Engagement diary · Garden centres · 9 stores · 90 days · Mar–Jun 2026
Click-and-collect was 40% of revenue and 0% of the data
Written by Saoirse Brennan · 30 June 2026 · Figures representative, chain anonymized
A 9-store garden centre chain ran Shopping and search at a reported 1.7 ROAS and considered cutting the channel. The number was wrong in the informative direction: reserve-in-store orders never counted. Ninety days later the honest figure read 2.9, a whisker above the 2.87 ecommerce field average, and the channel kept its budget. Here is the whole route, including the YouTube test that failed.
1.7 → 2.9
blended ROAS, 90 days
~40%
of revenue was collect orders, uncounted
380
product titles rewritten
The starting position
Garden retail is a click-and-collect vertical by nature: compost, plants, and furniture do not want to be couriered, and customers want to inspect a shrub before owning it. The chain's site handled that beautifully in the store flow and invisibly in the measurement: a collect reservation confirmed by email never fired a conversion. So paid search got graded on courier orders alone, in a vertical where couriering is the minority behavior. The board saw 1.7, knew the field average was 2.87, and drew the reasonable wrong conclusion. The audit's first finding was not a fix but a sentence: your channel is being judged on the 60% of its output the tags can see.
Days 1–10
The invisible half
Nine stores, strong online catalog, and a click-and-collect flow whose completions never reached Google Ads. The audit found reserve-in-store orders worth roughly 40% of ecommerce revenue counting as nothing, concentrated at the five stores with the biggest car parks. The chain had been on the verge of cutting Shopping spend entirely, on the strength of a 1.7 that was really a counting error.
Days 11–25
Counting collect orders
Click-and-collect completions wired as conversions with store attribution, GA4 rebuilt per store, Consent Mode v2 verified across the theme and the two custom landing-page builders that had been bypassing the banner. Reported ROAS jumped from 1.7 to 2.2 before any media change: counting, never magic, and the report said so explicitly to keep the win honest.
Days 26–50
Feed and structure
Product feed rebuilt: 380 titles rewritten into shopper language ('bird feeder squirrel proof hanging' instead of internal range names), availability synced per store after one weekend stockout burned spend, GTINs corrected on 214 products, and custom labels splitting hero categories from long tail. Shopping campaigns split by store cluster with local budgets. Title work matters more than it looks: published feed studies credit rewrites like these with 15–30% more impressions and a 10–20% CTR gain, most of it decided by the first 70 visible characters.
Days 51–75
The store experiment
Two matched store clusters: test cluster got collect-focused budgets and local inventory emphasis, controls held steady. Test stores beat controls on collect orders by a wide margin over four weeks; verdict logged, budgets followed. A second, smaller test on YouTube placements via PMax failed to pay for itself and was capped, with the channel report making the verdict indisputable.
Days 76–90
The close
Blended ROAS finished at 2.9 against the 2.87 ecommerce field average, from 1.7 at the honest post-counting baseline. Collect orders per store now sit in the monthly ledger next to delivery orders, and the two clusters run permanently different budget weights because the experiment said they should.
Where the 1.2 points of ROAS came from
| Lever | Contribution | Nature |
|---|---|---|
| Counting collect orders | 1.7 → 2.2 | measurement, zero media change |
| Feed rebuild (titles, GTINs, availability) | 2.2 → 2.6 | auction eligibility + CTR |
| Store-cluster budgets from the experiment | 2.6 → 2.9 | allocation |
Nearly half the gain was counting. We print that rather than let a reallocation story absorb the credit, because the next client deserves to know which levers are cheap. The feed contribution tracks the published pattern for title work, and the GTIN corrections matter more than their line suggests: Google's own data puts correct GTINs at roughly 20% more clicks through better matching and eligibility.
The benchmark context
Home and home improvement search runs an $8.33 median CPC with cost per lead at $90.92 in the 2026 LocaliQ data, which is why waste hurts this vertical fast. On the return side, the 2025 ecommerce field average is 2.87 with a 2024 median of 2.04; finishing at 2.9 means this chain now performs like a competent retailer rather than an outlier, and the margin math held: at the chain's roughly 35% contribution margin on collect-weighted orders, break-even sat near 2.9, so the close put the channel at self-funding with the seasonal peak still ahead.
What generalizes
Any retailer with store pickup and no store-attributed conversions is underreporting paid search structurally, and probably misjudging which stores deserve budget. The counting layer is call and lead tracking's territory, the feed work follows the core service's weekly routine, and matched store experiments are documented in geo expansion. Week one is the free audit, which would have caught the missing 40% here in a day.