Smart Bidding: when automation turns on you

By Miguel Maio · Published 27 October 2026

Smart Bidding doesn't fail because it's dumb. It fails because it's starved. And the one feeding it the wrong food is usually you - often on the advice of people who should know better. The promise is real when there's data, and the disaster is real when there isn't. The difference between the two lives in what Google says the algorithm needs and what it actually needs.

The strongest version of the Smart Bidding promise

Smart Bidding adjusts the bid in every single auction in real time, using signals you'd never catch in time: device, hour, audience list, search history, and the way all of it combines in one user. No human manager does this auction by auction, every day, without sleep. In its strongest form, the promise is real: you gain the granularity the manual hand kept losing.

And it's true. In accounts with volume and clean measurement, Smart Bidding beats manual bidding comfortably - not by magic, but by processing signals that slip past you. Optmyzr, which makes its living optimising other people's accounts and is therefore biased toward automation, reports consistent gains above certain conversion volumes. I take the figure with care given the source, but the direction matches what I see on the ground. When there's data, the algorithm sets the bid better than I do. Full stop. The problem was never the machine's intelligence. It's the diet.

The "minimum conversions" for Smart Bidding: the letter and the spirit

The letter: Google writes, in its own best-practices document, that the minimum number of conversions to use Target CPA and Maximize conversions is zero. The spirit is something else. The same Google, two lines down, warns that "you must be tracking conversions" - and Target ROAS, which has to learn values, already demands 15 conversions in the last 30 days (Google, Smart Bidding best practices).

The obvious question: if zero is enough for Maximize conversions, why does its cousin need 15? Because "minimum to switch on" and "minimum to work" are different things, and Google only tells you the first. Zero is what the button accepts, not what the algorithm needs in order to stop spending blind. It's like saying the minimum to drive is holding a licence: true, and useless to someone who has never sat behind the wheel. This is where automation turns on you - not because it lies, but because it lets you do something it knows will go badly, and calls it "supported".

The autopsy: from ~100 leads a day to 2 or 3

Real case, anonymised client, pure lead generation. A consultant from Google itself told the client to move a brand-new campaign - no conversion history in GA4, no consolidated audiences - to Maximize conversions. The client, with Google telling him to do it, did. The result: from around 100 leads a day to 2 or 3. For close to a week. Exactly the learning period window.

The algorithm did what it was told: it started learning from zero, in an auction where it had no pattern to tell who converts from who clicks and vanishes. It spent the budget exploring. And exploration, with no history, is indistinguishable from waste. I recovered by restoring my earlier setup from a Google Ads Editor backup - which I always take before letting anyone touch an account. Without that backup, the rebuild would have taken weeks instead of an afternoon.

The lesson isn't "Google gave bad advice", though it did. It's more uncomfortable: the interface let a structural change through - new campaign, new strategy, new goal - that triggers the learning period for every reason we're about to see, with no warning proportional to the risk. The letter ("minimum zero") covered for the spirit (this will break).

The layer nobody sees: modeled conversions in Smart Bidding

Here's the part almost nobody measures, and it decides whether your data diet is real or invented. A large share of the conversions Smart Bidding learns from, in consent markets like the EU, are not observed - they're modeled. Estimated by a model when the user refused cookies. The algorithm optimises toward a number that is, in part, a prediction about another prediction.

And modeling has hard thresholds, which Google publishes but almost no one reads. In Google Ads, modeled conversion backfill requires 700 ad clicks over 7 days, per country and domain grouping (Google Ads Help). In GA4, behavioral modeling for consent mode asks for at least 1,000 events a day with consent denied for 7 days, and at least 1,000 daily users with consent granted on 7 of the last 28 days (GA4 Help). Below these volumes you don't get worse modeling - you get none. You get the consent gap, raw, and the algorithm learning from half the truth.

One detail separates people who understand this from people who recite slides: modeled data does not enter the BigQuery export. It lives only in the reports and the Data API (GA4 Help). Anyone auditing through BigQuery, as I do when I want the raw number, sees fewer conversions and sometimes fewer users than the GA4 report. It's not a bug in either. One shows the modeled, the other the observed. If you don't know which you're looking at, you're comparing two languages and calling it a discrepancy. Automation eats the modeled number. If your volume doesn't clear the threshold, you're feeding the algorithm an estimate that doesn't even exist.

Fragmented data: the same trap, from the other side

Data starvation has a twin: data that exists but is broken, and therefore lies. Same law, opposite direction. Real case: Display campaigns segmented by device, one for desktop and one for mobile. Universal Analytics didn't attribute cross-device conversions - the user saw it on the phone and converted on the laptop, and UA counted zero for the mobile campaign. The apparent CPA inflated. Automation, reading that inflated CPA, cut exactly the campaign that was driving demand.

I fixed it by ending the two device segments and letting the signal consolidate. The CPA "dropped" - it had never really been that high, it was just miscounted. The irony is that Google today recommends precisely this: consolidate, don't fragment, to give the model the whole signal. A years-old field lesson became the official best practice, because the principle never changed: the algorithm decides only as well as your worst data lets it. Fragmenting by device, by campaign, by account or by duplicated conversion is all the same thing. You give it pieces and it makes whole decisions about pieces.

What I do before switching bid automation on

Before moving anything to Smart Bidding, I run a short, non-negotiable list - with the why behind each point, which is what the guides that just say "turn it on and wait" leave out. I automate detection, never the decision to switch on.

• Real conversion history on the campaign, not zero with Google's blessing. Without a pattern, the learning period is blind exploration and the budget burns.
• Consolidated audiences and signals, not a newborn campaign with no list - the algorithm needs to know who it has already won.
• Measurement clearing the modeling threshold, or the explicit assumption that I'll run below it and what that means.
• A Google Ads Editor backup before any structural change, because the change triggers the learning period and I want to be able to roll back in an afternoon, not a week.
• One change at a time, so the learning period has a single cause and I know what moved.

It isn't distrust of the machine. It's feeding it before I ask it to run. It's the same principle I apply in Google Ads scripts: automate detection, not strategy.

The cost of switching on too early

Automation doesn't punish you when it fails. It punishes you a week later, when the learning period has already burned budget that isn't coming back and you're explaining to the client why leads collapsed in the month you "modernised" the account. The algorithm is good - that was never the question. The question is whether you gave it what it needs, or only what Google says is enough.

Sources

• Google - Smart Bidding best practices (PDF)
• Google Ads Help - Learning period duration
• Google Ads Help - Conversions and consent mode
• GA4 Help - Behavioral modeling (consent mode)
• GA4 Help - Reporting surfaces comparison

Frequently asked questions