Fix: Bid strategy or target changed while the campaign was still in Smart Bidding Learning

A Smart Bidding model that is mid-Learning is, by definition, not yet calibrated against your account's auction reality. When an operator makes a documented reset-trigger change — a bid strategy swap, a target shift, a budget jolt, a conversion-action edit, an audience or geo scope change — while the campaign still shows the Learning badge, Google does not pause the previous learning and resume it; it discards the partial model and starts a new Learning window from zero [1][2]. This is a different failure pattern from a campaign that overshoots 14 days for no apparent reason — that finding lives in [[fix-campaign-learning-stuck]]. This finding fires when the auditor detects a known-reset-trigger change in Change history whose timestamp falls inside an active Learning window.

If you just launched this campaign, this audit needs at least 14 days of historical data — return in 2 weeks.

Why this matters

Smart Bidding is a Bayesian model: it accumulates conversion signal continuously, weights recent data more heavily than older data, and uses query-level features to bootstrap predictions when conversion volume is thin [2]. Google's official guidance is that calibration "can take up to around 50 conversion events or 3 conversion cycles" [1]. The model needs that volume to converge whether the campaign is brand new OR was reset by a mid-flight edit — there is no partial credit for the conversions accumulated before the reset.

Changing a bid strategy setting while the Learning badge is still active is the worst-timed edit you can make. Two distinct costs stack:

1. Doubled calibration window. The campaign had not yet converged on its first Learning run; you now require a second full Learning window of equal length before performance is trustworthy. For a tCPA campaign with ~30 conv/30d throughput, this can mean 6-8 weeks total before steady-state instead of 3-4.
2. Compounded noise. Performance during Learning is exploratory — wider CPA/ROAS variance is expected and is not a signal of strategy health [4]. Operators who interpret first-window noise as "this target is wrong" and tighten the target during Learning lock in a feedback loop where every weekly review triggers another reset. Search Engine Land calls this "perpetual learning" — campaigns that never demonstrate real performance because each panic-tweak resets the clock [4].

The blast radius is forecastability, not delivery. Smart Bidding does not pause while in Learning — it still bids — but the bids are exploratory, not converged, so any conclusions drawn from in-window CPA/ROAS are statistically invalid input for further changes [2][4].

For background on what the Learning phase IS, conversion-volume thresholds, and the broader trigger taxonomy, see [[smart-bidding-learning-phase]]. For the distinct case where Learning runs past 14 days with no operator-initiated trigger, see [[fix-campaign-learning-stuck]].

How to verify

1. Open Tools → Change history. Filter to the affected campaign and a date range covering the last 30 days. Export to CSV if the campaign has high change density.
2. Open Campaigns → Bid strategies column for the same campaign. Note the current status badge AND any "status reason" surfaced under the badge. Cross-check the campaign-settings audit log (gear icon → Change history) for bid-strategy edits specifically.
3. Identify the Learning window start date — this is either the campaign launch date, the most recent bid-strategy-type change, or the most recent change that surfaced a "Learning" status-reason note. Anchor your investigation to this date.
4. In the Change history export, flag any change-type rows falling after the Learning window start where the change matches a documented reset trigger:
   • Bid strategy type changed (e.g., Maximize Conversions → tCPA)
   • Bid strategy target changed by >15% (tCPA or tROAS value)
   • Bid-strategy setting modified (portfolio settings, eligible bid range, max CPC ceiling on portfolio tCPA)
   • Campaign added to or removed from a portfolio strategy
   • Conversion-action goal added, removed, or re-prioritised for the campaign
   • Daily budget changed by >20% (vs trailing-7-day average)
   • Geographic targeting expanded/contracted (country-level or major region add/remove)
   • Audience-segment targeting added or removed (not bid-modifier-only edits)
5. If you find one or more reset-trigger rows inside the Learning window, this finding applies. Note the timestamp of the most recent reset trigger — the new Learning window resets from that timestamp, not the original launch.
6. Cross-check trailing-30d conversion volume in Goals → Conversions → Summary. If volume is below the strategy's floor (~30 conv/30d for tCPA, ~50 conv/30d for tROAS), the campaign will also show Learning Limited — a separate root cause that does not resolve through waiting [3].

How to fix

Total time: 10-20 minutes to diagnose, then 14-21 days of operator discipline. The fix is stop tweaking — the campaign cannot recover its prior learning state, so the only productive action is to give the new Learning window an uninterrupted run.

1. Freeze the campaign for the full new Learning window. From the most recent reset-trigger timestamp, wait at least 14 days before any further bid-strategy, target, budget, or scope change. Treat the campaign as read-only for that window [3][4].
2. Document the reset-trigger event in the campaign description or your change-log tool. This prevents a teammate from making another reset-trigger edit during the window and forms an audit trail for next-quarter review.
3. If the reset-trigger change was speculative ("let's see if a tighter target holds"), revert it immediately. A revert is itself a setting change and re-resets Learning — so revert ONCE and then freeze. Do not toggle back and forth.
4. If the reset-trigger change was intentional (e.g., a new conversion action you must measure against), commit and wait the full window. The cost of the reset is already paid; further edits only multiply it.
5. Disable any automated rules or scripts that touch bid strategy / target / budget during the freeze window. A common silent killer is a "lower tCPA by 5% if CPA >target for 3 days" rule that fires inside Learning and resets the model. Audit the campaign's Tools → Bulk actions → Rules tab and pause any rule that touches the affected campaign.
6. At day 14, re-check the Learning badge. If it has cleared to Eligible, judge performance against the trailing-14-day window. If it is still Learning without further operator changes, escalate to [[fix-campaign-learning-stuck]] — the volume floor (≥30 conv/30d tCPA, ≥50 conv/30d tROAS) is the more likely root cause from this point [3].
7. Build a "reset-trigger pre-flight checklist" for future edits. Before any bid-strategy, target, budget, conversion-action, audience, or geo edit, check the bid-strategy status FIRST. If the badge reads Learning, defer the edit unless it is a true emergency. If the badge reads Eligible, the edit is safe to make in isolation but will likely re-trigger Learning — plan a 14-day judgement freeze starting from the edit.

How to confirm the fix worked

Diagnostic checklist — run at day 14 AND day 21 after the freeze started

• The Smart Bidding status badge has moved from Learning to Eligible (or Active).
• Change history shows no reset-trigger edits (bid strategy, target >15%, budget >20%, conversion-action, scope) during the freeze window.
• Automated rules and scripts touching the campaign's bid strategy / target / budget are paused or scoped to exclude the campaign.
• Trailing-30d conversion count is at or above the strategy's volume floor — ≥30 for tCPA / Maximize Conversions, ≥50 for tROAS / Maximize Conversion Value [3].
• Daily spend variance over the most recent 7 days is within ~30% day-over-day (a converged model produces smoother delivery than an exploring one) [4].
• Trailing-14d CPA or ROAS is within ±15% of the pre-reset baseline — if not, the new Learning may have reached a different operating point, and the next judgement window starts now (do not edit immediately).

If the badge has cleared, no further reset triggers occurred, and the conversion-volume floor is met, the finding closes.

Edge cases — when this rule does NOT apply

• Campaign launched <14 days ago. First-Learning windows are expected; this finding only fires when an operator edit occurs INSIDE that window, not for the launch itself. The audit will surface as "insufficient data" until the temporal window is met.
• Creative-only edits. Adding/pausing responsive search ad assets, adding/pausing PMax asset-group creatives, swapping image assets — none reset Learning [2]. They appear in Change history and should be excluded from the trigger filter.
• Negative-keyword additions. Account-level or campaign-level negatives do not reset Learning [2]. PMax search-themes additions do not reset Learning either.
• Small bid adjustments on segments (device, time-of-day, audience modifier ≤±15%) do not reset Learning [3]. Larger modifier swings — especially crossing the 100% threshold — are practically equivalent to a setting change and will show up under this rule.
• Reactivating a paused campaign is treated as a "new strategy" event per Google's documentation [1] — if the campaign was paused mid-Learning and resumed, the badge will read Learning again and this finding does NOT apply (no operator-initiated reset trigger; the badge is a fresh launch).
• Conversion-action edits forced by external events (e.g., Consent Mode v2 enforcement broke an existing event, you had to migrate to a new conversion action) are still reset triggers — they fire this rule, but the right operator response is "commit and wait," not "revert."

Industry benchmarks

• Google's published calibration window: up to 50 conversion events OR 3 conversion cycles, whichever comes first [1]. For most Search/PMax campaigns this corresponds to 1-2 weeks; for low-volume B2B lead-gen it can stretch to 4+ weeks even without operator interference.
• Operator practice — "do nothing" window after a known reset: 14 days minimum, 21 days for low-volume / long-cycle accounts [3][4]. Optmyzr's analysis recommends 2-3 weeks of stability specifically after a budget change before judging Smart Bidding performance [3].
• Reset-trigger thresholds (operator consensus, NOT Google-published): budget changes >20% day-over-day, target changes >15% week-over-week — both round-numbered defaults that align with Google's own conservatism in published settings UX.

Related

• [[smart-bidding-learning-phase]] — what the Learning phase IS, conversion-volume thresholds, full taxonomy of reset triggers (read first if Learning concepts are new).
• [[fix-campaign-learning-stuck]] — sibling finding for the distinct case where Learning overshoots 14 days with no operator-initiated trigger (volume floor or model-state issue, not edit-driven).
• [[fix-bidding-target-change-over-15pct]] — narrower sibling finding focused specifically on the >15% tCPA/tROAS target-shift case as a standalone reset trigger.
• [[fix-budget-change-over-20pct-detected]] — sibling finding for the >20% budget-volatility detector (any Smart Bidding strategy, not PMax-specific).

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. Google Ads Help — How our bidding algorithms learn. https://support.google.com/google-ads/answer/10970825 (accessed 2026-05-27)
3. 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)
4. 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)