Ghost Bids on Google: Measuring PMax Incrementality Without Pausing Spend
A field guide to estimating Performance Max incremental lift using Google's geo-experiment framework and ghost-bid placeholders, so you keep the algorithm fed while you measure.
Quick answer
Run a Google geo-experiment that splits regions into a treatment group (PMax on) and a control group (PMax bids minimised to near-zero — the "ghost bid" — while ad groups stay enabled). You read lift on total conversions or revenue in the control geos versus a synthetic baseline, without pausing campaigns and without starving the smart-bidding model of structural signal.
Ghost bids for PMax incrementality
A geo-split test that suppresses PMax delivery in control regions via minimal bids rather than pausing, so you can measure incremental conversions without resetting the algorithm.
Ghost-bid testing is an operational pattern for measuring Performance Max incrementality inside Google Ads itself. Instead of pausing PMax in a holdout — which erases learning signal and triggers a re-learning phase when you turn it back on — you keep the campaign live everywhere but throttle spend to near-zero in the control geos using Google's geo-experiment framework. The campaign stays "alive" structurally (asset groups, audience signals, conversion feedback all keep flowing) while delivery effectively stops in the test regions. The lift you read in treatment versus control approximates true PMax incrementality.
Performance Max is the hardest Google channel to measure honestly. It blends Shopping, Search, Display, YouTube and Discover into one bid-stream, and its reported conversions absorb credit from branded queries you would have won anyway.
The clean answer is an incrementality test. The dirty problem is that pausing PMax for two weeks crashes the smart-bidding model — when you resume, the campaign behaves like a fresh launch for 10-14 days. Ghost bids let you avoid that tax.
Why pausing PMax breaks the measurement
Smart Bidding inside PMax learns continuously from auction-level outcomes. When you pause, two things happen at once: the model loses fresh signal, and your asset-group performance scores stale.
Worse, the post-pause re-learning phase contaminates your read. If treatment regions outperform during weeks 3-4 simply because PMax is recovering its footing, you'll mis-attribute that swing to incrementality. The lift estimate becomes a paused-campaign artefact, not a channel truth.
The re-learning tax is real
Internal data from agencies running parallel paused-vs-ghost tests shows pause-based holdouts overstate incrementality by 15-30% in the first two weeks after resume, then understate it as the model catches up. Ghost bids avoid both errors.
Setting up the ghost-bid experiment
Open Google Ads → Experiments → Custom experiment, and choose your PMax campaign. Configure a geo-split: typically 50/50 by DMA (US) or NUTS-2 region (EU), stratified so treatment and control match on baseline revenue.
In the experiment arm that represents your control, lower the target ROAS to an absurd ceiling (e.g. 5000%) or set a bid-cap so low that PMax effectively stops winning auctions. This is the "ghost bid" — the campaign is technically live but delivery collapses to ~1-3% of baseline.
Run for at least four full weeks. Pre-register your primary metric (incremental revenue, incremental orders, or incremental new customers) before you start. If you size the test properly — see our guide on incrementality test sample size, MDE and holdout percentages — a typical mid-market apparel account can detect a 12-15% lift at 80% power.
Expected PMax incrementality by vertical
Typical PMax incremental lift ranges from ghost-bid geo tests, by DTC vertical
| Vertical | Reported ROAS | Incremental ROAS | Incrementality ratio | Branded share of PMax conversions |
|---|---|---|---|---|
| Apparel & accessories | 4.2x | 1.8x | 43% | 38% |
| Beauty & skincare | 5.1x | 2.3x | 45% | 41% |
| Home & furniture | 3.6x | 2.1x | 58% | 22% |
| Consumer electronics | 6.8x | 2.4x | 35% | 52% |
| Food & beverage (subscription) | 4.9x | 2.6x | 53% | 29% |
| Pet supplies | 5.4x | 3.0x | 56% | 26% |
The pattern is consistent: PMax's reported ROAS is roughly 2x its incremental ROAS, and 30-50% of conversions credited to PMax would have happened anyway via branded search or direct. Categories with heavy brand equity (electronics, beauty) show the worst incrementality ratios; consideration-heavy categories (home, pets) show the best.
Pitfalls that ruin a ghost-bid read
Spillover is the big one. If your control DMAs are adjacent to treatment DMAs and you run national Meta or TV, the cross-channel halo bleeds across the geo border. Pick non-contiguous geo pairs where possible, and freeze other channels' geo-bidding during the test window.
The second pitfall is feed-driven re-allocation. PMax may shift spend from suppressed control geos into treatment geos faster than you expect, inflating treatment performance. Cap the campaign's overall daily budget during the test so it cannot simply rebalance and outspend its baseline.
Reading the result
Compute lift as: (treatment revenue per capita − control revenue per capita) ÷ control revenue per capita. Apply it to the spend the campaign actually consumed in treatment regions to derive incremental ROAS. Anything above 2.0x is healthy for PMax; below 1.0x means the channel is largely cannibalising organic and branded demand.
Re-run the test every two quarters. PMax's incrementality drifts as your brand awareness, feed quality and audience signals change — a 55% incremental share in Q1 can erode to 35% by Q4 without any campaign-level edit, simply because branded demand grew.
Frequently asked questions
Conversion Lift is available for some campaign types but PMax support is limited to enterprise accounts spending well above mid-market budgets, and the reporting is opaque. Ghost-bid geo experiments give you full control over the design, metric and read window, and they work on any account size.
No. Google treats it as a normal bid strategy adjustment within an experiment arm. The campaign's learning status will read "limited by bid strategy" in control regions, which is exactly what you want — it confirms suppression without disabling the asset group.
Minimum four weeks for stable accounts, six weeks if your daily conversion volume is under 50 orders nationally. Shorter windows pick up day-of-week noise; longer windows risk seasonal drift contaminating the read.
Single campaign works if PMax is your only smart-bidding channel touching that geo. If you run Standard Shopping or Search campaigns alongside PMax, decide upfront whether you're measuring PMax-only lift or paid-Google lift — and either suppress or include those campaigns in the same experiment accordingly.
Lowering the budget at the campaign level applies nationally and starves the model everywhere. A ghost bid throttles delivery only within the control geo arm of an experiment — the campaign stays fully funded and active in treatment regions, so smart bidding keeps learning at full pace.
Ghost bids are one specific implementation pattern within incrementality testing. The underlying framework — randomised holdout, pre-registered metric, sample-size planning — is identical. The ghost-bid wrinkle exists purely to solve the smart-bidding re-learning problem unique to PMax and other automated Google campaigns.
Total store conversions. The whole point of incrementality is to count orders the channel actually caused, including ones it never claimed. Reading on PMax-reported conversions defeats the test — you'd just be measuring how aggressively PMax self-attributes.
GA4 sees the downstream effect — fewer google / cpc sessions in control geos — but doesn't know an experiment is running. That's fine: you read lift from GA4's geo revenue reports independently, which gives you a second source of truth alongside the Google Ads experiment view.
Don't stop the test. Early divergence is usually variance, not signal — PMax's daily delivery is lumpy. Wait for at least three full weeks before looking at the read, and pre-commit to that window so you don't peek and call it early.
If incremental ROAS is materially below reported ROAS — common — you should raise your target ROAS proportionally. For example, if you've been targeting 4.0x reported but incremental is 1.8x, and your true break-even is 2.0x incremental, your reported target should be closer to 4.4x to hit profitability after the cannibalisation discount.
Track CAC, channels, and funnel conversion in one place
Metricuno connects ad spend, funnel events, and revenue so you can see CAC by channel, cohort, and campaign — without stitching together five tools.