Performance Max for an Online Store: A Setup and Optimization Case Study

Performance Max stopped being an “experimental” format long ago: for most online stores it is the main sales channel in Google Ads. Standard Shopping campaigns have not gone anywhere, but it is PMax that gets priority in the Shopping auction and access to all of Google’s surfaces at once.

The catch is that PMax forgives few mistakes. A campaign with a messy feed, a single dumping ground of asset groups and a “let’s give it a try” budget will burn money just as confidently as poorly configured search — only figuring out where it went will be harder.

In this article I will walk through setting up and optimizing Performance Max for e-commerce using a real case from our agency — a jewelry online store in Ukraine where we raised ROAS from 2.8 to 5.1. If you need a general PMax guide (how it works, the formats, B2B scenarios), we have it separately: setting up Performance Max in Google Ads. Here the focus is specifically on online stores: the feed, the structure, the budget, optimization and the typical failures.

In short: Performance Max for an online store is a Google Ads campaign that pulls products from a Merchant Center feed and shows them across Search, Shopping, YouTube, the Display Network, Gmail, Discover and Maps at once. The foundation of the result is a quality feed and a well-thought-out asset-group structure; the algorithm needs 2–4 weeks of learning and 30–50+ conversions a month to reach a stable ROAS.

When a Store Needs PMax and When It Needs Standard Shopping

Performance Max is not an “improved version” of a Shopping campaign but a different tool with a different logic. Standard Shopping gives you control: you see the search queries, manage bids by product group and understand what you are paying for. PMax gives reach and automation but takes part of that control for itself.

The choice depends on the data you can feed the algorithm:

My practical rule of thumb is this:

1. A new account with no conversion history — start with standard Shopping. Collect 30+ conversions, clean up the feed, learn the real cost of an order, and only then turn on PMax.
2. An account with history and steady sales — PMax as the main campaign almost always wins on volume at a comparable ROAS.
3. Hybrid — the working scheme for mid-size and large e-commerce: PMax for the core catalog plus standard Shopping for products that need manual control (low-margin items, clearance of leftovers). Search Engine Journal in 2026 calls the PMax + Standard Shopping hybrid the most resilient e-commerce strategy — and our practice confirms it.

An important note on expectations: per the Triple Whale benchmark (2025, a sample of 18,000+ e-commerce brands), the average ROAS in Google Ads for e-commerce is around 3.7, and PMax campaigns average lower than branded search. That is normal: PMax works across the whole funnel, cold audiences included. Comparing its ROAS to a branded campaign is self-deception — something we will return to in the brand-exclusions section.

Case Study: A Jewelry Online Store — ROAS from 2.8 to 5.1

The central case of this article is a jewelry online store in Ukraine that came to us with a classic complaint: “The ads work, but break even.” Performance Max was already running and ROAS sat at 2.8 — for the jewelry niche, with its margins, that is the edge of profitability.

SEOquick experience. Over the course of the project we rebuilt Performance Max with regional campaign segmentation and raised ROAS from 2.8 to 5.1, while impressions grew 5.6x. The most striking result came from segmenting by city: in Odesa the campaign hit a ROAS of 116 — every hryvnia invested came back a hundredfold. The full breakdown is in our jewelry online store promotion case study (in Russian).

What it looked like “before”: one PMax campaign for the entire catalog and all of Ukraine, a shared budget, a single ROAS goal. In that configuration the algorithm does exactly what is most profitable for it: it spends the budget on the “easiest” conversions — cheap products and cities with high order competition — while expensive pieces and promising regions get no impressions.

What we changed — essentially, the rest of this article is a detailed breakdown of these steps:

1. Cleaned up the feed — attributes, images, GTINs, prices (the Merchant Center section below).
2. Split campaigns by region — separate campaigns for key cities with their own budgets and tROAS goals. This let us see that the conversion economics differ severalfold between cities, and reallocate the budget toward where it pays back better.
3. Rebuilt asset groups by category and margin — rings, earrings and chains each got their own group with their own creatives and audience signals.
4. Separated branded and non-branded traffic — so PMax would not “paint” itself a ROAS off people who were already searching for the store by name.
5. Gave the algorithm time — after each structural change we held out a relearning period without touching the settings.

Regional segmentation is a tactic I especially recommend for Ukrainian stores. A single “whole-country” campaign averages out the economics: Kyiv, with its expensive clicks and high competition, eats the budget, while cities with lower competition but real demand are starved of impressions. Separate campaigns by region make that difference visible and manageable. That is exactly how we found the “gold mine” in Odesa.

▶ Торговая кампания Google Shopping: Пошаговая настройка

Merchant Center and the Feed: The Foundation Everything Stands On

In Performance Max for e-commerce, the creatives are secondary. The feed is primary: it is where Google takes the product title, image and price from, and decides which queries to show the listing for. A weak feed means a weak campaign, and no settings will compensate for that.

A product feed is a structured file (or an API transfer) with data about your products: title, description, price, availability, image, identifiers. It is uploaded to Google Merchant Center, passes review and becomes the source for product listings.

The Minimal Set of Attributes That Must Be Perfect

• title — the most important attribute. Build it on the formula “Product type + brand + key features”: not “Ring 1023” but “Gold ring with a 0.1-carat diamond, 585 hallmark”. Google matches user queries against the title first and foremost.
• description — a detailed description with specifications and the words people search the product by. If you generate descriptions with a neural network, since 2025 Google requires marking such content with the structured_title/structured_description attributes, or you risk a disapproval for spec non-compliance.
• price and availability — must match the landing page to the penny. A price mismatch between the feed and the page is one of the main reasons for product disapprovals, and a systematic mismatch leads to an account suspension.
• gtin / mpn / brand — the product identifiers. For branded goods a GTIN is mandatory; reusing one GTIN across different products is a typical mistake that “drops” positions in the feed. For handmade and own-production items, the brand + mpn pair is enough.
• image_link — from 2026 Google is raising the minimum image-resolution requirement to 500×500 pixels (higher for apparel). Photos on a white background, with no watermarks or promotional captions.
• google_product_category and product_type — Google’s category and your own hierarchy. product_type matters especially: it is convenient for building ad groups and listing groups inside asset groups.
• sale_price, custom_label_0–4 — custom labels are your segmentation tool: margin, seasonality, bestsellers. Without them you cannot build a “by margin” structure (more on it below).

Typical Feed Mistakes We Find in Audits

1. The price and availability in the feed lag behind the site (the feed updates once a week, stock changes daily). The fix is automatic feed updates at least once a day plus automatic product-data improvements in Merchant Center.
2. Disapproved products hang there for months — no one looks at the Merchant Center diagnostics. And that is a direct hit to reach: a disapproved product simply does not show.
3. Truncated or “glued together” titles where the key words sit at the end and get cut off in the results.
4. A single “as-is” feed from the CMS without refinement — an export from the admin panel almost never meets Google’s requirements without attribute mapping.

In the jewelry case, working on the feed was the first stage: while disapproved products and broken titles hang in Merchant Center, optimizing the campaign is pointless — the algorithm gets garbage as input.

▶ Google Merchant: Основы продвижения интернет-магазина

Asset-Group Structure: By Category and By Margin

An asset group is a container inside a PMax campaign that combines creatives (headlines, descriptions, images, video), audience signals and listing groups of products from the feed. In essence it is the equivalent of an ad group, only for all surfaces at once.

Two extremes that are equally harmful:

• One asset group for the entire store. The algorithm shows rings for earring queries, generic creatives appeal to no one, and in the reports it is impossible to tell which category is pulling the result.
• Fifty asset groups for every little thing. The data is spread thin, no group gathers enough statistics to learn, and creatives get duplicated.

The working logic for an online store is segmentation along two axes:

1. By product category — as many groups as you have fundamentally different categories with different demand and a different audience. In the jewelry case that means rings, earrings, chains and so on: each category has its own creatives (wedding rings and stud earrings are bought by different people in different situations) and its own listing groups by product_type from the feed.
2. By margin — through separate campaigns. Here is the subtlety: the tROAS goal is set at the campaign level, not the asset-group level. So if you have products with a 15% margin and a 60% margin, they need different campaigns with different goals — high-margin items can afford a lower tROAS (aggressive growth), low-margin items get a strict goal. Segment them via custom_label in the feed.

Practical rules we apply:

• Fill out the creatives to the max: up to 15 headlines, descriptions, at least 3–5 images in different formats, video. If you do not upload a video, Google will generate one itself from the feed images — and that almost always looks worse than even a simple clip assembled by hand.
• The texts and images of an asset group must match its category. A “Wedding rings” group with a photo of a chain is a real example from an audit, and it works exactly as badly as it sounds. Incidentally, draft headlines and descriptions for each category are convenient to mass-generate with ChatGPT or Gemini — we have a collection of 50 prompts for ChatGPT and Gemini, some of which adapt nicely to ad copy.
• Build listing groups from the feed hierarchy (product_type), not manually by product IDs — otherwise new products will fall out of the campaign.

Audience Signals: A Hint, Not Targeting

An audience signal is not a limit on impressions but a starting hint to the algorithm: “begin with these people.” From there PMax expands reach to similar users on its own. So there is no need to fear “narrow” signals — but an empty signal slows down learning: the algorithm spends the budget on cycling through everyone indiscriminately.

What to put into the signal for an online store, in order of value:

1. Data about your buyers (Customer Match) — an export of emails/phone numbers from your CRM. The strongest signal: these are people who have already proven they buy from you.
2. Site visitors and abandoned carts — remarketing segments from GA4.
3. Custom segments by search query — people searching for “buy a gold ring”, “wedding rings price” and the like. For each asset group, the queries of its own category.
4. Google in-market audiences for the category topic.

Each asset group gets its own signal, thematically aligned with its products. A single “all buyers” signal across all groups is lost precision.

From 2026 Google is tightening up audience transparency: PMax is gaining full audience reporting (down to demographics) and first-party audience exclusions at the campaign level — keep an eye on these updates, they address part of the old “black box” complaints.

Brand Exclusions and Excluding Existing Customers

Two mechanisms without which you will not learn the real effectiveness of PMax.

Excluding Branded Traffic

By default PMax happily shows up for queries containing your brand — and credits itself with the conversions of people who were already coming to you. The campaign’s ROAS looks great, but there is no real lift in sales.

What to do:

1. Create a Brand Exclusion List: the store name, common typos, spellings in Latin and Cyrillic script.
2. Apply the list at the PMax campaign level.
3. Catch branded traffic with a separate cheap search campaign — that way you see the honest picture: how much it costs to attract new demand versus servicing existing demand.

In the jewelry case, separating branded and non-branded traffic was one of the key steps: before it, part of that “pretty” ROAS was driven by repeat customers searching for the store by name.

Excluding Existing Customers

If your goal is growing the base, turn on the New Customer Acquisition goal in the campaign settings. Two modes:

• New customer value — bids for new customers are raised, but impressions go to everyone;
• New customers only — impressions go only to those Google does not consider your customer (based on uploaded Customer Match lists, the tag and auto-detection).

For the “new only” mode, upload an up-to-date buyer list from your CRM to Google Ads and refresh it regularly — on a stale list the exclusion works like a leaky sieve. Remarketing to the existing base, meanwhile, should go into separate campaigns with their own economics: bringing a customer back is almost always cheaper than acquiring a new one, and mixing this money in one pot means fooling yourself with averages again.

Budget and Algorithm Learning: How Long to Wait and How Much to Pay

Performance Max is machine learning, and that requires data and time. The most common cause of “PMax doesn’t work” is that the campaign was strangled by edits or budget before it had a chance to learn.

Benchmarks confirmed by both practice and the 2025–2026 industry guides:

• Learning period: 2–4 weeks after launch or a major edit. In the first 1–2 weeks the CPA can be 30–50% above target — that is a normal calibration mode, not a reason to switch everything off.
• Conversions: at least 30 a month per campaign, comfortably 50 and up. If you do not have that many orders, do not split the budget across five campaigns — one that learns is better than five forever “raw” ones. Alternatively, train the campaign on micro-conversions (add-to-cart) before switching to purchases.
• Daily budget: a rule of thumb is 3× the target CPA. If an order costs 300 UAH, a daily budget below 900–1,000 UAH will almost certainly stretch learning out over months.
• Do not crank up the tROAS goal at the start. Set an achievable goal (slightly better than current reality), let the campaign stabilize, and raise it in steps of 10–15% no more than once every 2 weeks. A sharp goal increase = a new learning cycle.
• Change the budget gradually too — in steps of up to 20–30%, otherwise you will reset the accumulated learning.

And the main rule: one meaningful change at a time. If you simultaneously changed the goal, reshuffled the asset groups and uploaded a new feed, you will never know what exactly worked or broke.

In the jewelry case, patience turned out to be no less important than the settings: after rebuilding the structure we deliberately did not touch the campaigns during the relearning period, even though the numbers looked worse than expected in the first weeks. The ROAS growth to 5.1 came after things stabilized.

Optimization: What to Watch After Launch

PMax is no longer the “black box” it was in 2022. Google has gradually opened up control tools, and not using them means optimizing blind.

Channel Performance Report

Channel reporting shows a breakdown of results by surface: Search, Shopping, YouTube, Display, Gmail, Discover, Maps. The first thing to check on an e-commerce campaign: what share of the budget goes to Display and YouTube, and what it brings in. If the Display Network eats 30% of the budget at zero conversions, that is a signal of a problem with the creatives or the signals (and sometimes a reason to trim “parasitic” reach via placement exclusions at the account level).

Search Themes and the Search Terms Report

Search Themes are your way to hint up to 25 themes per asset group to the algorithm: the phrasings you want to show up for. They work as a soft analog of keywords. And the search terms report (a full one, like in Search — Google opened it for PMax) is the way to see what you actually showed up for.

A regular query-optimization cycle:

1. Once a week, review the search terms report.
2. The irrelevant — into negative keywords: since 2025 PMax supports up to 10,000 negative keywords at the campaign level (they work for Search and Shopping).
3. Strong queries that are not in your themes — add them to the Search Themes of the relevant asset group.

Insights and Asset-Group Reports

The Insights tab shows which audiences and themes bring conversions, plus demand trends. Asset-group reports show conversions, value and spend for each group: treat weak groups by replacing creatives and refining the signals, not by immediately turning them off. While there, watch the asset performance ratings (low/good/best) and replace “low” ones with fresh variants.

Product View

For a store, a regular products report is a must: which items get impressions and which are “asleep.” The typical picture is that 20% of products eat 80% of the budget. From there the decision is: “sleeping” products are either fixed at the feed level (title, price, image) or moved into a separate campaign with their own budget (the classic “zombie feed” scheme).

SEOquick experience. Shopping campaigns with the right segmentation give a store predictable economics: in our PPC optimization case study for a Shopping campaign (in Russian) we took an online store’s Shopping campaigns to a ROAS of about 9, and the conversion rate for Kyiv to 20.8%. The principles are the same ones described in this article: a clean feed, segmentation, regional analytics.

Typical PMax Failures in E-commerce

Here is the anti-rating — the mistakes we most often find in audits of online-store PMax campaigns:

1. Launching on a dirty feed. Disapproved products, broken titles, price mismatches. PMax amplifies whatever is in the feed — chaos included.
2. One “do-it-all” campaign — the entire catalog, the whole country, a single goal. An average across the board instead of managed economics. The cure is segmentation by category, margin and region (as in our jewelry case).
3. Branded traffic inside PMax. The campaign “paints” a ROAS off people who would have come anyway. Without brand exclusions you do not know the real effectiveness.
4. Panic during the learning period. Edits every three days, budget swings, switching it off after two weeks “because it’s expensive.” Every edit zeroes out the progress.
5. An unrealistic tROAS at the start. A goal of 800% against an actual 300% — the campaign simply stops spending and showing.
6. Empty asset groups. One headline, two images, no video. The algorithm has nothing to test — reach and channels get cut.
7. Ignoring the search-terms and channel reports. The control tools exist — people just do not use them. The budget leaks for months on irrelevant queries and empty Display reach.
8. Blind faith in “the automation will figure it out.” PMax automates execution, but not strategy: the structure, the feed, the signals, the goals and the analytics are still set by a human.

A separate word on measurement: without correctly configured e-commerce conversions in GA4 (with the order value passed through!) PMax optimizes toward nothing. Verify the transaction value pass-through before launch — an algorithm that learns on “0 UAH conversions” will chase cheap orders instead of profitable ones.

Instead of a Conclusion

Performance Max for an online store is not a “money” button but a system: feed → structure → signals → budget → optimization discipline. In our jewelry case, no single element on its own would have produced ROAS growth from 2.8 to 5.1 — what worked was the sequence: a clean feed, regional segmentation, the separation of branded traffic and patience during the learning period.

And remember: the search results are changing — more and more purchasing decisions begin in AI answers. How to prepare a store for this shift, read in our material on GEO optimization of a site for ChatGPT and AI search.

FAQ

What is better for an online store: Performance Max or standard Shopping?

It depends on the data. A new account with no conversion history is better off starting with standard Shopping: it gives control and reaches profitability faster. For a store with 30–50+ conversions a month, PMax usually brings a larger sales volume. The optimum for mid-size e-commerce is a hybrid: PMax for the core catalog, Shopping for items that need manual control.

How long does Performance Max take to learn?

Basic calibration takes 2–4 weeks, full stabilization up to 6–8 weeks. In the first weeks the CPA can be 30–50% above target — that is normal. Every major change (goal, budget ±30%, structure) triggers relearning, so change one parameter at a time.

What budget does PMax need for an online store?

The benchmark is a daily budget of at least three target CPAs, so the campaign gathers 30–50 conversions a month. If you have fewer orders, do not split the budget across several campaigns: one trained campaign beats five “raw” ones. Very small stores are wiser to start with standard Shopping.

Can you add negative keywords in Performance Max?

Yes. Since 2025 PMax supports campaign-level negative keywords — up to 10,000 per campaign (they apply to Search and Shopping). Plus there is the search terms report, brand exclusions for branded traffic and Search Themes (up to 25 per asset group) for hinting the themes you want.

Why does PMax have a high ROAS while sales don’t grow?

Most often the campaign cannibalizes branded traffic: it shows up for people who were already searching for your store and claims their conversions. Apply a brand exclusion list, move the brand into a separate search campaign and turn on the “new customer acquisition” goal — then you will see the real incrementality.

How many asset groups should a campaign have?

As many as you have fundamentally different product categories with different audiences — usually 3–7. Each group gets its own creatives, its own audience signal and its own listing groups by product_type. Split different margins into separate campaigns with different tROAS goals, since the goal is set at the campaign level.