How to move from tCPA to tROAS with conversion value prediction

We cover how to optimize paid campaigns for the most valuable users by leveraging predicted conversion values, like lead scoring

First things first, let’s introduce David Loris: David is a Data Scientist specialized in marketing. He has been working at the intersection of data and marketing for almost 20 years, having helped brands such as Booking, Expedia, GetYourGuide, Typeform and Zalando improve marketing performance through data and measurement techniques. Together, we help out companies to roll out their first conversion value prediction tests for paid media.

---

Harnessing the best out of ad platform optimization

Marketing teams spend millions optimizing toward sign-ups, even when sign-ups are not all created equal. Some churn quickly. Some convert late. Others bring in outsized revenue, renew again and again, and invite their entire company. You miss that nuance if you’re optimizing paid campaigns for sign-ups only by running tCPA instead of tROAS.

Businesses with long sales cycles or freemium models face difficulties moving from tCPA to tROAS because partial or complete revenue only comes after conversion lookback windows (always a maximum of 30 days). Without predicting the value of conversions (visitors, signups, or leads), marketers can’t harness the best of ad platforms’ smart bidding algorithms and are stuck targeting poor-quality and high-quality leads.

In this article, we’ll cover:

• What is the impact of conversion value prediction, and which businesses need it the most

• What are the different types of conversions you can predict value for, and which one(s) should you use

• What’s required to implement this technically

• How to run your first test—including setup, model design, and analysis

Whether you’re working with long sales cycles, freemium models, or just struggling to scale CAC efficiently, this guide will help you assess if conversion value prediction is worth implementing—and how to do it well.

In what contexts is bidding using predicted conversion value impactful?

In most cases, predicting conversion value is a complex activity involving both data and marketing teams. It’s not a necessary activity, so it shouldn’t be treated as a default playbook. But it can create a meaningful business impact when:

1. Conversion to a paying customer happens a long time after an ad click
   In long sales cycles, for example, for medical equipment, financial products, or enterprise tools, conversions may take weeks to months, making it too late for smart bidding to react.

2. LTV varies significantly across user segments
   SaaS is a clear example. Depending on seat count, pricing tier, and churn, one user might be worth 50x another. Optimizing based on cost per signup ignores this nuance. This is also relevant for mobile games, where you commonly see 20% of users are responsible for 80% of the revenue (“whales”).
   

Not every business should be investing in lead scoring

In these cases, revenue prediction helps bring valuable data back into the platforms—so you can bid higher for leads that matter, and lower for the ones that don’t.

Why predicted values can really impact paid performance

Paid platforms like Meta and Google no longer deliver the best results with manual audience targeting. Instead, performance depends on the quality of creatives and the signals you feed back—especially when running conversion or value-optimized campaigns.

B2B companies, for example, often bid based on a cost per lead target which treats all conversions equally (setting a target CPA bid in Google ads). But this leaves massive inefficiencies.

Let’s say you’re targeting a $5 CAC for all leads. But some users are only worth $3, while others are worth $100. If you're capping bids at $5 across the board, you're potentially blocking yourself from acquiring those $100 leads—because you’re not willing to bid more aggressively when it makes sense.

Imagine that winning a $100 lead would require a $10 CAC. Under your current ceiling, that user never enters the funnel—even though the return would be 10x. That’s money left on the table.

With conversion value predictions, you can:

• Run tROAS campaigns that bid based on predicted conversion value

• Adjust your bidding dynamically based on firmographic, behavioral or other data

• Power better optimization with stronger data signals, even before conversion to a paid customer

Example: Architecture of a conversion value prediction model for B2B

There are different actions in a user journey whose value can be predicted

In this example we have a typical funnel for a B2B company, with different user actions leading to revenue. We can build separate models to predict values for each funnel step using machine learning or other scoring methodologies.

1. probability of the visitor converting into lead

2. probability of lead converting to paid customer

3. the predicted revenue once the customer converts

Each model may have different data inputs. The further down the funnel, the more information we can collect which can be used in prediction.

Once we have developed these models they are then combined in order to find the predicted value of the lead.

The standard calculation for lead value

This value can be passed to smart bidding as conversion value for the lead allowing us to use ROAS instead of bidding on a cost per lead target.

What’s technically required for building a revenue prediction model?

For B2B, a common and useful conversion to predict the value of is “lead.” If you want to assign a predictive value to a whitepaper download or call booked, you’ll need a few things in place first.

1. Data Warehouse and Platform IDs

You’ll need to store user data—and tie it back to campaign identifiers—in a warehouse or structured environment.

Important considerations:

• You must link marketing identifiers (like gclid, fbclid) to in-house user IDs

• Meta doesn’t require fbclid, but Google does require gclid for conversion value import

This connection allows you to pass lead values back to platforms like Google and Meta and effectively run tROAS bidding strategies.

2. Additional Data Points

Data quality determines model quality. Common variables in conversion value prediction models include:

• Data collected at conversion: Country of signup, device type

• Behavioral data: Visits to pricing or product pages, number of sessions, pages viewed, or time on site

• Firmographic data: Company size, industry, revenue, tech stack

• Product data: High-intent behaviors (e.g. invited colleagues, downloaded assets)

3. Model

You’ll need a model to assign predicted values to leads. This can range from simple rules-based logic (e.g. pricing page visits + firmographic data) to machine learning models trained on past conversions.

If your company has no existing conversion prediction, start simple:

• Use only one data source, like enrichment tools (e.g. Clearbit, GA4 etc)

• Use a rule based model instead of machine learning (e.g bid double on leads based in the US)

• Start with a binary model (likely/unlikely to convert) and iterate

Over time, you can layer in more signals and increase complexity.

⚠️ Common pitfall: Overengineering the first model. Better to be roughly right and usable than precisely wrong and overfit.

An example of lead scoring using both website activity and firmographic data

Start with what you have. Then refine.

4. Upload process

You’ll need a way to push your conversion value predictions to the Ad Platform. This is usually done through Offline Conversion Upload API. For this you will create a table with the following information which is then pushed via API.

• gclid: Click ID allowing google to associate this conversion to a click

• Conversion time: Time at which the conversion occurs

• Conversion Value: Predicted revenue from your model

Make sure to run QA to verify that the number of conversions you are uploading matches the number of conversions tracked using pixel.

Example: Rolling out your first paid test with a predicted lead value

This isn’t a one-week project, but much of the work can happen in parallel. Here’s a practical rollout plan:

1. Assess if you need it.
   Not all products benefit. If your LTV doesn’t vary much, or your conversions happen within the attribution window, conversion value prediction may not be worth the effort.

2. Ensure your infrastructure can support it.
   Make sure your tracking system and data capture are properly set up to store the necessary data.

3. Link gclid to internal user data (for Google Ads).
   You need to be able to trace users back to the campaign or keyword level to assign value accurately.

4. Build ROAS reporting.
   For later assessing the impact on campaign performance. Start measuring ROAS and CAC early.

5. Create a revenue prediction model.
   Start small. Don’t overcomplicate your first version. Most teams don’t update models frequently, but you can test improved versions quarterly or when new data becomes available.

6. Run a test.
   Set up your platform to receive conversion values. Then compare performance using ROAS and CAC. Analyze how campaigns behave when optimizing for predicted value.

7. Analyze and iterate.
   Refine your model. Check for model drift. Monitor platform optimization quality. This is a long-term investment, not a one-off test.

Implement conversion value prediction with expert help

You need both the model and the marketing strategy. You need someone to wrangle the data—and someone who understands how ad platforms actually behave.

That’s exactly why I (Barbara) team up with David Loris, a marketing data scientist. Together, we’ve helped companies set up their first lead scoring systems and prove their impact with real media spend.

If you’re ready to:

• Define which conversion to predict the value for

• Understand if this is a worthwhile investment

• Run a first scoring test

👉 Schedule a free call with us to talk through it.