Case: the 5,000-SKU ecommerce site that moved from a single ROAS to managing by profit

The second typical case: an ecommerce site with thousands of SKUs, one "everything" PMax at target ROAS 4 that "works" (actual ROAS 4.3) and a profit that refuses to grow. The rebuild, in steps, towards an architecture that manages by margin rather than by revenue.

Starting point

• 1 PMax with the whole catalogue (5,000 products), target ROAS 4, €25,000 a month; 1 residual standard Shopping campaign (drained by the PMax); a brand Search campaign; 2 countries with one EUR feed for both.
• Conversion value = revenue including VAT; no returns deducted.
• Margins: from 12% (electronics) to 55% (own-brand textiles).

Step 1 · Economic truth (month 1)

• Values excluding VAT; adjustments for returns (by order ID).
• Margin table by category (from finance) → custom_label_0.
• POAS calculated per product from real sales (Shopping Ninja): the finding was that 28% of the spend was going to products with POAS < 1 (low margin, pushed by the PMax because they generated revenue).
• Brand excluded in PMax: PMax's "real" ROAS dropped to 3.4 once it lost the brand searches.

Step 2 · Structure by margin (months 2-3)

SEARCH  · Brand                                90% share
PMAX    · Own-brand textiles (high margin)     target ROAS 2.8  (listing group by label)
PMAX    · Home (medium margin)                 target ROAS 4.5
SHOP STD· Electronics (low margin)             target ROAS 9    (fine control; branches excluded)
SHOP STD· New products (< 30 days, no data)    maximise clicks → loose ROAS after a month
DG      · Dynamic catalogue remarketing        its own target CPA
Disjoint listing groups: each product in ONE campaign (PMax always wins)

Why: high margin deserves reach (PMax); low margin deserves control and a ROAS that only "self-serve" sales can meet (standard Shopping with branches excluded); new products get their grace period. Migration by label and by week (one family a week), with their own budgets and two weeks of observation.

Step 3 · Dynamic labels and the matrix (months 4-6)

custom_label_1 = performance (star / profitable / neutral / loss / no data) recalculated every night from POAS with hysteresis; the margin × performance matrix moves products between campaigns and targets (module 7 of this level). Effects: 310 loss-making products excluded (via checkbox) in the first month; 40 medium-margin "star" products moved to a loose target and grew sales by 35%.

Step 4 · Country feeds and titles (months 6-9)

• A dedicated FR feed (language, EUR with French VAT, FR shipping).
• Titles by template and by each market's vocabulary (the AI lesson in module 7): +45% impressions in the pilot category, with CTR stable.
• Merchant: seller ratings switched on; promotions during the sales.

Step 5 · Prospecting Demand Gen and video (months 9-12)

With buyer lists as the seed, narrow lookalikes and the product feed in Demand Gen: catalogue prospecting at its own CPA, judged on new customers and incrementality (a geo experiment across two regions). It contributed 9% incremental sales at ROAS 3.1 — acceptable at high margin, not at medium: it was confined to textiles.

Before and after (12 months, the same €25,000 spend)

ROAS fell and profit rose by 50%: proof that ROAS was the wrong metric.

Mistakes that were avoided

• Moving the whole catalogue in one go (it was done family by family).
• Keeping PMax and standard Shopping on the same products.
• Judging Demand Gen on view-through.
• One feed for two countries.

💡 Ninja trick: Shopping Ninja was the tool for steps 1 and 3 (POAS per product, traffic light, exclusions via checkbox, the performance label exported as a supplemental feed) and Ninja Shield kept the PMax and Demand Gen placements clean throughout. The architecture was designed by a person; it is maintained every night by a script.

What you should remember

• Economic truth first: values excluding VAT, returns, margins, POAS.
• Structure by margin: PMax for high, standard for low, new products with grace, dynamic remarketing; disjoint listing groups.
• Dynamic labels + the matrix move products on their own; country feeds and titles per market.
• ROAS fell and profit rose: the metric was the problem.