Skip to content
whitead

Fix: tCPA or tROAS target changed by more than 15% in the last 7 days

finding google ads updated 2026.05.28 9 min read
// contents · 8 sections
  1. Why this matters
  2. How to verify the issue
  3. How to fix it
  4. How to confirm the fix worked
  5. Diagnostic checklist — run within 14-21 days of the target change
  6. Edge cases
  7. Related rules and concepts
  8. Sources

You tightened the Target CPA from $40 to $30, or pushed tROAS from 400% to 520%, hoping the algorithm would absorb the new ceiling and keep delivering. Instead the campaign re-entered the learning phase: spend got erratic, CPA spiked, and your weekly review now contains noise rather than signal. Google's published guidance is to edit Smart Bidding targets in small increments — practitioner consensus puts the safe ceiling at ≤15% per week — because larger shifts force the bidding model to re-calibrate against a new objective and burn 7-14 days of exploration before the bidder converges again [1][2][3].

Why this matters

A Target CPA or Target ROAS is not a slider you can drag freely — it is the optimisation objective the Smart Bidding model is solving for. When you change the target by a material amount, the model treats the new value as a new objective and re-enters the learning phase to find the auction patterns that satisfy it [1][2]. Google's "How our bidding algorithms learn" reference lists target changes alongside bid-strategy switches, conversion-action edits and large budget swings as the events that re-trigger learning [1]. The percentage Google considers "material" is not published; the operating consensus across Optmyzr's 1,300+ account analysis and Search Engine Land's 2025 review of when AI bidding breaks is that target adjustments should stay within ±10% to ±15% of the current target per week, and that anything beyond that range should be treated as a deliberate reset rather than a tweak [2][3].

The blast radius is decision-making, not delivery. The strategy keeps bidding through the re-learning window, but the bids are exploratory rather than converged. Comparing CPA against a pre-change baseline during this window will systematically mislead — you read exploration noise as a sustained performance shift, panic-tweak the target again, and reset the 7-14 day clock a second time. This is the failure pattern behind "perpetual learning" — a campaign that never demonstrates steady-state performance because each weekly review triggers a corrective edit that restarts the model [3].

The cost compounds in two directions. Tightening the target too aggressively (e.g. cutting tCPA from $50 to $30 in one move) usually does not produce a $30 CPA — it produces throttled delivery, because the model can only honour the new target by excluding auctions where conversion probability is below the target's implied bid ceiling. Loosening the target too aggressively (e.g. pushing tROAS from 400% to 200% to "open up volume") usually does not produce proportionally more volume — it produces wasted spend on low-quality auctions while the model re-discovers which inventory still hits the looser ROAS. In both directions, the next 7-14 days of performance data is unsafe to read as steady-state.

How to verify the issue

  1. Open the affected campaign and check the bid strategy column. Confirm the strategy is tCPA, tROAS, or a portfolio variant of either. If the campaign is on Maximize Conversions or Maximize Conversion Value (no target), this rule does not apply — there is no target to change.
  2. Open Tools → Change history. Filter to the affected campaign and the last 7 days. Look for events of type "Bid strategy" → "Target CPA" or "Target ROAS" with a numeric change. Calculate the percentage delta: (new_target − old_target) / old_target × 100.
  3. If the absolute delta exceeds 15%, this finding fires. Cross-check whether multiple smaller edits stack — three sequential 6% tightenings inside a single week compound to roughly 18% and trip the same learning reset as one 18% move.
  4. Read the current bid strategy status badge. Learning confirms the model is mid re-calibration. Eligible means either the change has already stabilised (rare inside 7 days) or the model never registered the edit as material — in which case the finding is informational rather than urgent.
  5. Cross-check trailing-30-day conversion volume in Goals → Conversions → Summary. If the count is below the strategy's volume floor (~30 for tCPA, ~50 for tROAS), the re-learning window will be longer and noisier than for a volume-rich campaign — the same target change costs more when signal is scarce.

How to fix it

Total time: 5 minutes to diagnose, then 7-14 days of patience. The fix is almost never "change the target again" — most over-aggressive target moves resolve by either reverting (if speculative) or holding (if intentional) for the full re-learning window.

  1. Decide: revert, or commit and wait. If the target change was speculative ("let's see if tCPA can hold at 30% lower"), revert to the prior target now. Each subsequent change extends the re-learning window, so the cheapest path is one decisive move — either back to the prior target or forward to the new one — and then no further edits for 7-14 days.
  2. If reverting: open the campaign's bid strategy settings and restore the original target value. Save. Note the revert event in change history so the next audit has a clear timeline. Expect another 7-day learning window from the revert itself — the model treats the round-trip as two changes, not zero.
  3. If committing to the new target: freeze the campaign for at least 7 additional days from the change date. No further target edits, no budget changes greater than 20%, no conversion-action edits, no audience or geo shifts. Treat the campaign as read-only [4].
  4. Stage future target adjustments in ≤10-15% weekly increments. Going from tCPA $50 to tCPA $35 (a 30% cut) is two moves separated by a full learning window: $50 → $43 → wait 14 days → $43 → $35. The total elapsed time is longer than a single jump, but the bidder converges between each step and the final outcome is more stable.
  5. Anchor the target to observed performance, not aspiration. If the campaign's trailing-30-day CPA is $48, a tCPA of $40-$45 is a reasonable next step; a tCPA of $25 is aspirational and will throttle delivery regardless of how slowly you step toward it. Optmyzr's guidance: stay within ±10% of historical CPA as a baseline risk floor [2].
  6. If the campaign was near the volume floor before the change, the re-learning window may resolve into Learning Limited rather than Eligible. Address volume separately — see the campaign-learning-stuck article and the minimum conversions for auto-bidding article — before re-tightening the target.

How to confirm the fix worked

Diagnostic checklist — run within 14-21 days of the target change

  • The Smart Bidding status badge has moved from Learning back to Eligible.
  • No further target changes have been made since the original event.
  • Daily spend has settled — no day-over-day swings greater than ~30% in the last 7 days.
  • Trailing-30-day CPA or ROAS is within ±15% of the new target (if you committed to the change) or back within ±10% of the pre-change baseline (if you reverted).
  • Change history shows no operator-initiated target edits in the last 14 days.
  • If the campaign is on tROAS, value reporting is still firing correctly — a target reset does not fix a missing or broken conversion-value signal.

If the badge has cleared and the metrics are within tolerance for the chosen target, the finding closes. If the badge is still Learning after 21 days with no further changes, escalate to the campaign-learning-stuck finding — at that point the bottleneck is volume or signal quality, not the target change.

Edge cases

If you just changed your target, wait 7 days for Smart Bidding to stabilize before re-auditing. The detector compares the most recent target value against the value 7 days prior; a finding inside the first 7 days is expected behaviour and not actionable yet.

If your target moved because Google's automated suggestion was applied, treat it the same as a manual edit. Auto-applied suggestions count as operator changes from the model's perspective and trigger the same re-learning window.

If the campaign uses a portfolio Target CPA strategy across multiple campaigns, target changes affect every campaign in the portfolio simultaneously. The re-learning window applies to the portfolio as a whole, and the volume gate is the pooled conversion count across all campaigns in the strategy — not per-campaign.

If the change was a one-way ratchet during a seasonal event (e.g. tightening tCPA going into Q1 after Black Friday), the re-learning window is unavoidable. Plan it deliberately: schedule the change at the start of a low-volatility week and freeze edits for the full 14-day window.

Sources

  1. Google Ads Help — Duration of the learning period for campaigns and what affects it. https://support.google.com/google-ads/answer/13020501 (accessed 2026-05-27)
  2. Optmyzr — The Impact of PPC Bidding Strategies on Google Ads Performance. https://www.optmyzr.com/blog/impact-of-ppc-bidding-strategies/ (accessed 2026-05-27)
  3. Search Engine Land — When Google's AI bidding breaks – and how to take control. https://searchengineland.com/google-ai-bidding-breaks-take-control-466251 (accessed 2026-05-27)
  4. Google Ads Help — Edit your target CPA. https://support.google.com/google-ads/answer/6336096 (accessed 2026-05-27)
// was this useful?
// anonymous · no personal data stored
← Back to Wiki