Media acquisition
Give your ad platforms the context they can't see, so budget goes to the customers worth acquiring.
Ad platforms bid on the signal you send them. Report every signup as a conversion, and they learn to find more signups, whatever each one turns out to be worth. This page shows how to score signups and customers for the value you care about, how the scores reach your ad platforms, and how to measure the result.
What ad platforms don’t know
Ad platforms know clicks, impressions, and on-site behavior. They don’t know whether a company is growing or about to fail, what a customer is likely to be worth over time, or whether a prospect looks like your most valuable customers or your least valuable ones. So bids treat a valuable prospect and a poor one the same way.
Match your conversion events to the companies behind them, by CNPJ, then attach Avra scores to each one: for example, a score for your definition of a valuable customer. Your conversion events now carry context the platforms can’t see on their own.
The data
Avra RFM predicts your definition of a valuable customer from your own customer, revenue, and conversion tables. A typical schema ties each signup to the campaign that brought it and the company behind it, and each company to the transactions it makes once it’s a customer. The schema below is illustrative; see Relational data for how the RFM reads yours.
The click ID ties a platform’s event to your signup, and the CNPJ ties the signup to a company. A customer’s value comes from its transactions after it signed up. Every table also carries its snapshot time, so a backtest sees only what was known on each date. See Time and history.
| Data | Purpose |
|---|---|
| Customers with their value | Actual or estimated value per customer, for value-based optimization |
| Conversion events | Which leads became customers, and when |
| Match keys | CNPJ or other identifiers that tie platform events to companies |
For Brazilian companies, Avra GFM embeddings can add context from outside your data, such as a company’s owners and sector, to your own models.
The prediction problem
Each use of the score is a question with three parts: who is scored, the outcome, and the window:
| Use | Who is scored | Outcome | Window |
|---|---|---|---|
| Qualified conversions | Each new signup, when its CNPJ is captured | Becomes a valuable customer, as you define it | 90 days |
| Value-based bidding | Each new customer | Revenue in its first year, a number | 12 months |
| Exclusions | Prospects you can match by CNPJ | Churns early | 90 days |
You define “valuable”: a customer that transacts in its first months, crosses a revenue threshold, or stays past a date that matters to your business. The definition becomes the label, so it should be one you’d pay to acquire more of. See What you can predict.
From score to bid
Score each signup the moment its CNPJ is captured, and send the result back to the ad platform as a conversion event that carries a value. The platform learns what your valuable customers look like and bids for more of them.
The same scores improve the rest of the campaign. For better lookalikes, seed audiences with your highest-value customers, not just everyone who converted, so the platform finds prospects that resemble them. For value-based bidding, send predicted value as the conversion value, so bidding optimizes for valuable customers, not raw conversions.
Scores also show who to leave out: exclude prospects likely to churn early from acquisition campaigns, and suppress low-scoring leads from retargeting. When you retarget, rank the pool by score and show your strongest creative to the prospects most likely to be worth it.
Example: a value-optimized campaign
Score your customer list
Score it with a model that predicts your definition of value.
Upload it as a custom audience
Include a value column with each customer’s score.
Build a value-based lookalike
Seed it with that audience.
Optimize for value
Set the campaign to optimize for value instead of conversions.
The platform learns what your valuable customers look like and finds more of them.
How it’s evaluated
Evaluation happens twice. Before any spend, the model is backtested out of time on past signups, against the targeting you use today, and measured by how many valuable customers it places at the top of the ranking. In market, the campaign is compared with one that runs on platform-native targeting alone. See Evaluation.
| Stage | Metric | What it tells you |
|---|---|---|
| Before spend | Lift and precision at the top of the ranking | How many more valuable customers the top-scored signups hold than a random selection |
| In market | Incremental ROAS | The return on ad spend compared with platform-native targeting alone |
| In market | Cost per qualified lead | Targeting efficiency when bids use Avra scores |
| In market | Value of acquired customers | The average value of customers acquired with and without Avra scores |
Integration
| Moment | Path | How it’s used |
|---|---|---|
| A signup arrives | Live predictions | Score it when the CNPJ is captured, and send the conversion event with its value |
| Audience lists | Batch inference | Score your customer list for custom audiences and value-based lookalikes |
| Lookalikes by structure | Embeddings | Rank prospects by the cosine similarity of their GFM embeddings to your most valuable customers’ |
The embeddings technical guide has a complete similarity search to start from.