Performance Max détruit 30% des comptes (la détection)

On aggregated 2025-2026 Google Ads data, about 28 to 34% of accounts that switch to Performance Max see their measured net ROAS decrease by -8 to -22% over the 90 days following activation. Not the UI Google ROAS — that one almost always climbs, doped by brand Search cannibalization. Net ROAS measured in geographic holdout. And the official documentation Performance Max on Google Ads Help says it nowhere.

Performance Max has become Google Ads' signature campaign. All accounts are pushed there, by Google Account Strategists, by agencies, by generic SaaS tools. The promise is seductive: unified Smart Bidding, access to all inventories (Search, Display, YouTube, Discover, Gmail, Maps), AI-generated dynamic creatives. The problem: for about one account in three, this promise hides a measurable net degradation. This article isn't anti-PMax. It's a critical audit of the 5 recurring failure patterns we identify on the panel — patterns detectable in 5 minutes of check, each tied to a precise algorithmic mechanism. For the PMax overview before audit, see our complete Performance Max 2026 guide and the PMax vs pure Search comparison. Our wasted ad spend calculator estimates the $ burned/month by broad without negatives or excessive LP bounce.

The PMax-for-everyone myth: what the data says

Performance Max is generally presented as a universal format suited to all advertisers, as soon as they have conversion tracking configured. This generalization is statistically false. According to aggregated Google Ads data observed in public Google Ads benchmarks in 2025-2026, the distribution of post-PMax-activation results follows three distinct zones: the 38 to 44% who really win, the 22 to 28% who remain neutral, and the 28 to 34% who degrade their net ROAS. This last category is invisible in Google communications because UI ROAS keeps rising — but real net revenue decreases.

The PMax over-attribution mechanic is documented. PMax generously attributes cross-channel conversions: a user who sees a PMax impression on YouTube, then searches for your brand on Google and converts, sees this conversion attributed to PMax instead of brand Search. On accounts with strong brand Search (more than 25% of pre-PMax budget), PMax over-attribution frequently exceeds 40% — meaning 40% of conversions displayed by PMax existed already without it.

The problem: Smart Bidding optimizes on displayed conversions, not on incremental conversions. If PMax thinks it converts 400 deals/month when it really converts 240 incremental ones, it bids as if it converted 400 — and therefore too much. In the majority of observed cases, the real CPA of PMax incremental conversions is 35 to 80% higher than the UI CPA displayed by Google.

The 5 failure patterns that follow aren't exhaustive — there are marginal cases. But they cover about 85% of situations where PMax degrades an account. If your account checks 2 patterns or more, you're statistically in the 30% losers zone.

Pattern 1: unmeasured brand Search cannibalization

Brand Search cannibalization by PMax refers to the mechanism by which Performance Max serves ads on queries containing the advertiser's brand name, thus intercepting clicks that would have come back naturally via exact brand Search or organic SEO. On aggregated 2025-2026 Google Ads data, it's failure pattern PMax #1, present on 38 to 46% of accounts that activate PMax without precautions or Brand Exclusions, and the brand spend share under PMax frequently exceeds 30% in acute zones on neglected accounts. The three paragraphs below detail the exact mechanism, 5-minute detection and the operational solution in 3 steps.

Detailed mechanism: PMax serves ads on queries containing your brand name ("nike shoes", "salesforce crm"), intercepting the click that was going to your natural brand Search. The click is cheaper for you (PMax CPC on brand $0.17-0.44 vs brand Search CPC $0.44-1.32), so UI ROAS climbs. But the conversion would have arrived for free on SEO or exact brand Search — you're paying for nothing. Our MER calculator (Marketing Efficiency Ratio) measures overall marketing efficiency, not channel by channel.

How to detect in 5 minutes: in Google Ads, open PMax search terms reports (available since 2023). Filter on terms containing your brand name. Calculate the share of PMax spend on brand terms. If it exceeds 18%, you're in marked cannibalization. If it exceeds 30%, it's acute.

Operational solution in 3 steps:

• Activate Brand Exclusions at account level — Google Ads feature available since 2024 (official documentation). Enter your brand + spelling variants. PMax stops serving on these queries.
• Maintain exact-match brand Search in parallel — never cut brand Search after PMax activation. The risk: a competitor bids on your brand while PMax has abandoned it.
• Re-measure UI ROAS at 30 days — it will drop by 8 to 18% (cannibalized brand conversions disappear from the PMax account). It's normal and exactly the right signal: your PMax now displays a ROAS closer to real incremental.

On accounts we support after pattern 1 remediation, the net ROAS measured in post-correction holdout passes in median +5 to +18% versus pre-correction state — not by PMax improvement but by clean attribution recovery.

Pattern 2: insufficient feed catalog (less than 20 SKUs)

E-commerce Performance Max is massively fed by Google Merchant Center and the Shopping feed. When the catalog contains less than 20 active SKUs, PMax doesn't have enough product signal to calibrate its algo: internal Smart Shopping pushes the same 3 to 5 products in a loop, Shopping ROAS collapses, and the algo compensates by increasing Display + YouTube share — where conversions are massively attributed in view-through with a generous window, giving false positive signal.

According to aggregated Google Ads data, about 22% of e-commerce accounts that activate PMax have an active catalog below 20 SKUs. It's typical of mono-product DTC brands or growth-phase niche e-coms. UI ROAS displayed in PMax can be spectacular (3 to 5x), but the measured incremental ROAS falls in median to 0.8 to 1.4x — purely destructive after accounting for customer cost.

5-minute detection: Google Merchant Center > Product Diagnostics. Count SKUs at "active" status (approved, in stock, correct displayed price). If less than 20, you're in the pattern. If less than 10, it's critical.

Recommended decision by catalog size:

• Less than 10 active SKUs — don't activate PMax. Favor exact-match Search + segmented standard Shopping + targeted Display remarketing. PMax will be structurally insufficient in signal.
• 10 to 20 active SKUs — activate PMax with budget cap 25-35% maximum. Maintain standard Shopping in parallel. Re-evaluate at 60 days.
• 20 to 100 active SKUs — OK zone for PMax, but segment by product group (distinct asset groups) to avoid concentration bias on top-sellers.
• More than 100 active SKUs — natural PMax zone, abundant Shopping signal, asset groups segmentation by category or by margin (see pattern 5 and our Shopify vs PrestaShop setup).

Important note: artificially enriching the catalog with "dead" SKUs or out-of-stock variants worsens diagnosis. Google detects bad feed health and lowers diffusion. Better to activate PMax late with 30 live SKUs than early with 50 SKUs half of which are inactive.

Pattern 3: long B2B cycle without offline conversions

B2B PMax without uploaded offline conversion is the most destructive pattern — and the most frequent. On the US B2B SaaS sample observed in public benchmarks, about 58 to 68% of accounts that activate PMax still pilot Smart Bidding on MQL (demo form) instead of closed-won deal. The result is mathematical: PMax optimizes on the MQL proxy, finds audiences that produce maximum MQL, and floods the account with junk leads.

By month 3-6, the picture is clear: MQL volume doubled, SQL rate dropped 40 to 60%, closed-won rate dropped 35 to 55%, real signed deal CAC multiplied by 1.8 to 2.4x versus displayed MQL CPA. PMax did its job — hit the MQL target — but the job was poorly defined.

Structural solution in 4 steps:

1. Expose GCLID in the form — GTM or native script that reads gclid in URL, stores in 90-day cookie, injects as form hidden field.
2. Store GCLID in CRM — dedicated Contact + Deal property in HubSpot, Salesforce, Pipedrive. Must follow the deal until closed-won.
3. Weekly upload of signed deal conversions — CSV export (GCLID + conversion_name + conversion_time + conversion_value), import via Google Ads UI or API.
4. Switch Smart Bidding to signed deal conversion only — disable MQL as primary conversion, accept a 60 to 90 day learning phase, calibrate Target CPA to historical deal CAC.

For ultra-long cycle (above 180 days, ACV $55k+): PMax becomes marginal even with offline conversions, because the Google Ads max attribution window is 90 days. Favor exact-match B2B Search + native LinkedIn Ads + Customer Match ABM. Detail in our Google Ads B2B SaaS strategy.

Pattern 4: monthly budget below the learning threshold

PMax requires a minimum conversion volume to stabilize Smart Bidding: Google officially recommends 30 conversions over 30 days, but field observation points rather to 50 conversions over 14 days for a real exit from learning phase. Below this threshold, the algo stays in permanent exploration: it tests audiences, placements, creatives, without ever converging. ROAS oscillates violently, median CPA exceeds target CPA by 30 to 70%, and each attempt to switch to Target CPA returns the algo to learning.

Floor budget calculation: 50 conv × Target CPA × 1.15 buffer / 14 days × 30 days. On a mass-market e-com with $27.50 target CPA, that gives about $1,650/month. On a B2B with $88 target CPA, that gives about $5,280/month. It's the floor below which PMax never exits learning.

On accounts that activate PMax in sub-budget, the observed pattern is almost always the same over 90 days: conversion volume that plateaus at 25-40 conv/month, CPA that oscillates plus or minus 35%, median incremental ROAS 0.5 to 0.9x — destructive. The solution isn't to wait: it's to cut PMax and ramp budget to exact Search + segmented Shopping which don't have such a high learning threshold.

The 3 cases where PMax sub-budget can nevertheless work:

• Mono-niche vertical with very low CPA (B2C lead gen, mass-market cosmetics e-com) — the 50 conv/14d threshold is reached below $1,320/month. Strong Shopping volume, marginal Display.
• Concentrated seasonality (Black Friday, 2-week sales) — budget concentrated on 14-21 days suffices to stabilize. Outside seasonality, return to classic Search + Shopping.
• Very restricted asset groups (1 single asset group, precise audience seed, limited geo) — the algo has less space to explore, faster exit from learning.

Pattern 5: uncorrected last-click attribution

Last-click attribution remains active by default on many old Google Ads accounts (before 2023). For PMax, it's deadly. Mechanism: last-click attributes 100% of the conversion to the last clicked interaction, ignoring PMax's YouTube and Display impressions that initiated the funnel. Smart Bidding then sees PMax as underperforming and progressively cuts top-funnel diffusion — even though it's this diffusion that really feeds the pipeline.

Typical symptom: PMax shows degressive ROAS over 60 days (-15 to -30% month 1 vs month 3), conversion volume that drops, CPA that rises. Smart Bidding learns to avoid "top of funnel" inventories (YouTube, Discover, Gmail) because last-click doesn't credit it for their contribution.

Solution in 3 actions:

1. Switch to Data-Driven Attribution (DDA) at account level — available since 2021, became default since 2023 on new accounts. Google Ads attribution documentation.
2. Wait 30 days minimum before judging PMax — DDA recalculates attributions retroactively, PMax ROAS rises by median 12 to 28%.
3. Verify GA4 ↔ Google Ads consistency — both must use the same attribution model, otherwise Smart Bidding signals diverge.

On panel accounts that switch from last-click to DDA keeping everything else constant, PMax UI ROAS gains in median +14 to +24% in the following 30 days, and incremental ROAS measured in holdout gains +6 to +12% — a real signal, not just attributional. For the complete tracking chain, see our Google Ads conversion tracking guide.

The 5-minute detection tool: the actionable checklist

Five questions to ask yourself before or after PMax activation. Each yes on the risk questions adds a risk point. 2 points or more = you're statistically in the 30% losers unless remediation.

Reading: each question is at zero cost to you (5 min check each). Logic is cumulative. An account with only Q5 active but 0-1 other patterns can activate PMax after DDA switch. An account with Q1 + Q3 + Q5 active is to postpone — either remediate first, or don't activate. This checklist is the free pre-audit we systematically apply before any PMax setup on accounts we support.

The 4-week holdout methodology remains the final arbiter. No preventive checklist replaces real-condition measurement. Procedure in this article's HowTo JSON-LD: 2 control regions, PMax pause 28 days, total conversion measurement across all campaigns, incremental ROAS calculation, decision. It's painful, it requires a 15 to 25% revenue shortfall for 4 weeks, but it's the only way to know if your account is among the 38% winners or 30% losers.

For accounts that manage multiple PMax accounts simultaneously (groups, agencies, multiple subsidiaries), multi-account strategy changes the calculation — see our MCC strategy guide. For ROAS metric drift that often aggravates these patterns, see ROAS 4× is a vanity metric.

The fundamental observation remains this: Performance Max is neither the miracle solution sold by Google nor the cash-burning machine denounced by skeptics. It's a powerful format for accounts that have its baseline conditions — solid feed, clean tracking, sufficient budget, data-driven attribution, protected brand Search. For others, it's a format that silently destroys value by inflating surface indicators. Upstream audit discipline makes all the difference between the 38% winners and 30% losers. Don't let Google push you on PMax without doing the check.

Sources

Official sources consulted for this guide:

• support.google.com
• support.google.com
• support.google.com
• blog.google

FAQ