What you will find in this article
- What Target ROAS actually predicts: not the conversion rate, but the predicted conversion value of each single click
- The auction-time signals the algorithm combines to estimate that value, and which of them you can influence
- The complete matrix: which inputs feed the prediction and how much weight each one carries in practice
- What Google officially documents about value prediction, with three official sources separated from field observation
- Behaviours observed on real campaigns during the learning phase, clearly flagged as inferences
- The practical workflow to feed the algorithm clean value data so the prediction converges faster
Two campaigns, same budget, same products. One hits a 480% target roas in three weeks. The other stays stuck at 220% and never recovers. Same bidding strategy, different results. The reason is rarely the target you typed in: it is the quality of the value data the algorithm had to learn from before it could predict anything at all.
Target ROAS does not bid on clicks. It bids on a number it invents for every single auction: the predicted value of the conversion that click might produce. Understanding how that conversion value prediction is built, and how much it leans on accurate conversion data and historical data, is the difference between an advertiser who fights the algorithm and one who feeds it.
This article clearly separates what Google officially documents about value prediction from what emerges from running and observing Target ROAS campaigns over many learning cycles.
What Target ROAS predicts, and why it is not the conversion rate
The most common misunderstanding is treating Target ROAS like Target CPA with a different label. They optimise toward fundamentally different predictions. Target CPA predicts how likely a click is to convert. Target ROAS predicts how much a conversion will be worth, which is what makes it the natural fit when your business objectives are framed around return on investment rather than raw volume.
In practical terms, the algorithm builds an Expected Conversion Value for the click and divides it by your target return on ad spend (ROAS) to derive the bid in real time. A click with a predicted value of $50 and a target ROAS of 500% produces a maximum bid of roughly $10. The prediction is the variable. The target is just the divisor you control.
The crucial consequence: the prediction is only as good as the values you feed it. If every conversion reports the same flat value, Target ROAS has nothing to differentiate high-value from low-value clicks, and it degenerates into a volume strategy with a constraint bolted on.
The signals that feed the value prediction: what the algorithm actually weighs
The value prediction is not a single number pulled from history. It is recomputed at auction time by Google's algorithms, combining contextual signals with your reported conversion values. Some of these you can influence directly; others you cannot.
Prediction input matrix: what feeds Target ROAS and how much it weighs
This table summarises the current state (June 2026) for each input into the value prediction. Sources are indicated for documented information; for those inferred from running campaigns this is explicitly stated.
| Input | Officially documented | Advertiser control | Practical weight on prediction | Source |
|---|---|---|---|---|
| Reported conversion values | Yes | High | Very high | Official Google Ads |
| Conversion history depth | Yes | Medium | Very high | Official Google Ads |
| Conversion value rules | Yes (XMP) | High | Medium-high | Official Google Ads |
| Device / location / time | Yes | Indirect | Medium | Official Google Ads |
| Query / intent context | Partial | Indirect | Medium | Official + field analysis ⚠️ |
| Audience / remarketing signals | Partial | Indirect | Medium-low | Field analysis ⚠️ |
| Target ROAS value set | Yes | High | Throttle only | Official Google Ads |
| Data freshness / import lag | No | Medium | Medium-low | Field analysis ⚠️ |
What Target ROAS actually reads: Smart Bidding inputs and Shopping setup
The prediction is not a black box you cannot feed. The two levers Google documents most explicitly for Google Ads bidding are the conversion values you report and the conversion history the strategy can learn from. Everything else is context the algorithm layers on top. For Shopping specifically, the setup makes the value-prediction logic concrete at the product level.
The practical reading: Target ROAS does not approve or reject clicks the way ad review approves or rejects creatives. It scores each auction with a predicted value and bids proportionally. The advertiser's job is not to fight the score but to give it accurate, differentiated values to learn from.
Inferences from running campaigns and observed behaviour
The learning phase is a value-calibration phase, not just a data-collection phase. In the first one to two weeks after switching to Target ROAS, bids swing widely because the value prediction has high uncertainty. On accounts where I fed differentiated conversion values from day one, the prediction stabilised noticeably faster than on accounts reporting a flat value per conversion. The pattern is consistent: the model converges on what it can actually distinguish.
Flat conversion values are the single most common silent failure. Across audited accounts, the campaigns stuck below target almost always reported one fixed value for every conversion. Target ROAS cannot predict a value distribution it has never seen, so it spreads spend evenly and the realised ROAS hovers around the account average regardless of the target typed in. Switching to true revenue values, where available, was the change that moved the needle most.
Aggressive target jumps reset the calibration. When I raised a target ROAS by a large step in a single move, the prediction effectively retrenched: volume collapsed and the algorithm narrowed to only the most confidently high-value auctions, often under-spending the budget. Smaller incremental moves let the prediction re-centre without throwing away learning. This is a guardrail effect, not a relearning of value.
The target value is a throttle on the prediction, never a teacher. Images of Target ROAS as "telling Google to find better customers" are misleading. The target does not improve the prediction of value; it only filters which predicted values are worth bidding on. Every account where performance improved did so because the input values got better, not because the target got higher. The distinction between the value prediction and the efficiency guardrail is the most useful mental model I have for diagnosing a tROAS campaign.
Practical workflow: how to feed the value prediction so it converges
1. Send true, differentiated conversion values before switching strategy. Before applying Target ROAS, make sure conversion tracking reports the real value of each conversion, not a flat placeholder. For ecommerce this means actual transaction revenue, ideally weighted by profit margin where you can model it; for lead gen it means value tiers that reflect lead quality. The prediction can only differentiate what your data differentiates.
2. Clear the conversion-history gate before expecting stable predictions. Each campaign type has a baseline of recent conversions the strategy needs to predict reliably. Below that threshold the prediction is wide and bids are erratic. If you are under volume, run Maximize conversion value without a target first, the same conversion value bidding strategy minus the constraint, to build history, then add the target.
3. Use conversion value rules to encode what your reporting cannot. Where device, location or audience genuinely change the value of a conversion to your business, apply value rules so the prediction sees those differences:
Tools → Conversions → Value rules → New rule (device / location / audience) This
sharpens the predicted-value distribution without touching the target.
4. Move the target in small steps and read the average target ROAS report.
Adjust target by 10-15% → wait one full learning cycle → check Average target ROAS column The average target ROAS metric shows the traffic-weighted target the strategy actually optimised for, which tells you whether your changes took hold or were averaged away.