Hypothesis
If Performance Max runs alongside the existing Search and Shopping campaigns in the treated arm, total account conversion value there will rise beyond the control, because it reaches inventory and audiences the other campaigns do not; if not, it is mostly reclaiming conversions they already earned. The decision behind it is concrete: add or scale Performance Max alongside the existing campaigns, or hold it back.
Google describes Performance MaxA Google Ads campaign type that runs one goal across all Google inventory from shared assets. as a single campaign that can serve across its channels and inventory from one set of assets and one goal. That reach is the source of both the opportunity and the risk. It may find buyers on placements the account never bought before, and it may also bid on the same searches and shoppers the existing campaigns already convert, then report those conversions as its own. Both outcomes produce an attractive campaign report; only one grows the business. The Performance Max structure guide covers how to set the campaign up well; this test decides whether, once set up, it adds anything.
Metrics and guardrails
The primary metric is incremental account conversion value: total conversion value across all campaigns in the treated regions, compared with what the control regions and the pre-period predict it would have been. Reading the account total is the point. If Performance Max takes a conversion from Shopping, the account total does not move, and the test correctly shows no gain. Count value in the business's own order data where possible, because the platform's attribution is the thing under question. Divide the incremental value by the new campaign's spend to get Incremental ROASRevenue caused by advertising, measured against a holdout, divided by the ad spend., and compare it with the account's break-even return using the ROAS calculator. An enquiry-led account reads qualified leads from its CRM in place of conversion value, the measure the guide to Performance Max for lead generation judges the campaign on.
Three guardrails explain the result. The first is brand search conversions claimed by Performance Max: if much of what it reports came from Branded SearchSearch queries that contain a company's or product's brand name. queries, it is harvesting demand the brand campaign already met. The second is new-customer share of conversions, read from the business's own customer records by region: a campaign that adds value mostly from returning customers may be buying back loyalty the business already had. The third is blended cost per conversion across the whole account, which shows whether the extra spend made every conversion dearer even if the total rose.
Design
Randomise at region level, or use the platform's own experiment split where one is offered for the setup. For regions, choose areas that can be targeted separately and carry steady shopping and search volume, match them into groups on total account conversion value over a pre-period, and assign one in each group at random to the treated arm. In treated regions, Performance Max launches on top of the existing campaigns; in control regions, the current campaign set runs unchanged.
| Design element | Choice | Reason |
|---|---|---|
| Unit of randomisation | Region, within matched groups, or the platform's experiment split | Performance Max overlaps every other campaign, so only an account-level comparison is fair |
| Arms | Current campaigns plus Performance Max versus current campaigns alone | One change: the new campaign |
| Primary metric | Total account conversion value per region | Counts gains and reallocated credit together |
| Brand exclusions | Applied as the business intends to run the campaign | Tests the real setup, not a worst case |
| Feed and promotions | Frozen for the test window | Feed changes move every shopping campaign at once |
| Learning period | Run but excluded from the read | Early delivery is not representative |
Other campaigns keep their budgets and targets in both arms. If Search or Shopping spend falls in the treated regions because Performance Max is winning their auctions, that is part of the effect being measured, not a reason to rebalance mid-test. The Performance Max versus Standard Shopping comparison sets out how the two campaign types differ in control and reporting; this test does not choose between them but asks whether adding one grows the total.
Sample size reasoning
The comparison is total account value per region, so what limits the test is the number of regions and how noisy their combined value is, not clicks. Total value is a skewed, lumpy figure: a few large orders can swing a region's week. Decide the smallest account-level lift that would justify the campaign's spend, which the break-even return defines, and check against the incrementality testing guide whether the regions can detect it given how closely they tracked each other in the pre-period.
IncrementalityConversions caused by marketing. at account level is harder to see than a campaign report suggests, because the new campaign's spend is often small next to the account's total value. If the regions cannot detect a lift large enough to pay for the campaign, enlarge the test (more regions, a longer window, a larger budget for the new campaign in treated regions) or accept that the decision rests on judgement.
Duration
Run a matched pre-period, then the campaign's learning period, then at least one full buying cycle of stable delivery, all clear of feed changes, sales and seasonal peaks. A sale that lands mid-test raises every campaign's results and can make a small effect impossible to see; a feed change alters what every shopping campaign can show. Fix the read window before launch.
Reading the results and the decision rule
Read the test once, at the end of the planned window. Compare treated regions' total account value with the control prediction, report the interval, compute incremental return on the new campaign's spend, then check the guardrails. Write the decision rule into the plan before launch:
| Result | Decision |
|---|---|
| Total value rises by enough to clear break-even return on the campaign's spend | Keep or scale Performance Max; repeat the read after a material budget increase |
| Total value rises but below break-even | Hold the campaign at a lower budget or return the spend to the campaigns it overlapped |
| Total value flat, campaign reports many conversions | Treat its conversions as reallocated credit; return the budget to Search and Shopping |
| Value rises but mostly from brand queries or returning customers | Tighten brand exclusions and audience signals, then retest before scaling |
The scaling question that follows a positive result is a different test. The budget increase incrementality test asks what more spend on one campaign adds, and is the natural next step once Performance Max has earned its place. The academy exercise on auditing Shopping and testing Performance Max for an aquarium store practises designing this regional test for a store that planned to double Performance Max on the strength of its own report.
An outdoor gear retailer tests Performance Max by region
Kestrel Ridge Outfitters, an invented online outdoor gear retailer, runs Search and Standard Shopping across sixteen regions. It matches them into eight pairs on twenty weeks of total online revenue and launches Performance Max with brand exclusions in one region of each pair, spending £9,000 over six weeks after learning. The new campaign reports £41,000 of conversion value. Against the control prediction, total account revenue in treated regions rises by about £14,000, with an interval of roughly £3,000 to £25,000, an incremental return near 1.6 against a break-even of 2.5. New-customer share is unchanged. Kestrel Ridge holds Performance Max at a lower budget and returns the rest to Shopping.
Pitfalls
Reading the campaign's own report is the pitfall that the whole design exists to avoid, and it is tempting because the report is detailed and the account total is noisy. Brand search absorbed into the campaign is the second: without brand exclusions, a share of its reported value is brand demand. The third is feed changes mid-test. Product titles, prices or availability updated during the window change every shopping campaign's results, and if the change reaches regions unevenly through stock or delivery areas it biases the read; the product feeds guide explains which feed edits matter.
Spill-over is the regional design's own weakness. People shop across region boundaries and location targeting is imperfect, so some treated-region ads reach control-region shoppers. Prefer regions with clear boundaries and leave buffer areas out where travel between them is heavy.
Common mistakes
Performance Max tests are expensive, so mistakes cost weeks. The experiment library lists the other designs, and the Google Ads Help Centre documents the campaign's settings and reporting.
- Judging the campaign on its own conversions. Read the account total against control.
- Launching without brand exclusions. Brand demand then appears as the campaign's success.
- Rebalancing other budgets mid-test. Overlap with other campaigns is part of the effect.
- Changing the feed during the window. Every shopping campaign moves together.
- Reading inside learning. Early delivery is not how the campaign will run.
- Ignoring returning customers. Value bought back from existing buyers is not growth.
