Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › Intermediate

Setting Guardrails for Automated Campaigns at a Garden Tools Retailer

Practise steering automated campaigns through the few levers the platforms leave you. You read an invented garden tools retailer's campaign and shop data, confirm that its spend is leaking to existing customers, then set the conversion signals, exclusions, campaign split, asset groups and weekly checks that point delivery back at new customers and profitable products.
ExerciseIntermediate levelModule 1: AI in ad platforms5 minUpdated

Why this matters

The guide to AI in ad platforms makes the central point of this module: the platforms optimise exactly what they are told to, and the advertiser controls the frame: what counts and what it is worth, what the campaign may not do, what it may use and how it is structured. When automated campaigns drift towards existing customers or thin margins, the fix is one of those frame decisions, not a setting inside the automation.

The retailer and its automated campaigns

Spadewell Garden Tools is an invented online retailer of hand tools, watering kit and cordless power tools. It moved to one Performance Max campaign and a Meta Advantage+ sales campaign. Both report a strong return on ad spend (ROAS), yet new customer numbers in the shop have not moved.

Illustrative example — not a real result

Spadewell: one month of automated campaigns against the shop's own data

CampaignSpendPlatform-reported ROASOrders the platform claimsShare of those orders from returning customers (shop data)
Performance Max, all products£9,000500%60064%
Meta Advantage+ sales£4,000420%28060%
Product bandGross marginShare of Performance Max spend
Hand tools46%30%
Watering and irrigation38%25%
Cordless power tools18%45%

Current set-up:

  • Performance Max goals: purchases valued at revenue, plus newsletter sign-ups and add-to-basket.
  • One asset group for every product; its audience signal is a list of all past buyers.
  • No brand exclusion; final URL expansion on, with clicks landing on the blog and returns page.
  • Customer list uploaded 14 months ago, matching about 8,000 of 22,000 customers.
  • Meta receives purchases from the pixel only; the shop's Conversions API integration is off.
  • Meta suggestions: a lookalike of all buyers and the newsletter list; no excluded audiences.
  • Finance can afford up to £30 of ad cost per new customer.

Choosing inputs and exclusions

Work through the levers in order. Signals first, because every other lever is downstream of them. Then the hard limits: exclusions, brand and URL controls, and the customer acquisition goal, which the Performance Max structure guide describes as the only documented way to shape spend on existing customers. Then structure, split by target rather than by taxonomy. Audience signals and Meta suggestions are hints, not targeting. Returning customers get their own Remarketing campaign and budget, so structure, not delivery, decides who sees which ad.

Task

  1. Confirm the leak from both sides: what the shop's new versus returning split and the Meta audience segment breakdown would need to show.
  2. Work out the cost per new customer for each campaign and compare it with what finance can afford.
  3. Rewrite the conversion signals for both platforms, including how Meta should receive purchases.
  4. Set the exclusions and new-customer controls on each platform, and say what happens to buyer-based signals and suggestions.
  5. Redesign the Performance Max structure: how many campaigns, which targets and which asset groups.
  6. Write a weekly and a monthly check routine with the report each check reads.

Worked answer

Confirm the leak. The shop data already shows most claimed orders coming from returning customers. The guide to Advantage+ audience reaching existing customers asks for a second source: read the audience segment breakdown over several weeks and check it agrees with the shop's split. If it labels many buyers as new, the customer definition is stale.

Illustrative example — not a real result

Spadewell: cost per new customer and break-even returns

CampaignNew-customer ordersCost per new customerAgainst the £30 ceiling
Performance Max600 × 36% = 216£9,000 ÷ 216 ≈ £41.67Over by about £11.67
Meta Advantage+ sales280 × 40% = 112£4,000 ÷ 112 ≈ £35.71Over by about £5.71
Product bandBreak-even ROAS (1 ÷ margin)Proposed target ROAS
Hand tools1 ÷ 0.46 ≈ 217%300% (shared campaign)
Watering and irrigation1 ÷ 0.38 ≈ 263%300% (shared campaign)
Cordless power tools1 ÷ 0.18 ≈ 556%600% (own campaign)

The single 500% target sits below break-even for power tools, which take 45% of the spend.

Signals. Sign-ups and add-to-basket come out of the campaign goals, because a campaign told that a sign-up counts will find sign-ups. Purchases keep revenue values, with margin handled by the targets above. On Meta, switch on the Conversions API integration beside the pixel and confirm the shared event ID deduplicates them, as the Meta Conversions API guide explains.

Exclusions and new-customer controls. Refresh the customer list with every identifier Spadewell lawfully holds, on a schedule. In Performance Max, set the customer acquisition goal to bid only for new customers, which depends on that list; replace the all-buyers signal with a custom segment of people researching garden tools; add a brand exclusion and leave brand to an exact-match Search campaign; exclude blog, returns and account pages from URL expansion or switch it off. On Meta, add the customer list and a purchase audience as excluded custom audiences, which Meta documents as honoured controls, use a new-customer setting where offered, and remove buyer-based suggestions.

Structure. Two Performance Max campaigns split by a margin label in the feed, so each product sits in one place. Inside the first, separate asset groups for hand tools and watering, because their images, headlines and landing pages differ. A small remarketing campaign on each platform carries returning customers.

Checks. Weekly: the shop's new versus returning split for paid orders, cost per new customer against the ceiling, search terms insights for brand leakage and landing pages used. Monthly: the Meta segment breakdown against the shop, products with no impressions, asset ratings, and the exclusion list's size and refresh date. Change targets in measured steps so learning is not restarted every week.

Common mistakes

  • Fixing the automation instead of the frame. The campaign found what it was told to find.
  • Customer list in the signal or the suggestions. That points delivery at buyers rather than away from them.
  • Pixel and server events without a shared event ID. Purchases count twice and the reported return rises.
  • Judging success on platform ROAS. The test is cost per new customer from the shop's own data.
  • Splitting by category rather than by target. More campaigns, less data each, no better control.

Next, take the checkpoint on steering AI in ad platforms or return to the intermediate AI in marketing level.

Pages this lesson applies

Terms used on this page

Also related

Related