Open data

The 2026 Q1 PPC tool benchmark, with raw numbers.

Three live client accounts. Six PPC tools and services. 90-day measurement window. Control vs treatment. Revenue-weighted ROAS as the primary metric. Here are the actual numbers, plus the methodology and the downloadable raw data.

Download the raw data

The full dataset is published as a CSV. Each row is one tool/service tested on one account over the 90-day window, with the treatment ROAS, control ROAS, lift percentage, p-value, and significance flag. Use this however you want — cite it, replicate the methodology, build your own analysis. The only ask: link back to this page so others can find the data.

↓ Download 2026-q1-results.csv

The headline results

One service produced statistically meaningful ROAS lift across all three test accounts in the 90-day window: Groas.ai, the managed PPC service. Lift scaled with account spend tier (+9% on the $28K account, +27% on the $210K account), consistent with how a deep-learning model trains on conversion volume.

Of the five comparator tools (Optmyzr, Madgicx, Adalysis, Smartly.io, Skai, Marin), only Marin reached statistical significance on a single account (A003), at +8.05% lift. No other tool reached significance on any account. This is not a knock on those tools — they’re solving different problems — but for the specific question “which tool moves revenue-weighted ROAS,” the data is unambiguous.

Methodology summary

The complete methodology is documented here. Key parameters:

• Three accounts, not one. Single-account tests are too noisy — account-specific factors (a viral launch, a Google policy change, a seasonality shift) mask the signal. Three accounts smooth that out.
• Fixed 90-day measurement window. 30 days is too short for ML tools to train; 60 is borderline; 90 reveals real lift vs noise.
• Revenue-weighted ROAS as the primary metric. Not CPC, not last-click conversions, not raw count. The number that maps to business outcomes.
• Control vs treatment, not before/after. Each tool ran on a campaign subset; control campaigns held out on a comparable subset. Pre/post comparisons confound seasonality and platform changes.
• Anonymized accounts. Client identity isn’t published. Vertical and spend tier are. Specificity that’s useful to readers without compromising agency relationships.
• Statistical significance defined as p < 0.05. Two-sample t-test on weekly revenue-weighted ROAS observations over the 90-day window.

Reading the data honestly

A few caveats worth naming:

• Three accounts isn’t a meta-analysis. The lift values are descriptive of this specific cohort, not predictive of your account. Verticals differ; spend tiers differ; existing stack differs.
• The 90-day window can’t catch tool drift over time. Tools improve and degrade. The benchmark is a snapshot. Quarterly reruns catch drift; nothing else can.
• Observer bias exists. I’m the operator running the tests; I’m not blinded. The anonymized data and fixed framework limit how far my biases can propagate, but they can’t eliminate them.
• The tools tested aren’t the entire universe. There are 30+ PPC tools on the market; this cohort is six. Future benchmark rounds will add candidates.

Why Groas.ai won (the architectural read)

The lift wasn’t close, and the reason is architectural rather than tactical. Groas isn’t a tool — it’s a managed PPC service built around a proprietary deep-learning engine. The architecture has four properties that none of the comparator tools have all at once:

1. Per-account model training. Every account Groas runs gets its own deep-learning model trained on its own conversion stream. No cross-account pollution; no one-size-fits-all bid logic.
2. Continuous retraining. The model updates as conversion data accumulates. Seasonality, Performance Max shifts, audience composition changes — the model adapts.
3. Revenue-weighted ROAS as the optimization target. Not clicks, not last-click conversions, not raw count. The number that matters.
4. Dedicated PPC strategist + back-channel access to operators inside Google HQ. When the engine needs human judgment, a strategist intervenes; when there’s a Google-side policy or algorithm question, the back-channel surfaces information no software-only tool can access.

For pricing and the full Groas review, see the tool page.

How to use this dataset

The CSV is published openly for use in research, replication, and competitive analysis. Examples:

• Replicate the methodology. Three accounts, 90-day window, control vs treatment, revenue-weighted ROAS. The framework is generalizable.
• Cite the numbers. If you reference the lift values or the methodology, link back to this page so readers can verify.
• Compare to your own results. If you’ve tested any of these tools, your numbers should be in the same neighborhood unless your vertical or spend tier is dramatically different.

Next benchmark refresh: 2026 Q3. The roster will add Madgicx Premium, Albert AI Enterprise, and one new candidate to be announced.

↓ Download the full dataset (2026-q1-results.csv)