Smart Bidding looks like magic because nobody shows you the machinery. This lesson opens it just enough for you to know what it is doing, when to trust it and what breaks it. You don't need to understand the statistics; you do need the mental model.
What it does in each auction
For every search that could trigger your ad, the algorithm estimates the probability of conversion (and, if it bids for value, the expected value) from the signals available at that moment:
- Device, operating system, browser.
- Precise location and time of day / day of the week.
- The exact search term and its intent.
- Whether the person has already visited your site, is a customer, is on a list.
- Behavioural history of similar users.
- The ad and the page it is about to show.
With that probability it works out the bid: if you want a CPA of €30 and it estimates a 10% chance of conversion, it bids around €3 for that click; if it estimates 1%, €0.30. Millions of times a day. No person can bid like that; that is why manual bidding has lost.
Where the estimates come from: your conversions
The algorithm learns from the conversions you teach it. Every recorded conversion tells it «this combination of signals ended in a sale». Two consequences you cannot dodge:
- With no conversions it does not learn. That is why strategies with a target ask for volume: the official benchmark is around 50 conversions for a strategy to leave learning and stabilise; with fewer it works, but with more variance.
- It learns what you give it, good or bad. If you count a click on «show phone number» as a conversion, it will optimise towards people who click «show phone number». Tracking (Module 6) is the steering wheel of Smart Bidding.
The «Learning» status
It appears when the strategy is calibrating, for three reasons:
- A new strategy (or one just reactivated).
- A change of target (CPA/ROAS) or a significant budget change.
- A change of composition: campaigns, ad groups or keywords added to or removed from the strategy.
How long it lasts depends on how many conversions you get and how long a conversion takes from the click (the conversion cycle: minutes in an ecommerce, weeks in a complex sale). Official guidance: up to 50 conversions or 3 cycles. During learning, performance is unstable; judging it then is judging an apprentice on their first day.
Why changes restart it
Every relevant change forces a recalibration. Changing the target CPA by 30% on a Monday and another 30% on the Thursday keeps the campaign in permanent learning: it never gets to perform. This is where the most important rule of the module comes from, developed in lesson 5: small, spaced-out, measured changes.
What counts as a relevant change: the target (CPA/ROAS), a big jump in budget, the strategy, the campaign's composition, mass changes to keywords or ads, and pausing/reactivating the campaign.
How to live with it
- Give it volume: if a campaign doesn't reach 30 conversions a month, group it into a portfolio with others or use «Maximize conversions» with no target until it grows.
- Give it clean data: one clear primary conversion action; the secondary ones set to observation only.
- Give it time: at least 2-3 weeks (or a full conversion cycle) before you evaluate a change.
- Give it context when you are the one who knows: for sales periods or one-off events there are seasonality adjustments (you warn it that the conversion rate will change for a few days) and for broken data there are data exclusions (you tell it to ignore a period in which tracking failed).
💡 Ninja trick: the conversion you teach it is everything. In lead generation accounts, our L1/L2 methodology separates browsing conversions (viewing a price, tapping the phone number: L1) from real contact ones (form submitted, call: L2) and requires L1 to weigh less than half of the total value — otherwise Smart Bidding learns to bring in browsers. And Lead Rating goes one step further: it sends the real value of each lead back to Google so it learns from the good ones, not from all of them.
What you should remember
- Smart Bidding estimates the probability of conversion per auction with signals you could never process, and bids in proportion.
- It learns from your conversions: with no volume it doesn't learn; with dirty data it learns badly.
- «Learning» = calibrating: up to ~50 conversions / 3 cycles. Don't judge during learning.
- Every relevant change restarts it: few changes, small, spaced out.