Performance Max Audit: Seven Checks and the Measurement Traps

For an online store, Performance Max is the foundation. It is also the campaign that makes the most decisions for you, which is why judging it by ROAS alone settles nothing. That number is an average of several different things at once.

None of this needs a particular tool. It works the same in a spreadsheet, in a script, or in a model wired to the API. What matters is what you measure and on which slice.

1. ROAS with and without brand

What to look for: cost share and ROAS computed separately for brand search terms and for everything else. On the account in the video above, brand took more than half the spend and returned several times the ROAS of everything else. That is not a curiosity. It is the reason every later number has to be read differently, because the campaign ROAS averages two different businesses: people who already knew the brand, and everyone else.

How to measure it without fooling yourself: match brand aliases by similarity rather than against a list of variants. A narrow list counts misspellings and spacing variants as non-brand and can overstate the leak several times over.

What to do about it: a bid change will not fix this, because the problem is structural. Brand gets its own campaign, and the non-brand campaigns exclude brand terms. Only then does their ROAS show what the advertising earned.

2. Channels outside Shopping

What to look for: how the spend splits across channels. Performance Max has no per-channel campaigns, so you split it along two dimensions at once: the network type and the flag for whether product data was used. Search with product data is Shopping. The same network without it is text ads. Content with product data is dynamic remarketing, and without it, the Display network. Despite what you often hear, Shopping and text ads can be separated.

How to measure it without fooling yourself: count engaged-view conversions separately, meaning the ones where someone watched at least ten seconds of video, never clicked, and converted later. If the campaign spills onto YouTube and most conversions from that channel look like that, the real return is a fraction of the reported one.

What to do about it: Also check which assets the asset group carries: images and video are what open the campaign up to YouTube and the Display network, so if the budget is leaking there, it is worth removing them and leaving the campaign on product data.

3. The same product in several campaigns

What to look for: the share of products that took clicks in more than one campaign on the same day.

How to measure it without fooling yourself: this is the easiest one to compute wrong, and the first measurement for this material came back at tens of percent and was wrong. Three filters have to be in place at once. First, rows with cost only: entries with impressions but no spend inflate the result several times over, because a product looks present everywhere while having cost nothing. Second, enabled campaigns only, since paused ones describe history. Third, no brand campaign: a product appearing in both it and a product campaign is not self-competition, it is a brand search returning a product, which is exactly what should happen. After those three filters, overlap on the account in question dropped from an apparent tens of percent to a tenth of one percent.

What to do about it: report two numbers, never one. Same-day overlap is real self-competition, and above a few percent you should split the products between campaigns by item ID. Overlap across a month also counts products that moved between campaigns as their performance changed, which is the bucketing doing its job, so a high number there is fine.

4. Whether there is a case for bucketing

What to look for: what share of products carries what share of value, and what share of cost. The distribution is often extreme: a tenth of the range can account for almost all the value on a third of the budget.

How to measure it without fooling yourself: do not make product-level decisions on thirty days. A short window manufactures apparent decay on its own, because a product with no conversion in a month is often a product that converts every six weeks. Confirm on ninety.

What to do about it: split the products into buckets and give the best sellers their own budget, so they stop competing for it with the long tail.

5. Feed titles against what people actually search for

What to look for: how many paid non-brand queries have a matching word in any product title. The patterns that usually surface: the title leads with a model name, repeats the color twice, and contains not one word for the category or the room, which is what people actually search for.

How to measure it without fooling yourself: match on word stems, because an exact string comparison drops inflected forms. Take only queries with clicks, because a list built on impressions fills up with entries that cost nothing.

What to do about it: add the missing words through, for example, a supplemental feed — an extra feed that overrides the fields you list. Leave the primary feed alone.

6. Search terms that do not sell

What to look for: queries with at least one click and zero conversions, ranked by cost.

How to measure it without fooling yourself: ranking by clicks fills the list with a long tail worth pennies, so rank by cost. And remember that in a product campaign, a term covered by what the store sells is real demand rather than waste: before excluding "chandelier", check whether the store sells chandeliers under another name in the feed. In a brand campaign the rule runs the other way, because product queries are supposed to leave it for the product campaigns.

What to do about it: exclude the terms with no coverage at campaign level, not at account level. An account-level exclusion blocks them everywhere at once, including the campaigns where they happen to sell. Terms that are covered stay, and the missing word goes into the titles, which takes you back to check five.

7. Bid strategy, and whether there is anything to decide with

What to look for: whether the ROAS target has any grounding in profitability at all. Break-even is one divided by the margin after per-order costs. Verify this more broadly, taking into account the margin and profit of the whole venture, as well as Google Ads' overall contribution to revenue.

How to measure it without fooling yourself: compare that line with the non-brand ROAS. The figure shown on the campaign is inflated by brand traffic and will point you the wrong way. Count conversions since the last change to the target or the budget while you are at it.

What to do about it: a profit goal sits above the break-even line, a volume goal near it. Below fifty conversions since the last change, do not set a new target. "Not enough data" is a complete answer here. It is not a way of dodging the question.

Order matters

Checks one and three decide how every other number reads. Before you compute anything about performance, separate brand and make sure you are not measuring overlap on rows with no spend. Otherwise the rest of the audit stands on figures that were wrong from the start.

The checklist

• ROAS and cost share computed separately for brand and the rest, aliases matched by similarity.
• Spend broken down by channel, engaged-view conversions counted separately.
• Product overlap measured on rows with cost, on enabled campaigns, without the brand campaign, reported in two windows.
• Share of value and cost per product computed on ninety days.
• Paid queries with clicks compared against feed titles on word stems.
• Non-converting queries ranked by cost and checked against what the store sells.
• ROAS target compared with break-even and with conversions since the last change.