Smart Bidding for online shops: what it optimises and what it misses

Smart Bidding maximises the number you report

Smart Bidding covers the bid strategies that, according to Google, bid for conversions or conversion value in every single auction: Target CPA, Target ROAS, Maximise conversions and Maximise conversion value. Google assigns Target ROAS and Maximise conversion value to the business goal "Increase profit". In my view that only holds under one condition: the conversion value has to reflect profit. If the shop reports basket revenue, the strategy chases revenue, even when profit falls in the process.

Suppose a shop sells two products at €100 net each and runs a target ROAS of 4, which means €25 of ad spend per order. The figures are invented, the arithmetic is not:

In the account both show a ROAS of 4, and to the bidding logic both orders are worth the same. The algorithm is not making a mistake here. It has simply been given the wrong task.

The rule has a clear limit. If every product earns roughly the same share of revenue, revenue is a usable proxy for profit, and switching changes little more than the target figure. You can check this in ten minutes: work out the break-even ROAS for each category, which is 1 divided by the contribution margin rate. A category at 40% sits at 2.5, one at 20% at 5. If your shared target is below the break-even of a high-revenue category, every order there is subsidised by the other categories. Then the work on the value pays off. If not, there are more urgent problems to fix.

Which signals does Smart Bidding use, and what is missing from all of them?

According to Google, signals are identifiable attributes of a person or their context at the moment of the auction. These are the ones relevant to shops (list as of November 2026):

These signals help predict how likely a conversion is and what value it brings. What is left of the order after purchasing, shipping and returns appears in none of them. Nor do they show whether the purchase would have happened without the ad. For me that is the second big limit alongside margin: the system learns from attributed conversions, and an attributed order from a loyal customer looks to the bid like a genuine additional sale. A high ROAS on brand terms or with existing customers can therefore say little about the additional profit.

Which strategy suits which shop?

The classification in the table is mine, not Google's. For most shops with widely varying order values, it points towards value-based bidding.

How to choose between the two value strategies is covered in Maximise conversion value with or without a target ROAS. Target ROAS can optionally be combined with Smart Bidding Exploration: according to Google, a ROAS tolerance lets the campaign reach additional search queries, and for Shopping ads this runs as a beta. Since bids are then allowed to drop below the target, I measure the tolerance against break-even, not against the target. In an Optmyzr analysis of 14,584 customer accounts, Maximise conversion value performed clearly better than Maximise conversions on ROAS, CPA and CPC. These are observational data from a vendor, measured on ROAS and CPA, not on margin. All I take from them is that nothing argues against value-based bidding as the default.

For evaluation, Google gives a clear figure: judge results over longer periods with at least 30 conversions, or 50 for Target ROAS, so roughly a month or more. That is not a threshold for starting, but it does have a consequence for structure. If you split by margin into five campaigns and want to judge each one separately, you need 250 conversions in the same period. A correct value in the conversion does part of that splitting without an extra campaign. That is why I would work on the value first and on the structure afterwards.

The most expensive data errors never trigger an alarm

A missing tag gets noticed within days because conversions collapse. A wrongly defined value never gets noticed, because every report looks consistent. My ranking by damage, which looks different if margins are uniform and returns are low:

1. Revenue instead of contribution margin as the value. Acts in every auction and favours low-margin items as soon as they sell more easily.
2. Returns missing from the bid. Categories with high return rates permanently look better in the account than they are.
3. New customer and repeat order valued the same. The system takes the easier sale, and the customer base does not grow.
4. Technical faults. Expensive while they last, but visible and limited.

The fourth point has its own tool. Google names tagging problems, website outages and upload problems as typical causes that data exclusions are meant to absorb. They only affect what the bid strategy reads. Your reporting stays unchanged, and Google does not guarantee stable performance as a result. Set the exclusion quickly and choose the date range so that it covers the usual delay between click and purchase. Used constantly, exclusions harm performance, according to Google itself, so they are no remedy for accounts where measurement breaks regularly. There, the repair comes before any bidding work.

Returns: the correction usually arrives too late for the bid

Google allows value restatements and retractions, ideally via the order ID. According to Google, however, these corrections are only readable for automated bidding up to 7 days after the conversion was first recorded, while adjustments overall are possible for up to 54 days. Set against this is the withdrawal period: in Germany it is 14 days and, for the sale of goods, usually only starts when the goods are received.

Suppose an order arrives on day 3. The period then runs until day 17. If the customer withdraws on day 15 and the returned parcel has been checked on day 20, the correction lands 13 days after the window for the bid closed and well within the window for the report. If you offer longer voluntary return periods, the gap grows further. My conclusion: the retrospective upload mainly repairs the report, hardly the bid. It catches early cancellations, but not the typical return.

If you want returns reflected in the bid, you have to price them in at the time of purchase, as an expected value per category:

Reported value = net revenue × product margin × (1 − return rate) − return rate × cost per return

Again with invented figures, per €100 of net revenue and assuming the returned goods go fully back into stock:

On margin, A comes first. On expected value, B does. The value will rarely be exactly right for a single order. It only has to be right on average per category. That is why the rate needs updating from the inventory system regularly, otherwise the bid quietly goes stale.

There is a trap in the combination. If you already deduct the expected rate in the value and then upload every actual return as a retraction, you count returns twice in the report. In that case I would reconcile actual returns in the shop report and adjust the rate from that. The upload still makes sense for cancellations before dispatch, as long as the rate does not include them, and for accounts that keep bidding on revenue. The technical route is described in sending returns to Google Ads. With a low return rate and uniform margins, the whole effort hardly pays off, and a flat deduction in the target is enough.

New customers only count for more in the bid once you say so

Without instructions, a first order is just another order to the system. Google offers customer lifecycle goals for this. New Customer Value bids higher for new customers than for existing ones, works with Target ROAS or Maximise conversion value and is Google's recommendation for purchase goals. New Customer Only bids exclusively for new customers. Google only recommends it for strict acquisition budgets or non-purchase goals, and existing customers then need a campaign of their own. For Search, PMax and Shopping, Google is also rolling out retention goals and High Value Mode in Shopping. These let you value returning customers separately, and that too is a value setting you should justify from your own repeat purchase data.

How much higher is a commercial figure, not a platform question. Suppose a first order brings €30 of contribution margin and your own cohorts show, on average, one further order worth €30 within twelve months. Then the new customer is worth €60 and the uplift is €30. Only count repeat purchases you can see in your own data. An uplift based on hope lets the system spend money that never comes back. The calculation works out more readily for consumables than for a sofa. The implementation in detail is covered in the new customer goal in Google Ads.

How to switch the value without losing control

Moving from revenue to margin lowers the reported value per conversion, and the target has to move with it. The calculation is simple: new target = old target × (new value / old value). A target ROAS of 4 (400%) becomes 1.4 (140%) with a value of 35% of revenue. If you report contribution margin before advertising, the figure reads differently afterwards: 1.0 is break-even, and anything above it is what is left over. How to derive the level of the target is covered in deriving the target ROAS from margin.

The arithmetic is clean. The algorithm's response is less so. The vendor ProfitMetrics writes that there is no way to predict how the algorithm will react to the change. I know of no independent controlled study on profit values as a bidding basis. I would therefore treat the switch as an experiment:

• Clarify beforehand whether the value passed can be read in the browser and whether you are happy for your margin to be visible there.
• First record the new value in parallel as a secondary action and check it against the inventory system, and only then make it the primary action.
• Start with a campaign whose revenue you can afford to lose, and keep the daily budget close to the previous average.
• Judge after the 50 conversions at the earliest, and measure against contribution margin after advertising, not against ROAS, which has changed anyway.

Details on the pitfalls are in margin as the conversion value. With very few conversions I would postpone the switch, because you cannot judge the result within a reasonable period.

What stays with you, and how to start

I would not want the bid in the individual auction back. The definition of value, the level of the target and the offer itself are decisions the machine cannot make. One check is enough to begin with: take last month's orders attributed to Google Ads and put the reported conversion value next to the contribution margin after returns, per category. If the ratio differs markedly between categories, you are currently optimising for a different goal from your own, and every further piece of bidding work reinforces that.