Smart Bidding Only Works When Your Conversion Data Is Honest

Google's bidding algorithms are not smart in the way the name suggests. They are obedient. You name an event, and the machine spends your money finding more people who will do that event. Name the wrong one and it will find those people faster than any human buyer could, and your reporting will look excellent the entire time.

Most teams feed Google garbage and then blame the algorithm.

I have never fixed a badly performing automated campaign by changing the bid strategy. The fix is almost always upstream, in the definition of what got counted.

What counts as a conversion is a decision, not a setting

I sort conversions into three levels before I let anything optimize toward them.

Macro is the event you invoice: closed revenue, the job completed, the subscription paid for. Meso is qualified progress toward it, like a sales-accepted lead or a demo somebody actually held. Micro is engagement, so scroll depth, a video view, a newsletter signup.

The rule I run is to optimize on the highest level that clears the volume floor, and to step down only as far as volume forces. Most accounts do the reverse. They pick a micro event because it is abundant and because it ends the learning phase quickly, then act surprised when they have bought a great deal of engagement and very little revenue. Abundance is what makes micro events tempting and it is the same thing that makes them expensive. You reach significance on the wrong outcome fast.

There is a costlier version of this. A form fill counts as a conversion even when you cannot serve the person who filled it. If your business has any constraint on who it can deliver to, a coverage area, an inventory limit, a qualification bar, and you pass every raw form fill to the platform, you are paying an algorithm to get better at finding people you have to turn away. Reported cost per acquisition falls. Revenue does not move. Either filter those submissions before the event leaves your site, or import the fulfilled outcome back from your CRM and optimize on that.

That decision, which event leaves your site, is worth more than any bid strategy you will pick this quarter. If you want a second read on where your account is lying to you, that is the kind of thing I unpack on a 20-minute call.

The volume floor is a real number

Smart Bidding needs somewhere around 30 to 50 conversions per campaign per 30 days before its predictions carry weight. Below that it guesses confidently. Data-driven attribution is hungrier, closer to 500 conversions a month before the model can tell paths apart, which is why a lot of smaller companies are reading a data-driven report that is mostly noise in a nice interface.

Both numbers change what you do on Monday, because starvation is usually self-inflicted. When a campaign cannot reach thirty conversions in a month, the reflex is to add budget. Budget is rarely the constraint. Structure is. Accounts get split by geography, by product line, by whatever the org chart looks like, until every campaign is individually too small to learn anything.

Segment by data sufficiency, not by the map. If splitting a campaign would put either half under the floor, do not split it. If it is already split, merge back until each campaign clears thirty a month, and keep the granularity you wanted in the reporting layer, where it costs you nothing.

Honest also means nobody else is taking the credit

On The RealReal's paid social, the team could not prove the channel was driving revenue. Every time somebody saw a Facebook ad and later searched the brand name, Google brand search took last-touch credit for the sale. On paper, paid social looked like a rounding error and brand search looked like a genius.

Re-modelling the attribution would not have settled that, because both models were reading the same tainted click path. So we turned Facebook off in matched geographies and left it running everywhere else. Six weeks later the geos without it were down 35% on net-new customers. That was the answer, and 25% of the budget moved.

Geo-split tests want four to six weeks and matched markets. When two channels are fighting over the same conversion, it is the measurement I reach for first, because it reports what the spend is adding rather than what it is claiming. Add up the conversions every platform reports and you will land at three to five times your real sales. That is not a bug in their reporting. It is their business model.

Then there is the plain garbage

At GlacierGrid, then called Therma, about 60% of traffic was landing in direct. Nobody types a B2B URL that often. Direct is where untagged links go to be forgotten, and a bucket that size means you are running paid off a map with the roads rubbed out.

We tagged everything properly. LinkedIn turned out to be driving three times the demos anyone believed, so $15,000 a month moved off Google and onto it, and SQLs were up 600% over the following six months. Nothing about the market had changed. We had just been buying against a picture of it that was wrong in a specific, fixable way.

The rule I use now: if more than a third of your traffic lands in direct, that is your tagging, not your customers.

Getting there is unglamorous. At PartnerSlate we found fourteen different UTM variations describing a single campaign, because fourteen different people had built links by hand. Every one of them fragmented the same campaign's performance into a slice too small to act on. A URL builder that nobody can bypass took attribution accuracy from around 40% to 95%.

Campaign names go [quarter]_[initiative_type]_[brief_description], so q1_webinar_growth_masterclass. Lowercase, underscores, no spaces, no personal shorthand. It is a boring convention and it is load-bearing, because a naming scheme only works if it is the single easiest way for a busy person to build a link.

What I would do first

Pull conversions per campaign for the last 30 days and sort ascending. Everything under thirty is not learning, whatever the interface says about it. Merge until it clears.

Then open your conversion actions and ask what each one would be worth if a thousand of them arrived tomorrow. If the honest answer for the event you are optimizing toward is "not much," you have found the problem, and no bid strategy is going to fix it for you.

I am not against automation here. I run Smart Bidding on nearly everything, and it beats manual bidding at a scale no human can watch. It is just that automation multiplies whatever you hand it. Hand it an honest number and it compounds. Hand it a form fill you cannot fulfill and it will scale that with the same enthusiasm.

If you want to see where the money is actually leaking before you touch the account, the Pain Ladder Diagnostic is a decent place to start.