Google Ads & Performance

Performance Max Learning Phases — Why Campaigns Underperform at First

PMax campaigns go through four phases. Understanding the mechanics helps you make better decisions and waste less budget. A practical guide.

EARNST · March 13, 2026 · 10 min read

Days 1 to 3 look fantastic. Day 5, everything collapses. Day 7, the budget gets halved. This is the most common and expensive mistake in Performance Max management — and it happens not out of ignorance, but as a perfectly understandable panic response to a systemic behaviour of Google's algorithm.

This is not an isolated case. Every PMax campaign goes through this pattern. Understanding the mechanics behind it means making better decisions, waiting at the right moments, intervening at the right moments, and wasting less budget.

The exploration-exploitation dilemma

Google's bidding algorithm faces a fundamental problem at every campaign launch: it has no data. No conversion signals, no audience insights, no knowledge of which asset combination works for which audience.

In computer science, this is known as the Multi-Armed Bandit Problem. Imagine a room with a hundred slot machines. Each has a different payout rate, but you know none of them. The only way to find out which machines are worth playing is to play and observe.

That is exactly what the algorithm does. It distributes impressions broadly, measures reactions, and sharpens its strategy with each data point. Every "wasted" impression in the first days is not a bug — it is the system learning. Exploration is not a malfunction but a necessary investment.

The four phases of a PMax campaign

Phase 1: Honeymoon (days 1–3)

The first days almost always look good. CPA sits well below target, conversion rate looks impressive. But the numbers are misleading.

Google serves the easiest audiences first: brand queries, the existing retargeting pool, and users who most closely match your audience signals. These conversions are low-hanging fruit — a large share of them would have happened without PMax.

Google shows you the best results first. This is not coincidental but part of the system logic: an immediately disappointing dashboard would cause advertisers to shut down the campaign before the algorithm even gets a chance to work.

Phase 2: Crash (days 4–7)

The low-hanging fruit has been harvested. Now the real prospecting work begins, and the bidding model shifts from "show what works" to "test systematically." CPA often rises to three to five times the target value.

Simultaneously, conversion delay kicks in: users who clicked on days 1 to 3 are only now converting — while clicks from days 4 to 7 have not generated conversions yet. The dashboard shows an apparent total failure, even though the pipeline is full.

This is the moment when most mistakes happen.

Phase 3: Recovery (days 8–21)

The algorithm has now collected enough positive and negative signals. Recovery is a filtering process, not a building process. Google is not primarily learning whom to target — it is learning whom to exclude.

Fluctuations get smaller, CPA approaches the target value. But performance never reaches honeymoon levels. That was never the real value of the campaign.

Phase 4: Steady state (from day 21–28)

The algorithm has converged. Performance oscillates around the actual value of the campaign. If that value falls below your target, the problem is not the learning phase — it is the foundation: budget, assets, landing pages, or audience strategy.

The three bottlenecks of the learning phase

Three factors determine how long the learning phase lasts and how stable the outcome will be.

Conversion volume. 30 to 50 conversions in 30 days is the minimum for stable bidding models. Staying below this threshold keeps the algorithm in a permanent learning phase. Insufficient budget is the most common reason.

Signal dimensionality. Audience times channel times asset times device times time-of-day times geography — that produces hundreds of thousands of combinations. The more variables the algorithm must optimise simultaneously, the more data it needs. Separating asset groups by theme reduces complexity significantly.

Conversion delay. The longer the purchase decision process, the longer the algorithm flies blind. For an online shop with impulse purchases, it is hours. For hotels or B2B services, 7 to 28 days can pass between first contact and conversion. During this time, the model lacks the feedback it needs to learn.

The biggest trap: panic reactions

What 90% of advertisers do in phase 2: halve the budget. Swap assets. Change bidding strategy. Pause the campaign and restart.

Every single one of these actions resets the algorithm to day 0. This creates an endless loop: honeymoon, crash, panic, reset, honeymoon, crash. The campaign never gets past phase 2, and the entire budget flows into exploration — without ever reaching the exploitation phase where actual optimisation happens.

The only correct response in phase 2: do nothing. Look at trends no earlier than day 14 — not at individual days. Making daily decisions on PMax is like stock trading based on the 5-minute chart: technically possible, practically ruinous.

When is a PMax campaign actually dead?

Not every bad performance is a learning phase. There are clear criteria for when shutting down is the right decision.

Hard kill criteria — one is enough

6+ weeks with under 30 conversions total. The algorithm does not have enough data to converge. Either the budget is insufficient, or the conversion event sits too far down the funnel.
CPA after 4+ weeks still over three times the target value. The campaign has converged — just not where you need it.
90%+ of budget flows into Display and Discover instead of Search and Shopping. This is the display trap: PMax uses cheap display impressions to inflate conversion numbers.
Brand cannibalisation over 50% of conversions. PMax claims conversions that would have come through brand search anyway. This is not customer acquisition but an accounting trick.

Soft warning signals — three or more means critical

ROAS stagnates below target after week 4
Impression share in parallel search campaigns drops
Asset performance shows barely any "Best" ratings
Audience signals are completely ignored by the algorithm
Strong fluctuations after 6+ weeks (the algorithm has not converged)
Conversion quality is poor (bounce rate over 70%)

AI Max for Search — same pattern, different level

AI Max for Search has been available since May 2025, globally with text guidelines since February 2026. The bidding engine is identical to PMax — and so is the honeymoon-crash cycle.

The difference lies in the risk profile. PMax expands across channels: the biggest risk is budget silently draining into display placements. AI Max expands across queries: the biggest risk is Final URL Expansion. Google sends users to pages never intended as landing pages — the careers page for a product keyword, the privacy policy for a service query.

The recommendation: use the 50/50 experiment split that Google offers. And restrict Final URL Expansion — or disable it entirely until you have enough data to assess the impact.

The right funnel architecture

PMax is a conversion optimiser. Not an awareness channel. The display and YouTube impressions that PMax serves are a byproduct of exploration — not controlled reach.

Building an awareness pool deliberately works better through Demand Gen (YouTube, Discover) or Meta (Facebook, Instagram). Feed these audiences into PMax as first-party signals. The chain looks like this:

Demand Gen and Meta build the awareness pool (15 to 20% of ad spend). These audiences get fed into PMax as signals. PMax converts the warm traffic — and works significantly more efficiently because the algorithm starts with qualified signals instead of cold traffic. Search captures organic intent — users actively searching for your product or service.

Running PMax without an upstream awareness funnel forces the algorithm to simultaneously create awareness and generate conversions. This can work, but systematically costs more and takes longer.

What you can do differently tomorrow

Five points for your next PMax launch:

Budget for at least 30 conversions per month. Below this threshold, the algorithm stays in the learning phase. If the budget is insufficient, define micro-conversions as intermediate goals.
Build audience signals on first-party data. Customer lists, website visitors, engagement audiences. The stronger the input signals, the faster the algorithm converges.
Separate asset groups by theme. One asset group per product category or service. No mixed groups that confuse the algorithm.
Define micro-conversions as intermediate goals. Newsletter signup, product page visited, configurator started. This gives the algorithm faster feedback, especially with long conversion delays.
Hands off for four weeks. Look at trends no earlier than that — not at individual days. Intervening in phase 2 means paying for the learning phase twice.

If after four weeks you find that the kill criteria apply, shutting down is the right decision. But make that decision based on data — not based on day-5 panic.