When Performance Max Fails: Mistakes and Fixes

A Performance Max campaign that underperforms rarely announces why. The interface shows you a blended number across six surfaces, the machine reports that it is optimizing, and the ROAS just sits there being disappointing. I have written about building the campaign and the guardrails to set before launch. This post is the third situation: the campaign is live, the numbers are bad, and you need to find out what is actually wrong before you burn another month of budget on guesses.

My diagnostic order below is deliberate. It runs from the explanations that cost nothing to check toward the ones that require rebuilding, because most “Performance Max failed” stories I hear turn out to be one of the cheap ones.

First: is it failing, or still learning?

Before touching anything, establish that the campaign has had a fair trial, because the most common Performance Max mistake is judging it mid-calibration and reacting. Google’s learning period documentation says the bid strategy can take up to around 50 conversion events or 3 conversion cycles to calibrate, and the evaluation guidance I covered in the setup post wants six weeks before judgment, ramp excluded.

The corollary is the part people skip: every significant change restarts the clock. New targets, big budget moves, swapped conversion goals - each one puts you back in calibration while the dashboard shows you the chaos of a system re-learning. I check the change history first on any “failing” campaign, and if it shows weekly interventions, the campaign has never actually run. It has only ever been learning. The fix is organizational, not technical: stop, pick an evaluation date, and let one configuration finish a trial.

Mistake 1: there was never enough conversion volume

Performance Max is a machine-learning system, and it starves quietly on thin data. The largest independent study I know of - Optmyzr’s analysis of 24,702 campaigns across 9,199 accounts - found that accounts clearing roughly 60 conversions per 30 days did fine, and those below that threshold “saw worse performance across the board,” with CTR the only metric spared.

So the check is arithmetic: does this campaign (with account history helping) plausibly generate 60 conversions a month? If not, no amount of asset tinkering rescues it. The honest fixes are structural - a higher-funnel conversion action with real volume, a consolidated campaign instead of three thin ones, or the admission that a keyword-based Search campaign fits your volume better. I would rather run a small account on campaign types that work at small scale than feed a learning system too little to learn from.

Mistake 2: the numbers were inflated, and reality caught up

The second pattern is a campaign that looked great for months and “suddenly” got worse. Usually nothing changed except the audit. A Performance Max campaign without brand exclusions absorbs your branded searches and books them as wins. One with loose conversion goals optimizes toward newsletter signups priced like purchases. Both produce a flattering blended number that collapses the moment anyone measures properly.

Two checks take fifteen minutes. Pull the conversions by conversion action and see what the campaign is actually buying - if the cheap secondary actions dominate, the goals are the problem, not the machine. Then check whether brand exclusions are set, and if they are not, assume a slice of the reported performance was traffic you would have gotten free. This is not a fix that makes the campaign better. It makes the campaign honest, which changes what you decide to do with it - and it is the same self-graded-homework problem I flag in every platform’s claims about its own AI.

Mistake 3: reading the blend instead of the channels

For most of the format’s life you could not see where the money went, and Optmyzr’s gap analysis found 82% of advertisers were running Performance Max alongside other campaign types without channel-level visibility. That excuse expired: channel performance reporting now breaks out cost, clicks and conversions by Search, Shopping, YouTube, Display, Discover, Gmail and Maps.

On an underperforming campaign this report is where I look longest. The classic finding is a campaign that is really a Shopping campaign wearing a costume - conversions from Search and Shopping, budget bleeding into Display and YouTube impressions that convert nothing. That diagnosis has a direct fix: tighten the creative that feeds the weak channels (a campaign with no video gets an auto-generated one, and it serves), push spend discipline through account-level placement exclusions, and judge the campaign on what it demonstrably is for your account rather than on the six-surface promise.

Mistake 4: asset groups that give the machine nothing

The machine assembles ads from what you feed it, and thin or monotonous asset groups produce weak ads on every surface at once. The two failure signatures I see: one asset group carrying every product with fifteen near-identical headlines, and image-and-video slots filled by Google’s auto-generated remixes because nobody supplied real creative. Optmyzr’s gap analysis is interesting here too - among advertisers using full creative sets, 57% saw only average results, which I read not as “creative doesn’t matter” but as “volume of assets is not the lever, distinctness is.”

The fix is the asset-group discipline from the setup walkthrough: one theme per group, genuinely different angles across headlines, and your own video even if it is modest - the specs and the pruning cadence are in the assets guide. If the asset report shows one headline dominating every combination, the other fourteen were filler, and the ad copy prompts can generate real variety faster than a team brainstorm.

Mistake 5: expecting it to beat the campaign it replaced

Sometimes the campaign is running exactly as designed and the design was the mistake. The independent benchmarks are more modest than the sales pitch: Tinuiti’s Q3 2025 benchmark report found Performance Max conversion rates running 2% higher than standard Shopping - recovered from a 10% deficit earlier in the year - but with CPCs 7% higher, netting out to slightly lower ROAS for Performance Max. And the big Optmyzr study found accounts pairing Performance Max with Search and Shopping campaigns often saw the traditional formats win.

I take two rules from that. Performance Max is not an upgrade you install, it is a trade: broader reach and automated assembly, paid for in higher CPCs and less control, and whether that trade profits depends on your account’s data and creative. And a Performance Max campaign should be compared against the specific alternative it displaced, on your own margin math, not against its own previous inflated report.

Fix or rebuild?

My rule of thumb after the diagnosis. Fix in place when the problem is goals, exclusions, budget thrash or assets - those are adjustments the campaign can absorb, ideally batched into one change so learning restarts once instead of four times. Rebuild when the structure was wrong: conversion volume that will never feed the machine, one campaign doing jobs that need different goals, or a lead-gen account bidding on values that are fiction. And retire the format without shame when the channel report shows it doing, expensively, what a simpler campaign type already does well. The machine does not have feelings. Your budget does.

FAQ

How long should I wait before deciding Performance Max is failing? Six weeks of stable configuration, with the first one to two weeks of ramp excluded - and “stable” is the operative word, since every major change restarts the learning period. If you have never gone six weeks without an intervention, you do not yet have a result to judge.

Why did my Performance Max campaign stop converting suddenly? Check, in order: a conversion tracking change (the most common culprit and the least visible), a big edit that restarted learning, a budget or target change, and seasonality in your own demand. Genuine sudden decay of a healthy campaign is rarer than a measurement or configuration event that looks like one.

Should I add more asset groups to fix performance? Only if you have genuinely distinct themes to give them. More thin groups spread your conversion data thinner, and the machine learns less from each. Consolidation is the more common fix in the accounts I see, not expansion.

Is Performance Max just wrong for small accounts? Below roughly 60 conversions a month, the independent data says the format underperforms across the board, so I treat that as a real floor rather than a soft guideline. Small accounts usually get more from tightly-built Search campaigns plus Smart Bidding at the strategy level that matches their volume - and when that bidding layer itself misbehaves, the bid strategy diagnostic is the companion to this post.

This post completes the Performance Max trio in the paid media guides: the setup, the guardrails, and this diagnosis. Meta’s parallel automation has its own troubleshooting logic, and the wider practice starts at the complete guide to AI marketing.