The account had grown the way most five-year-old accounts grow: a new campaign for every season, promotion, and idea, none ever retired. The result was 14 live campaigns competing against each other in the same auctions.

The numbers, before and after

Days from restructure launch to a stable CPA, by our definition of 14 consecutive days inside a ±15% band: 19. The attribution window never moved, so the two columns are the same yardstick.

What was actually wrong

Nothing about the creative was broken, and the audience settings were ordinary. The problem was structural. With 31 ad sets sharing a $38k budget, most ad sets never collected enough conversions in a week to exit Meta's learning phase, so 38% of spend sat permanently in the platform's calibration mode. Fragmentation also meant the retailer was bidding against itself for the same local buyers.

Meta rebuilt its ad-delivery machinery around exactly this failure. The Andromeda retrieval engine, announced in December 2024 and fully rolled out by October 2025, selects audiences from creative signals at auction time. Accounts restructured for it, in the consolidation pattern of one to three ad sets with 10 to 20 creatives, saw around 8 to 10% gains in published panels before any creative work at all.

What we changed, in order

• Weeks 1–2: the standard audit. Counting rules agreed: showroom-visit leads and online checkouts both count, walk-ins do not, window frozen at 7-day click plus 1-day view.
• Week 3: 14 campaigns consolidated into 2: one Advantage+ sales campaign carrying 80% of budget, one campaign for the two launch-market exclusions. Ad sets cut from 31 to 3.
• Weeks 3–8: creative volume raised from 6 to 16 live ads, with four new concepts entering per month on the weekly calendar. Losers retired on schedule.

The scary part for the client was week one after launch, when CPA briefly rose to $67 while learning reset. Caleb Nguyen's launch checklist plans for that dip in writing, which is the only reason nobody panicked and reversed the change on day four. By day 19 the CPA sat at $44 and stayed inside the band.

What we deliberately left alone

Restraint was part of the design. The offer stayed untouched: same financing promotion, same landing pages, so the before/after isolates structure. The attribution window stayed frozen at the account's existing 7-day click plus 1-day view, because switching windows mid-test is how agencies manufacture improvements. And the two launch markets kept their own campaign with pinned budgets, a business rule the automation had no standing to overrule: the owner had lease commitments there, and lease commitments outrank cost per result.

We also declined to promise the CPM anything. Ecommerce CPMs ran a $15.06 median over the year with 13% annual inflation, and a six-showroom retailer does not move that market. The gains had to come from the share of spend doing useful work, which is exactly where they came from: the learning-phase column went from 38% of budget to 9%.

Where it went next

Week 9 onward is maintenance mode: four fresh creative concepts a month against the fatigue data, budget steps of roughly 20% when the stable band holds, and the weekly two-ledgers report reconciling Meta's claimed revenue against the retailer's point-of-sale system. The ROAS column read 2.1 at day 56 under the frozen window; the number the owner actually watches, showroom-visit leads per week, went from 31 to 55 over the same stretch.

What it means if your account looks like this

Campaign sprawl is the most common structural fault we see, and the cheapest to fix: the furniture account gained a 28% CPA improvement before a single new photo was shot. If your Ads Manager shows double-digit campaign counts and a learning-phase column that never empties, the same consolidation applies. The method is documented in the management service and the audit that starts it is described on the process page.

For the measurement half of this engagement, including why we trusted the before-numbers at all, see Measurement & CAPI.