Performance Max Optimisation: Troubleshoot in Six Steps

When Performance Max isn’t working, check it in a fixed order. First, whether the conversions it learns on are real purchases with real order values. Second, whether the decline is a trend or an ordinary monthly swing. Third, whether each campaign gets enough conversions to learn. Fourth, whether your products reach the auction and where the spend goes. Fifth, whether your structure splits that volume too thin. Only then change targets, budgets and assets. Change settings on top of bad numbers, and the campaign retrains on the same errors.

Why the order matters more than any single fix

In most of the online stores in our data, Performance Max is the advertising itself rather than one campaign among many. In GetProfit data, Performance Max runs in 141 of 146 stores and takes a median 95.7% of their ad budget (June 2025 – June 2026). When it underperforms, the whole account underperforms, and every change you make lands on almost all of your ad spend at once.

Changes also cost time. The Google Ads API documentation (BiddingStrategySystemStatus) lists the reasons a bid strategy goes back into learning. Among them: the strategy was recently created or reactivated, its budget or settings changed, or the conversion types and settings it depends on changed.

Google’s API guide Compare performance with an existing campaign recommends running Performance Max campaigns for at least 6 weeks before comparing their performance, to give Google AI time to ramp up. During that learning period, results are unstable and tell you little.

So this guide follows one rule: fix what the campaign learns from before you change how it learns. If you set a new target on top of broken tracking, the campaign trains on the same broken data, and you spend two weeks finding that out. The six steps go from the cheapest, most decisive checks to the most expensive ones. Performance Max for online stores covers how to build the campaign in the first place.

Step 1: Can you trust the conversions the campaign learns on?

Performance Max bids towards the conversion actions you tell it to value. Google’s API guide Campaign-level Conversion Goals puts it the same way: in Performance Max, conversion goals guide how Google’s algorithms serve ads and allocate budget. A campaign with no goals set explicitly uses all of the account’s conversion goals. If those actions are wrong, the campaign learns to do the wrong thing well.

Four checks, in this order:

1. The event. The primary goal is a purchase, not an add to cart, a checkout start or a page view. Tell a campaign to find add-to-carts, and it finds people who add to cart.
2. The value. Each order sends its real conversion value. If every order carries one fixed number, revenue in the report is the order count times that number, and the ROAS built on it means nothing.
3. Double counting. Only one primary goal counts the purchase. A second one — an import from analytics, an old tag, a goal from your store’s platform — fires on the same order and inflates the count.
4. Outliers. No single order is many times larger than your usual order. A wholesale shipment, a test order or a checkout glitch looks like a great sale, and Google goes looking for more buyers like that.

Google flags part of this itself: in the API, a bid strategy can carry a “misconfigured” status. That happens when the conversion settings it depends on are set up wrong, or when the account lacks conversion types that could be reported against it. But a clean status proves little. To Google, a fixed value or a duplicated purchase is set up correctly, and it still feeds the campaign false numbers.

The portal’s conversion check runs the same four checks and weighs the result by spend: a campaign counts as much as the money that runs through it. It treats a conversion value more than 30 times the account’s median order value as an anomaly, lists it separately and leaves it out of the results. It also flags a product campaign whose average conversion value is below 0.30× or above 8× the account’s median order value, once that campaign has at least 10 conversions.

The portal’s store score treats tracking as a gate. When tracking is broken, the parts of the score that depend on conversions — structure, changes and product performance — count as zero. The feed and the assortment keep their points. The score drops, and it says why.

If any check fails, stop here. Fix tracking, then give the campaign time to learn on clean data. When a campaign spends but doesn’t sell, the same applies: check the numbers before you change targets. Our conversion tracking audit gives the full sequence for tracking.

Step 2: Is it a real decline or an ordinary month?

Before you look for a cause, make sure there is something to explain. The ROAS of a healthy store moves a lot from month to month on its own.

For the median store, a 16% swing is an ordinary month

In GetProfit data on 110 stores with at least eight months of history (June 2025 – June 2026), ROAS typically deviated from the store’s own median by 16.0% a month. For a quarter of the stores, the typical swing was 24.1% or more. At the median, the best month beat the worst one by 3.1×. One weak month is noise.

The portal raises a trend alarm when ROAS falls by 15% or more, or revenue by 10% or more, two months in a row. After one weak month, the portal waits for the next one.

Example store, not client data.

A tableware shop has a ROAS of 4.5 in July and 3.9 in August. That is a 13% fall: an ordinary swing, and the portal stays quiet. If ROAS goes 4.5 → 3.8 → 3.2 instead, it falls 15.6% and then 15.8%. Two falls in a row make a trend, so now look for a cause.

Product data before and after 15 June 2026 isn’t comparable

According to Google’s Retail Campaign Performance documentation, from 15 June 2026 the API’s product report includes all Performance Max networks, and metrics may show a one-time increase. The Google Ads Developer Blog post Product reporting changes for Google Ads starting June 15, 2026 adds that requests for dates before the launch will not contain the expanded data. Compare product totals on one side of that date only: a jump in mid-June may be a reporting change rather than growth. The details are in product data after 15 June 2026.

When the decline is real, find which part moved

ROAS equals conversion rate times average order value divided by cost per click. Within a store, ROAS moved together with conversion rate and average order value far more than with cost per click (GetProfit data, 1,360 store-months, June 2025 – June 2026). These are correlations, not causes, but they tell you where to look:

Then look at what changed in the account just before the decline: your own edits and anything applied automatically. If the decline begins right after a change, look at that change first.

Step 3: Does each campaign get enough conversions to learn?

Smart Bidding needs conversions to learn from. When there are too few, Google marks the bid strategy as limited by data: “not enough conversion traffic over the past weeks” (BiddingStrategySystemStatus, LIMITED_BY_DATA).

The portal treats 30 conversions a month per campaign as the minimum and 50+ as comfortable. In the portal’s view, more conversions per campaign mean less spread in results and less drift away from the target you set. Count pace, not totals: 50 conversions over six months is about 8 a month, and the portal reads that pace as a campaign that never finishes learning.

A quick test: divide your monthly conversions by 30. The result is roughly how many campaigns your volume can feed.

Example store, not client data.

A tableware shop gets 300 orders a month and runs 12 Performance Max campaigns. That is 25 conversions per campaign on average, below the minimum of 30. If two busy campaigns take 150 of the orders, the other ten share the remaining 150, or 15 each. At 30 per campaign, 300 orders can feed up to 10 campaigns; at a comfortable 50, six.

If most of your campaigns sit under the bar, the problem is the number of campaigns: there are too many for the volume. Merge campaigns before you change any target.

Step 4: Do your products reach the auction, and where does the spend go?

A campaign can only sell what gets shown. Two groups of products need a look: products that get no impressions at all, and products that get budget and never sell.

Unshown products: check campaigns and stock before the feed

Many products get no budget. In our study of 1.4 million products across 130+ stores, 55.0% of products received no budget at all over 13 months.

A broken feed is a rarer cause than it seems. In GetProfit data on 142 stores (data bank collected in August 2026), Merchant Center flagged less than 1% of the catalogue in 109 of them. In a sample of 400,000 flagged products, the most frequent flags were “not eligible in any campaign” and “out of stock”. Those are decisions about campaigns and stock, not feed errors.

So check your listing groups and product exclusions first, then stock, and only then the feed attributes. If a whole campaign barely spends or gets almost no impressions, work through why Performance Max is not spending.

Spend with nothing back is large: lower its share instead of switching it off

The study of 1.4 million products also found that stores spent a median 40.5% of their budget on products with not a single sale in 13 months. This is spend with nothing back: money that goes to products and never returns as a sale.

In the same study, 35.4% of products that had spent at least the cost of one conversion without a sale did convert in later months, after a median of 2 months. So reduce these products’ share of the budget instead of cutting them to zero.

Step 5: Is the structure splitting your volume too thin?

Structure comes after volume because structure decides how campaigns share the volume. The portal uses three working rules for campaigns and asset groups:

• split a campaign when it brings more than 100 conversions a month, or when parts of it differ in ROAS by more than 50%;
• split only if both resulting campaigns will still get at least 30 conversions a month;
• merge asset groups that get fewer than 5 conversions a month.

When a split is justified, the portal’s approach is to keep the proven core where it is. The products that sell steadily stay in the existing campaign with their learning history. Only the uncertain tail moves into a new campaign, so only that new campaign starts learning from scratch.

For typical structural mistakes and what to do instead, see the most common Performance Max mistakes.

Step 6: How should you change targets, budget and assets?

Budget and target changes send the bid strategy back into learning, so make them one at a time and in small steps. The portal works with these limits:

The portal also flags more than three edits a week on the same campaign or asset group. Frequent edits keep a campaign in learning all the time.

Judge each change on its own window. The portal compares the 14 days before and after a change.

It counts the change as working if ROAS rose by more than 20% while conversions stayed at 80% or more of the earlier level. It counts it as harmful if ROAS fell by more than 20% or conversions fell by more than half. Anything in between is unclear. With under 7 days of data, or no conversions, the portal holds the verdict.

Assets have a reporting catch of their own. Performance Max builds ads from the assets in each asset group. As the API’s Asset Group Level Performance page notes, you can’t view individual ad performance. Instead, you judge assets through the asset and asset group reports, and reading the asset report has its own rules.

Which report answers which question?

Google’s API guide Performance Max reporting splits reporting into several reports: campaigns and placements, asset groups, assets, retail (product) performance, locations and search terms. Each one answers its own question.

Our Performance Max reporting guide covers how each report is built and what it leaves out.

The six steps as a checklist

1. Conversions. Run the four checks: event, value, double counting, outliers. If one fails, fix tracking, change nothing else and give the campaign time to learn on clean data.
2. Trend. Compare the decline with the ordinary swing. One weak month is noise; ROAS down 15% or more, or revenue down 10% or more, two months in a row is a trend. Then find which part of ROAS moved (conversion rate, average order value or cost per click) and what changed in the account just before the decline.
3. Volume. Count conversions per campaign per month. Under 30 on most campaigns means too many campaigns for the volume.
4. Products. Find what gets no impressions, and why: listing groups and stock first, feed attributes second. Lower the share of spend with nothing back instead of switching it all off.
5. Structure. Split a campaign only if both parts will still get at least 30 conversions a month, and merge asset groups that get fewer than 5 a month.
6. Settings. Change targets, budget or assets one at a time, in small steps. Wait until the bid strategy has finished learning before you judge a change: the portal judges each change on 14 days of data, and Google recommends running a new campaign for at least 6 weeks before you compare its performance.

If you are taking over an account, or you need a full review rather than a fix, run a Performance Max audit instead.

Find out your account score. Sign in with Google in one click. The portal changes nothing without your consent.

Get your account score →

Sources

• Campaign-level Conversion Goals — conversion goals guide ad serving and budget allocation in Performance Max; without explicit goals a campaign uses all account-level goals. Checked 2 October 2026.
• Compare performance with an existing campaign — run Performance Max campaigns for at least 6 weeks before comparing performance. Checked 2 October 2026.
• BiddingStrategySystemStatus — what puts a bid strategy into learning; the “limited by data” and “misconfigured” statuses. Checked 2 October 2026.
• Retail Campaign Performance — from 15 June 2026 the product report includes all Performance Max networks; one-time increase in metrics; network breakdown from API v23. Checked 2 October 2026.
• Product reporting changes for Google Ads starting June 15, 2026 — requests for dates before the launch don’t contain the expanded data. Checked 2 October 2026.
• Asset Group Level Performance — individual ad performance can’t be viewed in Performance Max. Checked 2 October 2026.
• Performance Max reporting — the reports available for Performance Max, including search terms. Checked 2 October 2026.
• Google Ads API: Metrics — Auction Insights impression share is measured against the impressions your ads were eligible for. Checked 2 October 2026.
• GetProfit data: 146 stores, June 2025 – June 2026 — Performance Max share of budget.
• GetProfit data: 110 stores with at least eight months of history, June 2025 – June 2026 — month-to-month ROAS swing.
• GetProfit data: 1,360 store-months, June 2025 – June 2026 — what moves together with ROAS.
• GetProfit data: 142 stores, data bank collected in August 2026 — Merchant Center flags.
• GetProfit study: 1,404,808 products, 130+ stores, 13 months — products without budget, spend without sales, later sales.
• GetProfit portal methodology — conversion checks, trend alarm, per-campaign conversion bar, structure rules, change steps and the 14-day verdict.