The bidder believes everything you tell it
Target CPA and target ROAS do not watch your account the way a person does. At every auction they estimate how likely this click is to convert, and what it is likely to be worth, using your conversion history plus the signals available at auction time: the query, the device, the location, the hour, the audience lists the user sits in. Then they bid accordingly.
The training data for all of that is the Conversions column. And to the bidding system, every recorded conversion is a fact. It has no concept of "that one was a bot", "that one fired twice" or "that one was a newsletter signup somebody promoted to primary three years ago". It optimises toward what you counted, not what you meant.
This is why polluted conversion data is not a reporting problem. A skewed report you can mentally discount. Bidding cannot discount anything: it acts on the pollution at auction time, with your budget. Bad conversions do not just make the account look wrong, they steer money toward more of themselves.
The three usual pollutions
1. Spam and fake leads counted as conversions
Bot traffic and AI-written fake enquiries submit forms exactly like real prospects do, so they fire your conversion tag exactly like real prospects do. I have written up what these fake leads look like and how to stop them at the form; the short version is that the convincing ones pass every surface check a tag can see.
2. Duplicate fires
The same real conversion counted more than once. The classic causes: a tag firing from both a GTM trigger and a hardcoded snippet someone forgot to remove, a thank-you page that recounts on refresh or revisit, or a lead action left on "every" counting when it should be "one". The account appears to convert better than it does, and CPA targets get set against a number that does not exist.
3. Vanity actions set as primary
Newsletter signups, brochure downloads, scroll milestones and click-throughs promoted to primary conversion actions, usually with good intentions and no follow-up. Once they are primary, the bidder treats them as the goal. They are cheap to buy, so it will happily buy lots of them.
What each one does to delivery
Spam creates a feedback loop. Junk submissions tend to arrive through the cheapest inventory: low-grade placements, broad queries, the traffic nobody else wants. When those clicks "convert", the bidder reads the junk as its best-performing segment and shifts spend toward it, which buys more junk, which strengthens the signal. The account settles into a state where platform CPA looks the best it ever has while the sales team's opinion of the leads gets steadily worse. That gap is the tell.
Duplicates warp the comparison between campaigns. Double counting is rarely uniform: one form double-fires, another does not. Target CPA then quietly favours whichever campaign inflates its own numbers hardest, and budget drains from the honest campaign into the broken one. Nothing in the interface looks wrong, because the interface is built on the same counts.
Vanity primaries dilute the goal. When a real enquiry and a scroll event are both primary, the system meets its target with the cheapest mix available, and the expensive action, the one the business actually wanted, falls out of the mix. The campaign hits its CPA and produces nothing anyone can invoice.
The fix, in order
First, audit the conversion actions themselves
List every conversion action in the account and decide what genuinely deserves to steer money. In Google Ads, primary actions feed the Conversions column and bidding; secondary actions are observation only. Most polluted accounts are fixed substantially by demoting soft actions to secondary and leaving one or two actions primary that map to real commercial intent.
One trap while you are in there: custom goals override the primary and secondary settings. A campaign optimising to a custom goal uses whatever actions that goal contains, whether they are primary or not, so an action you demoted can be pulled straight back into bidding by a goal nobody remembers building. Check what each campaign's goal actually contains, not just the labels on the actions.
Second, fix the mechanics
One source of truth per action: if the tag fires from GTM, remove the hardcoded copy. Lead actions count one per click, not every. Purchases carry a transaction ID so a refire does not recount. None of this is glamorous, all of it changes what the bidder learns. The container side of this, the broken triggers, consent gaps and stranded enhanced conversions data that starve tags in the first place, is its own post.
Third, stop the spam at the source and wall off the damage
Form protection is covered in the fake leads post. For the window where bad data already got in, Google Ads has data exclusions: you mark a date range and Smart Bidding ignores the conversion data in it. Two things to know before using them. Exclusions apply to click dates, so allow for conversion delay when choosing the range. And they are a tool for genuine outages and pollution windows, not routine housekeeping; expect some delivery wobble after applying one, because you have just changed the ground truth under the model.
Then leave it alone
Do not fix the data and change the targets in the same week. If you clean the conversions and drop the target CPA together, you will never know which change did what. Clean the inputs, hold everything else still, and give the system a couple of conversion cycles to relearn before judging it. On large accounts I script the audit trail for this rather than clicking through every action and goal, the same argument that pushed my restructure work and keyword research into tooling: repeatable beats heroic.
What cleaning your data will not do
Honesty about the limits. Cleaning conversion data will not rescue an account that barely converts: with a handful of real conversions a month, the bidder has too little to learn from whether the data is clean or dirty, and the fix there is consolidation or different goals, not hygiene. It will not make a weak offer strong: accurate measurement of a funnel people do not want is just a sharper picture of the same problem. And it will not pay off instantly, because the relearning period comes first and it is rarely comfortable.
What it does do is make every other decision in the account meaningful again. Bids, budgets, negatives, ad tests: all of them are downstream of whether "conversion" means anything.
Where to start
Pull up your conversion actions and answer one question per action: if the bidder bought a thousand more of exactly this, would the business be better off? Every action where the answer is no and the status is primary is budget leaking right now. Then look for the gap between what the platform celebrates and what sales sees, because that gap is usually this post in miniature.