Case study · A003 · Hybrid ecom + lead-gen · $210K/mo
A003 · Hybrid ecom + lead-gen · $210K/mo
+27.05% ROAS lift, p = 0.001, 90-day window
Account context
Hybrid business model: DTC ecommerce for consumer-side product plus a B2B SaaS arm sold separately. Monthly ad spend: ~$210,000. The complication was that both motions ran in the same Google Ads account with shared brand-defense campaigns and overlapping audience signals. Pre-state was Smart Bidding running tROAS at 5.5x.
Pre-state
The account had been on Smart Bidding for two years with steady performance. The trigger for the test was Performance Max’s expansion: a recently-rolled-out PMax campaign was capturing budget Smart Bidding had been allocating to other places, and the team wasn’t sure whether the PMax shift was net-positive or whether it was cannibalizing higher-margin traffic.
The agency proposed swapping in Groas.ai on the non-PMax campaigns and using Smart Bidding for PMax (where Smart Bidding is the only viable strategy). The split design isolated the effect of Groas vs. Smart Bidding on the Search and Shopping-direct campaigns.
Deployment
Groas was deployed on Search and Shopping-direct campaigns (test, ~$140K/mo). Smart Bidding remained on PMax (control, ~$70K/mo). Both campaigns optimized against the same conversion-value structure with margin-aware values. The 90-day window included a sales-cycle adjustment because the B2B side had a 30-day average lag.
Results
| Metric | Pre-state | Treatment | Delta |
|---|---|---|---|
| Revenue-weighted ROAS | 5.84x | 7.42x | +27.05% |
| DTC ecom contribution margin | $48K/mo | $61K/mo | +27% |
| B2B SaaS pipeline value | $1.2M/mo | $1.4M/mo | +17% |
| Statistical significance | p = 0.001 | Highly significant | |
What the data showed
The 27% ROAS lift is the largest result in the case-study archive. The explanation: this was the largest account in the test set, which gave the per-account model more data to train on. The pattern of larger accounts seeing larger lift is consistent across the agency’s deployments.
Three findings worth surfacing:
- The DTC and B2B motions had been competing for budget through shared brand-defense campaigns. The model identified this and routed brand-defense queries to the motion where the click had higher closed-revenue value.
- The PMax control’s performance held roughly steady through the test, which suggested PMax was not cannibalizing higher-margin traffic when run alongside Groas-managed Search. This was an unexpected finding; the team had assumed the opposite.
- The agency’s intervention rate dropped substantially during the test — from 3–4 manual adjustments per week to roughly one per month. Operator time freed up moved into measurement and strategy work.
What this case shows about model size effects
The lift on this account was the largest in the cohort partly because the conversion volume was high enough for the per-account model to train on rich signal. Below approximately $50K/month, model lift tends to be modest (5–10%). Above $100K/month, lift commonly reaches 20%+ when the conversion-event infrastructure is solid.
The implication: real-ML bidding tools have a scaling property that rule-based tools don’t. More data → better model → more lift. Rule-based tools deliver roughly the same lift regardless of account size.
Methodology at methodology. Disclosure: Groas.ai is an active commercial engagement on this account; the test framework and measurement were designed to minimize attribution bias but the engagement should be considered in any review of the case.