Why your Google Ads "performance collapse" is usually a measurement problem
The story arrives the same way every time. A Google Ads account that performed well for months — sometimes years — starts drifting: CPAs creep up, conversion volume sags, tCPA campaigns throttle their own spend. The team responds the way the industry has trained them to: new campaign structures, fresh keywords, an agency review. Sometimes a full rebuild.
In our audits, the majority of these "collapses" trace back to something far less dramatic: the conversion signal changed, and nobody noticed. The campaigns were fine. The data feeding them was not.
The pattern
Modern Google Ads is a closed loop: your conversion tag reports outcomes, Smart Bidding learns from those outcomes, and bids follow the learning. Anything that distorts the reporting side of that loop distorts the bidding side within days. The usual suspects, roughly in order of frequency:
Silent tag breakage. A site migration, a consent-banner update, a tag manager "cleanup" — and the purchase tag fires on 60% of the transactions it used to. Reported CPA doubles; actual CPA may be unchanged. Smart Bidding, seeing fewer conversions, retreats from auctions it should be winning.
Consent-related signal loss. After Consent Mode changes or a stricter banner, EU conversions drop out of the observed data. Without modelling configured correctly, the account effectively goes blind in some of its markets while looking healthy in aggregate.
Duplication in reverse. Counterintuitively, fixing a duplication problem produces the same symptoms: if the account was double-counting for a year and someone corrects it, reported CPA "worsens" overnight, targets that were calibrated to inflated numbers strangle delivery, and everyone blames the market.
Conversion action drift. Someone adds a new conversion action, sets it to Primary, and suddenly the bidder is optimising toward newsletter signups blended with purchases. This one hides for months because total conversions look great.
Why bidding amplifies measurement faults
A human media buyer with a broken report makes bad decisions slowly. An automated bidder makes them at auction speed, thousands of times an hour, and then compounds them: less spend where signal was lost → fewer observed conversions there → still less spend. The feedback loop that makes Smart Bidding powerful when fed clean data makes it destructive when fed corrupted data. This is why measurement faults present as performance collapses rather than as reporting errors — the bidder converts one into the other.
It is also why the standard responses fail. Restructuring campaigns resets learning on top of the same broken signal. Adding keywords feeds the same distorted loop. Only fixing the signal fixes the loop.
The five checks to run first
Before touching a single campaign setting, spend a day on these:
1. Reconcile conversions against source of truth. Pull thirty days of conversions from Google Ads and thirty days of actual transactions from your payment provider or CRM, matched by day. A healthy account tracks within a stable ratio. A ratio that shifted on a specific date is your smoking gun — and the date usually corresponds to a deploy.
2. Check conversion action settings history. The change history filtered to conversion changes takes five minutes and catches action drift, counting changes, and window edits that nobody remembers making.
3. Inspect consent behaviour by market. Compare conversion rates for EU versus non-EU traffic before and after any banner or Consent Mode change. A divergence that opens on the change date is signal loss, not market weakness.
4. Look for duplicates directly. Export conversions with order IDs where available and count repeats. No order IDs in your conversion data? That is itself a finding — see our deduplication article for the fix.
5. Trace one conversion end to end. Make a test purchase and follow it: tag fire, dataLayer contents, consent state, Google Ads registration, value and currency. Twenty minutes, and it surfaces the faults dashboards never show.
Fixing it in the right order
When you find the fault — and if the collapse was sudden, you usually will — resist the urge to fix measurement and restructure simultaneously. Restore the signal first, ideally moving revenue events server-side where they stop depending on browser conditions. Then re-base your targets on corrected data: the "real" CPA after a fix is often materially different from what the account reported before, and old targets calibrated to bad data will strangle delivery against good data.
Give bid strategies a deliberate re-learning window — two to six weeks depending on volume — before judging anything. Only then, with a trustworthy baseline, does structural work make sense.
The general lesson costs nothing to apply: when performance moves sharply and the market didn't, suspect the measurement before the media. It is the cheaper hypothesis to test, and it is usually the correct one.