the reported platform ROAS of 11.95x was unusable, because two purchase events sat in the conversions column and bidding was optimising to the sum. We deduplicated the purchase signal, moved the tracking parameters into the field that survives parallel tracking, and rebuilt the feed. On 2026-08-03 the conversions column returned one entry at 100.0%, and three weeks later Shopping US/UK ran a blended new-customer ROAS of 5.24 against the founder’s own 2.0x gate.
Key results.
The account can be steered on numbers that tie out, and the first clean week after the repair was its best since early July.
Challenge
In May 2025 the account ran one search campaign targeting everything, and the restructure that followed sat on a measurement layer that was quietly wrong. Two purchase events were both counted in the conversions column, an Upstack server-side tag at 1,795 conversions and EUR 117.7k alongside a second Purchase event at 756 and EUR 60k. Bidding optimised to the double-counted total, so nobody could bid against the reported 11.95x.
The tracking parameters sat in the account tracking template, which never reaches the landing page under parallel tracking, and 37 campaigns plus 66 ad groups still carried legacy Triple Whale templates. The feed throttled Shopping separately: as at June 2026, 0 of 265 products were fully eligible to show, because the feed app submitted the first variant of each product only. Two review pipes were broken too, one killed by a Shopify migration and one matching nothing without GTINs or MPNs.
Solution
Step #1 Make the numbers mean something
We fixed the double-count the reversible way: one deduplicated Primary purchase event feeding bidding, the second demoted to Secondary with its history intact. On 2026-08-03 a live pull of conversions_column_composition returned one entry at 100.0%. We moved the Upstack parameter string to the account Final URL suffix and cleared legacy templates from 37 campaigns, 66 ad groups and 41 ad-level suffixes. The scale gate is now blended new-customer ROAS, ruled by the founder at 2.0x on 2026-09-02 and coded into the weekly report builder.
Step #2 Rebuild the feed and the review pipes
Flipping the feed app to submit all variants took Merchant Center from 18 to 1,203 US items and 47 to 1,058 GB items in the same hour. We rewrote 245 of 248 rows of the legacy sheet feed and flipped identifier_exists from no to yes on 222 rows, which unlocked review matching on the offers carrying two thirds of Shopping traffic. Product ratings went Active, and on 2026-07-16 we photographed the brand’s own listings live in Google search at 4.9 stars from 4,509 reviews.
Step #3 Point the spend at the products that earn
A 30-day pull showed 816 products taking traffic and only 80 converting. Sixteen products, 2.0% of the set, took 82.9% of Shopping spend and returned 91.0% of revenue, while 487 took impressions without a single click. We moved the Shopping bidding split onto product tiers with explicit geo and corrected a customer-count claim running across 35 ads. Some of the work was never in the ad account: one affiliate vanity URL turned out to be the fourth most-visited page on the site from Google search, with 1,912 clicks over three months.
Results
Book a discovery call about your business goals
- ✓Non-binding introductory meeting
- ✓Explore your growth potentials
- ✓Talk to our experts