Performance Max campaigns in B2B SaaS accounts commonly drift into a low- quality conversion trap: the algorithm is generating high form fill volume, CPAs look acceptable on paper, but the leads being produced are overwhelmingly unqualified. When you map those form fills to pipeline, almost nothing converts to SQL or closed-won. The standard response is to rebuild the campaign from scratch, which is almost always the wrong move when a significant learning history exists.
This post covers the retraining approach — changing the conversion signal the algorithm is optimizing toward without triggering a learning reset. It is the harder path to set up correctly but produces better outcomes and avoids the 4-6 week learning phase penalty you pay every time you rebuild. For the go/no-go decision on whether PMax belongs in your account at all, see the guide to Performance Max for B2B SaaS.
Diagnosing micro-conversion optimization
Before attempting to retrain a PMax campaign, confirm that the problem is actually conversion signal quality and not something else. Run these three checks. First, pull your conversion action breakdown from the campaign view (Columns → Conversions → All conversions by conversion action). Check what percentage of conversions the algorithm is optimizing toward are form fills versus downstream events. If form fills are above 80% of primary conversion volume, you have a signal quality problem.
Second, check your PMax campaign's asset group performance. If the highest- performing asset groups (rated "Best") are using creative combinations with broader, awareness-oriented messaging rather than your conversion-optimized bottom-funnel copy, the campaign has migrated toward upper-funnel audiences. This is a symptom of the algorithm finding form fills on awareness placements rather than buyers on Search.
Third, look at the search term report for your PMax campaigns (available under Insights → Search terms). If the queries triggering your PMax ads are broad informational terms — "what is [category]," competitor name queries, generic category terms — rather than high-intent evaluation queries, the campaign's audience composition has drifted away from active buyers. All three signals together confirm that the campaign is in a micro-conversion trap and needs to be retrained.
The retraining approach: shifting signal without resetting
The retraining approach preserves the campaign's learning history while progressively shifting what the algorithm optimizes toward. It has four components that should be implemented in sequence, not simultaneously — doing everything at once is more likely to trigger a learning reset than a controlled adjustment.
Step 1: Set micro-conversions to secondary. In your conversion settings, change the form fill conversion action from Primary to Secondary. Secondary conversion actions still record data and are visible in reporting, but they do not influence Smart Bidding. This removes the signal the algorithm has been over-optimizing toward without deleting the historical data. Do this alone and wait 48-72 hours before making any other changes.
Step 2: Set a higher-quality primary event. If you have offline conversion data flowing in from your CRM — SQL-stage, demo-booked, or qualified lead — set that as the new primary conversion action. If you do not have offline conversion tracking set up, this is the prerequisite step: implement your offline conversion stack first and let it collect 4-6 weeks of data before attempting to retrain PMax. The algorithm cannot shift toward a signal that does not exist.
Step 3: Adjust conversion value weighting. If your new primary conversion event is too low-volume for the algorithm to learn from quickly (under 30 events per month per campaign), implement a value ladder: assign different monetary values to different conversion stages (e.g., form fill = $5, MQL = $50, SQL = $400, closed-won = $2,000). Set all stages as primary with values attached and switch your campaign bidding to Target ROAS or Maximize Conversion Value. This gives the algorithm a multi-stage signal with enough volume to learn from while weighting optimization heavily toward higher-value events.
Step 4: Tighten audience signals. Simultaneously with the conversion action changes, update your PMax audience signals to include your CRM customer list (closed-won accounts as customer match), your SQL remarketing list, and your pricing/demo page visitor segments. Remove broad interest-based audience signals if any were set — these tend to push PMax toward awareness placements that generate high form fill volume but low SQL rates.
What to expect during the retraining period
The first two weeks after changing primary conversion actions will show conversion volume dropping and CPAs rising. This is expected — the algorithm is recalibrating from optimizing toward a high-volume signal (form fills) to a lower-volume signal (SQLs or qualified demos), and it needs time to find the traffic patterns that produce the new target outcome.
Do not adjust bids, budgets, or targets during the first four weeks of retraining. Changes during the learning phase extend the instability period rather than shortening it. The correct action if you see CPA rising is to wait, not to intervene. If after six weeks the campaign has not stabilized to within 20% of your target CPA, the problem is insufficient conversion volume for the new signal — either the primary event is too rare, or offline conversion imports are not uploading consistently.
By week 6-8, you should see the campaign's asset group composition starting to shift: asset groups with bottom-funnel, conversion-oriented creative will show improving performance ratings, and the query mix in the search term report will shift toward higher-intent evaluation terms. Both are leading indicators that the retraining is working before the CPA metric itself visibly improves.
Using negative keywords and placement exclusions during retraining
Negative keywords in PMax campaigns operate at the account level (not the campaign level), which means any negative keyword list you apply affects all campaigns including PMax. During a retraining period, it can help to add account-level negatives for the query types that have been driving the highest volume of unqualified form fills — typically generic category terms ("what is [software category]"), educational queries ("how to [task]"), and non-commercial informational queries.
Placement exclusions are available for display network placements within PMax. Run a placement report (Insights → Placements) and identify any URLs or app categories generating high impressions with zero conversions over the last 30 days. Add those to your placement exclusion list. Be conservative with exclusions during retraining — removing placements changes the traffic mix the algorithm is learning from and can interfere with the signal shift you are trying to create. Focus exclusions on clear junk placements (mobile game apps, content farms, non-English content if you are targeting English) rather than any placement that has not yet converted.
When rebuilding is actually the right choice
Retraining works when the existing campaign has meaningful learning history and the root issue is conversion signal quality. Rebuilding is the right choice in three specific scenarios: the campaign was launched less than 3 months ago (insufficient learning history to protect), the campaign's structure is fundamentally misconfigured (e.g., all products and audiences in a single asset group with no segmentation), or the conversion data in the campaign history is corrupted by a major tracking error that imported thousands of false conversions.
If you do rebuild, do not start the new campaign with form fills as the primary conversion event — that recreates the same problem. Start with either a value ladder (if you have the offline data ready) or with a high-quality micro-conversion like a specific page view (demo page, pricing page) that is closer to buyer intent than a raw form fill. The PMax audit guide covers how to structure a new campaign to avoid the micro-conversion trap from the start.
Measuring retraining success
The success metric for PMax retraining is not CPA or form fill volume — it is cost per SQL or cost per qualified demo. Pull this number weekly for the first 8 weeks of retraining by joining your Google Ads cost data (from the Google Ads UI or via the Ads Reporting API) with your CRM SQL data. Calculate total Google Ads spend attributed to the PMax campaign divided by the number of contacts that were created from PMax traffic and reached SQL stage within the same 30-day window.
If cost per SQL is improving week-over-week while form fill volume declines, retraining is working — the algorithm is finding buyers rather than form fillers. If cost per SQL is flat or rising while form fill volume declines, the algorithm has not yet found the new optimization pattern and you should check offline conversion import consistency. If form fill volume is growing again after week 4, the micro-conversion signal has re-entered the primary slot — audit your conversion settings to confirm the primary/secondary designation is still correct.
Most B2B SaaS PMax campaigns that go through a proper retraining show a 50-70% reduction in cost per SQL over 8-12 weeks. The temporary CPA spike during weeks 1-4 is the price of that improvement. The framework for optimizing toward SQLs rather than leads covers the broader account-level approach to conversion signal quality that makes PMax retraining stick long-term.