Every few months a client asks us some version of the same question: should we let Google's automation run the bidding, or keep our hands on the wheel? The question is fair, but it is usually asked too late. By the time someone is comparing manual CPC against Target ROAS, the decision has already been made by whatever state the account measurement is in. Automated bidding is a model that learns from the conversion data you feed it. If that data is thin, delayed, or mislabelled, the model will optimise faithfully towards the wrong thing, and the post-mortem will blame the algorithm rather than the plumbing.
So our team stopped treating this as a bidding decision and started treating it as a readiness check. Below is roughly what we run through before we change a bid strategy on an account we have inherited.
The measurement layer is the training data
Google is explicit about what its bidding systems use. Smart Bidding sets bids at auction time and factors in signals such as device, location, time of day, browser, operating system and language, along with combinations of those signals that individually would not show up in a manual bid adjustment. You can read the official description in Google's own documentation on Smart Bidding. What that documentation cannot tell you is whether the conversions you are sending back are worth optimising against.
This is where most of our early hours go. We open the conversion actions list and ask three unglamorous questions. Which actions are set to "primary" and therefore actually used for bidding? Are any of them counting the same event twice, for example a thank-you page fire alongside a form-submit event? And is anything in there that should not be a conversion at all?
That last one causes more damage than people expect. Accounts we take over frequently have a phone-click event, a newsletter signup and a PDF download all sitting in the primary bucket next to actual purchases. The model does exactly what it was told: it goes and finds cheap PDF downloads. Nobody wrote a bad strategy. Someone just left a checkbox in the wrong state eighteen months ago.
Volume, and the patience it demands
The second question is whether there is enough signal for the model to learn anything at all. There is no universal threshold we trust, and we are suspicious of anyone quoting one with confidence, because it depends on how variable your conversion rate is and how long your consideration cycle runs. What we do instead is look at conversion lag directly.
Google reports this. Segment the campaigns view by conversions and days to conversion, and you get a distribution rather than a guess. The practical guidance in Google's conversion lag reporting documentation is to exclude recent days from any assessment where fewer than 90% of conversions have been reported. For a business selling a low-consideration consumable, that might mean ignoring the last two days. For a B2B service with a multi-week sales cycle, it can mean that nothing you look at inside a month tells you anything reliable.
We have had to walk clients back from strategy changes made on four days of data more times than we would like. The account was not underperforming. It had not finished reporting.
Deciding what the conversion is worth
The bigger unlock, in our experience, is not switching from manual to automated. It is moving from counting conversions to valuing them. A lead form on a page for a small maintenance job and a lead form on a page for an annual contract are the same event to a system that only counts. Once you pass differentiated values back, the bidding target stops being "more of these" and becomes "more of the profitable ones".
This is where the work leaves the ads interface and turns into a data problem. Someone has to decide what a lead is worth, and that decision has to survive contact with the finance team. In practice we usually start with something crude and defensible: close rate by product line multiplied by average order value, computed from the client's own CRM export, refreshed quarterly. Crude and directionally correct beats precise and imaginary.
As an illustrative example of the shape this takes, and not a description of a specific client: a services company sending a flat value of 1 for every form submission will see its campaigns drift towards whichever service line produces the most enquiries, which is almost always the cheapest one. Feeding back a value derived from historical close rates changes which auctions the system fights for, without anyone touching a bid.
Test it as a change, not as a belief
When we do make the switch, we run it as an experiment rather than a flip. Google's own guidance is to run a bid strategy experiment for at least four to six weeks, and longer where conversion delay is long, which matches what we have seen: the first two weeks tell you almost nothing except that the learning period is uncomfortable.
Two disciplines matter here more than the setup itself. Freeze other changes for the duration, because a budget adjustment or a new landing page halfway through means you have tested nothing. And write down beforehand what result would make you revert. Deciding the success criteria after seeing the numbers is not evaluation, it is storytelling.
Where we still say no
We do not recommend automated bidding on every account. Campaigns with genuinely sparse conversion volume, accounts undergoing a pricing or positioning overhaul mid-flight, and short seasonal pushes that will end before any learning period completes are all cases where we have kept manual control and been glad we did. There is also a category we treat separately: accounts where the client cannot tolerate the variance. A model that improves average cost per acquisition while widening the daily spread is a bad trade for a business managing week-to-week cash flow, however good the monthly chart looks.
The pattern generalises past Google Ads, which is why we keep writing about it. Most of the value in applying machine learning to an operational process comes from work that happens before the model: defining the outcome precisely, instrumenting it honestly, and agreeing how you will know whether it worked. The model itself is usually the least negotiable part of the system and the least interesting decision on the table.
If you are about to change a bid strategy, our suggestion is to spend the first week on the conversion actions instead. It is duller work. It is also the part that decides the outcome.