The Algorithm Tax: What Pausing or Restarting Ads Actually Costs You

The Hidden Cost of Budget Changes

Every significant budget change comes with a tax you don't see on your invoice. Platforms call it the "learning phase." Practitioners call it wasted spend.

When you change budget by more than 20%, both Google and Meta throw out their accumulated optimisation data and start exploring again. CPCs spike. CPA jumps 25-40%. Conversions become erratic. This lasts 7-14 days — sometimes longer.

Meta's own documentation states that ad sets need approximately 50 optimisation events within 7 days to exit the learning phase. Google Ads Smart Bidding documentation cites a similar threshold (~50 conversions) before Target CPA or Target ROAS stabilizes. Eric Seufert at Mobile Dev Memo has documented the dynamic for years in mobile UA: every bid-strategy change is effectively a new auction with no prior data, and CPM volatility persists until the optimizer accumulates enough signal to converge.

If you make another change during this period, the clock resets. You pay the tax again.

Understanding this cost is essential before acting on any budget recommendation.

Google Ads Learning Phase

Standard Search Campaigns

Google's learning phase depends on conversion volume, not click volume. A campaign with 1,000 clicks but 8 conversions per month will never optimise well, regardless of budget.

Performance Max (PMax)

PMax takes longer — 4-6 weeks despite the "Learning" status disappearing earlier. The algorithm continues internal calibration for roughly 45 days. Performance during weeks 2-4 often looks worse than during the initial learning status period.

Don't judge PMax performance until week 6.

What Resets Google Ads Learning

Any of these changes restart the learning phase:

• Budget changes exceeding 20%
• Switching bid strategy (e.g., Target CPA → Target ROAS)
• Adjusting bid targets significantly
• Changing conversion actions (adding, removing, or modifying tracking)
• Major ad copy changes (multiple ads rewritten simultaneously)
• Audience or geographic targeting changes
• Pausing and reactivating campaigns

The key insight: changes under 20% generally don't reset learning. Conservative practitioners use 15% as their threshold.

Recommended: incremental adjustments of 10-15% with at least one week between changes. If you need a 40% increase, do it as two 20% steps with a week gap.

Meta Ads Learning Phase

Standard Campaigns

Meta's learning phase is shorter than Google's but more fragile. Small changes that wouldn't affect Google can reset Meta.

What Resets Meta Learning

• Budget changes over 20-25% in a single adjustment
• Changing bid strategy or bid cap
• Editing target audience (even minor refinements)
• Switching optimisation event
• Adding or removing creatives from an ad set
• Pausing an ad set for 7+ days

That last one is critical. Pausing a Meta ad set for more than 7 days effectively kills it. When you restart, you're starting from scratch — no accumulated learning, no optimised delivery, no audience signals.

The Pause Penalty

Meta now shows a useful indicator: "You can increase your budget to $X without restarting learning." Use this. It tells you exactly how much headroom you have before paying the tax.

CPA Impact During Learning

Expect CPA to rise 25-40% during any learning phase reset. This isn't the algorithm failing — it's deliberately exploring different audience segments, placements, and creative combinations. The instability is the learning.

The temptation is to intervene. Resist it. Every intervention during learning resets the clock.

Other Platforms

TikTok

• Learning phase: 50 conversions needed, similar to Meta
• Highly variable results — coefficient of variation of 1.26 in benchmark tests (vs 0.7-0.8 for Meta)
• More volatile by nature; longer evaluation windows recommended (8-12 weeks)
• Minimum: $20/ad group, $50/campaign daily

LinkedIn

• Slower learning, higher minimum costs ($10/day minimum, $25-50 realistic)
• 90-day click attribution window (longest of any major platform)
• B2B conversion cycles are longer — give LinkedIn 12+ weeks before judging
• Minimum lifetime budget: $100

Pinterest

• Similar learning mechanics to Meta
• Smaller audience = slower learning
• Minimum: $0.01/day technically, $500/month realistically

Adstock: Why Sales Don't Drop Immediately

You pause all ads on Monday. By Friday, sales look... fine. So the ads weren't working?

Not necessarily. This is adstock — the lingering effect of advertising after you stop spending.

How Adstock Works

Advertising builds awareness that decays gradually, not instantly. The decay rate depends on the channel:

The Adstock Trap

If you pause a channel and sales hold steady for a week, the natural conclusion is "that channel wasn't working." But if the channel has 2+ weeks of adstock, you're measuring during the carryover period. Sales haven't dropped yet because awareness hasn't decayed yet.

The true impact shows up 2-4 weeks after pausing, by which time you've already reallocated the budget and can't easily run a controlled comparison.

This is why incrementality testing uses geo holdout experiments with 30-day minimum durations — long enough for adstock to fully decay and reveal the true incremental impact. The Dropbox IEEE Access paper (Chivukula et al., 2026) used month-long geo-blackouts on the US market precisely because anything shorter still has adstock in the carryover window. Recast and Mutinex fit per-channel adstock curves directly from spend-vs-revenue data; Meta's open-source Robyn library does the same. The half-life that comes out of the model is the one to use, not the rule-of-thumb table above.

What Adstock Means for Budget Decisions

• Don't judge a pause too quickly. Wait the full adstock half-life (channel-dependent) before concluding a channel wasn't contributing.
• Search is the exception. Search ads have near-zero adstock. If you pause Google search and traffic drops immediately, that's expected.
• Brand effects last longer. If you've been running brand awareness campaigns, the effects linger for weeks. This is good — it means you don't need to spend continuously at maximum. But it also means you can't run a quick pause/restart test and draw valid conclusions.

The Safe Change Protocol

A framework for changing budget without paying unnecessary algorithm tax:

Step 1: Size the Change

Is the total adjustment under 20%? If yes, proceed directly. No learning phase disruption expected.

If over 20%, break it into 2-3 steps of 10-15% each.

Step 2: Change One Variable

Never change budget, bid strategy, and audience simultaneously. Change budget first. Wait 7-14 days. If performance stabilises, make the next change.

Step 3: Wait the Full Learning Period

Don't judge performance during this window. Don't make additional changes. Don't panic.

Step 4: Evaluate with Enough Data

After the learning period, give the new budget level 2-4 additional weeks before evaluating. You need at least one full conversion cycle (for B2B, this might be 30-90 days) to see the real impact.

Minimum Viable Budgets by Platform

If you're considering a new channel or restarting a paused one, these are the practical minimums for meaningful data:

Related Reading

• When to Change Your Marketing Budget — the decision framework for when a budget change is actually warranted
• Diminishing Returns: When More Spend Stops Working — how to spot saturation before you waste budget
• How to Test Budget Changes — from simple checks to statistically significant experiments
• How to Reallocate Marketing Budget Using Attribution — the full reallocation process

Holly Henderson

Co-Founder, mbuzz

Holly Henderson is Co-Founder of mbuzz. With 10+ years in marketing including roles at Westpac, Avon, and Forebrite, she's obsessed with making measurement actually useful.

Harvard Extension School Forebrite Westpac Avon