Google Ads Manual CPC vs Target CPA, 60/40 experiment: $416 vs $152 per lead (p=0.009)

On the afternoon of 24 August, someone opened a mid-sized Google Ads account and, in the space of ten minutes, changed the bid strategy on four battery campaigns from Maximise Conversions with a $200 target to Manual CPC. The next day a panel campaign followed. Eleven new campaigns were created that week, all on Manual CPC. Dozens of exact-match Target CPA campaigns were paused. The brand campaign was throttled to a single ad group. Every lead form extension link in the account was paused.

The change history reads like a demolition.

The stated rationale, as far as we could reconstruct it, was that the automated bidding had been "learning from the wrong conversions", which, as it happens, was true, and is covered elsewhere in this series. The response to that diagnosis was to remove the automation entirely and rebuild by hand.

What happened next

Nine days before the change against nine days after, every campaign in the account including the paused ones:

The account spent $13,465 more and produced 144 fewer leads. Cost per lead more than doubled.

By category, under Manual CPC:

For context, the account's lifetime blended cost per lead is $128, and its non-brand search figure is $152.

The confound

That before/after comparison is real, and it's damning, and it does not prove that Manual CPC caused it. Too much changed at once.

Any one of those could produce a large part of the effect. The restructure is a natural experiment with five simultaneous treatments and no control group. You can see that something went badly wrong. You cannot see which thing.

This is the central problem with restructures as a way of learning anything. They're expensive, they're irreversible in practice, and they don't isolate variables.

So we ran the actual experiment

One battery campaign, the highest-volume one. Control arm stayed on Manual CPC. Treatment arm went to Maximise Conversions with a $120 target. Sixty-forty split, control-heavy, because the campaign couldn't afford to lose more than it already had.

That's a dramatic result and it's too dramatic: the account's conversion counting was inflated at the time, for reasons the previous post covers. So we recomputed with only the conversion actions we trusted: the GA4 web form, phone calls, and Google-hosted lead form submissions.

Still significant at the 1% level, with the inflation removed. Google's own experiment readout agreed:

The experiment ran for seven days. It cost a fraction of what the restructure cost, it touched one campaign instead of sixteen, and it produced an answer with a p-value attached.

What was applied

The treatment settings were applied to the base campaign on 4 September.

Two things worth noting about that configuration.

The target is aggressive. The campaign's trailing-seven-day cost per lead before the switch was $253. A $150 target is 41% below that. Google's own guidance is to move targets in steps of 20–30%; steeper cuts usually cost impression share. The argument for $150 is that the treatment arm demonstrably ran at $152 deduplicated on these same keywords. The argument against is that a 41% step is a 41% step. The decision was to hold at $150 and watch impression share rather than tighten further.

Thirty-three conversions in thirty days is exactly the threshold. Smart bidding wants around 30. This campaign has 33. One bad week puts it under, and if the new target throttles volume, it may put itself under. That is the strongest reason not to touch the target again for a fortnight.

What the experiment does not prove

Battery. Only battery. The panel category ($28,331 at $689 per lead), was not tested, and a separate panel experiment that was running told a different story.

Manual CPC was beating Target CPA on panel, at least on the campaign it was tested on. That experiment had its own problems. It was sitting on a $31-a-day campaign when the $154-a-day metro campaign was the obvious host, and it had eleven clicks in three weeks, but the direction is a warning. Battery and panel are different categories with different query mixes, and the battery result does not transfer.

The correct next step for panel is its own experiment, on the right campaign. Not an extrapolation.

What the nine days actually cost

Against that, the experiment: one campaign, 40% of its traffic, seven days, and a decision with p = 0.009.

What the campaign looked like day by day

The base campaign through the transition: Manual CPC until 4 September, then Maximise Conversions with the $150 target:

Seven-day trailing: $5,815 spend, 23 conversions, $253 per lead. That's the figure the $150 target is being asked to beat. The daily swing ($142 one day, no conversions at all the day before), is a reminder of how little a single day tells you, and why the target shouldn't be adjusted on anything shorter than two weeks of data.

The broad-match risk that Manual CPC was hiding

One thing the switch back to automated bidding exposes. The campaign carried twelve broad-match keywords across two ad groups:

Under Manual CPC these were bid-capped by hand and recorded almost no spend, the search terms report over ten days showed broad match at essentially zero. Under a Target CPA strategy, the algorithm bids on whatever it believes will convert, and broad match on solar or rebate in a battery campaign is a very wide door. Nothing had come through it by day two, but it's the first place to look if the campaign's search terms start drifting after the switch.

The general version: Manual CPC hides structural problems by letting you cap them by hand. Automated bidding will find them. Audit the match types before you switch, not after.

The lesson

A bid strategy change is a hypothesis. The hypothesis "Manual CPC will outperform Target CPA in this account" was entirely reasonable to hold: the automation was learning from contaminated conversions, and manual control is the obvious response to that.

But a hypothesis is tested, not deployed. Google Ads has a built-in experiment framework that clones a campaign, splits its traffic, and reports a winner with a confidence interval. It exists precisely so that nobody has to restructure sixteen campaigns to find out whether a bidding idea works.

Use the experiment. If it wins, apply it. If it loses, you've spent 40% of one campaign's budget for a week and learned something. Either way the other fifteen campaigns kept running.