Performance Max isn't the problem. Your feed is doing the bidding.

Performance Max gets blamed for everything. CPA goes up, blame PMax. ROAS dips, blame PMax. Spend goes to Display, blame PMax.

We manage 200+ ecom accounts and the data is boring: in 9 out of 10 underperforming PMax campaigns, the bottleneck has nothing to do with the campaign type. The feed is bidding on the wrong products, the wrong queries, at the wrong margin. Fix the feed and the campaign starts behaving like the case studies promised.

Median CPA

-23%

Median ROAS

+19%

Time to lift

21 days

Brands shipped

18

What "the algo learns" actually means

When Google says PMax learns, what it learns is the asset combinations and the audience signals that produce conversions for the products in your feed. The algorithm is not learning your business. It is learning your inventory at the moment the feed is read.

If your feed treats your hero product the same as your tail SKU, the algo treats them the same too. It will pour spend into whatever converts in the next 14 days, including the SKU you sell at 4% margin because the title is generic enough to match a high-volume query.

The myth is that Performance Max is a black box. It is not. The black box is the part of your account you control most directly: the data layer feeding it.

What the feed actually decides

The feed decides three things every PMax bidder pretends are out of their control:

1. Which queries you compete on. Title and description determine query eligibility. A bracelet titled "14k Solid Gold Cuban Link Bracelet, Womens, 7 inch" gets shown for 12 distinct query intents. The same bracelet titled "Cuban Bracelet" gets shown for 2.
2. Which audience the algorithm builds. Google clusters product types and uses the cluster to seed the lookalike. A messy product_type tree produces a messy lookalike, and a messy lookalike produces wasted spend.
3. Which products carry the campaign. Custom labels, price tiers, margin bands. Without them, every product gets the same priority. With them, you can let your top 20% of products subsidize the rest, or split them out so the algorithm stops treating winners and losers as one population.

None of this needs Google's permission. It happens in the feed.

The Stage 4 reader problem

If you have read three blog posts about Performance Max, you already know the talking points. Add quality assets. Use audience signals. Set a smart tROAS. The whole industry is at Stage 4 awareness on PMax: aware of the solution, comparing tactics.

The market caught up. None of those tactics are differentiators anymore. Everyone running PMax has decent assets and a tROAS. The next frontier, the one that still produces 30 to 50% efficiency lift, is the data layer underneath.

This is the play that splits operators from button-pushers. Button-pushers tweak bid strategies. Operators rewrite the feed.

What we do in the first 14 days

When we onboard a brand and the diagnosis comes back "PMax inefficient," we do not touch the campaign for 14 days. We rebuild the feed.

Step one: title rewrites for the top 100 products by impression share. Generic titles get product-specific. Specs go in. Variants get explicit labels. We do this by querying search terms in Merchant Center, finding the queries the product is already winning, and writing a title that better matches the intent we want to keep.

Step two: custom labels. Three of them, minimum:

• margin_tier (high, mid, low)
• seasonality_band (peak, shoulder, off)
• velocity_30d (new, ramping, mature, dying)

Step three: product_type tree cleanup. We collapse anything more than 4 levels deep. The algorithm clusters poorly when the tree is jagged.

Step four: a price-tier split if the catalog spans more than a 10x range. Sub-100 and over-100 get separate treatment in PMax via custom label segmentation, because the auction dynamics are different.

None of this involves a single bid change.

The result, said plainly

Across the 18 brands we ran this play on in Q1, the median CPA dropped 23% within 21 days. The median ROAS lifted 19%. Spend held steady or went up because the campaigns finally had room to scale into product groups that were previously underweighted.

The honest caveat: 3 of the 18 saw flat results. Two of those had feeds that were already clean. The third had a category problem we could not fix from the feed (the brand was bidding on a saturated commodity). Feed work is not magic. It is just the highest-ROI lever most brands ignore because it is not labeled "Google Ads."

When this fails

This play does not work if any of these are true:

• Your catalog is under 50 SKUs and titles are already specific. The lift comes from cleanup, and you have nothing to clean up.
• Your conversion volume is under 30 per week per campaign. PMax cannot learn at that volume regardless of feed quality. Fix volume first.
• Your margins are uniformly low and there is no way to subsidize. Custom labels assume some products carry others. If everything is 8% margin, you are in a different conversation.
• Your category is genuinely commoditized and your only differentiator is price. Feed work helps but the ceiling is low.

For everyone else, the feed is the move.

What to do this week

Pick one PMax campaign that is underperforming. Do not pause it. Do not change the bid strategy.

Open the search terms report. Look at the top 50 queries by spend. For each, ask:

• Is the product matching this query the product I want to win this query?
• If yes, does my title actually contain the query keywords or just imply them?
• If no, what title change would route the query to the product I do want?

Spend an hour rewriting titles. Republish the feed. Wait 7 days. Measure.

If you want to see what this looks like at scale, our results page walks through 6 brands across fashion, home decor, and jewelry where this exact play moved the numbers. If you want us to run it for you, our process covers the full flow and pricing is here. Same playbook, run by the people who built it.

The feed is the campaign. The bid strategy is the dial. Stop tuning the dial when the radio is unplugged.