Lookalike Modeling Built on Deposit Value, Not Pixel Fires

Model on funded players. A seed built from registrations will reliably find more registrations, and in the accounts we have taken over, six to eight of every ten registrations never funded anything, which means the model was trained accurately and expensively to find people who fill in forms. The seed decides the model. Everything after the seed is arithmetic. The seed we will build from: Second deposit is the anchor event. Not first deposit, which is still bonus-shaped behaviour and teaches the model to find bonus hunters, and never registration. We weight by ninety-day net revenue decile and let the top three deciles carry the seed. We will not model from fewer than five thousand matched records, and we will not model from a seed whose median tenure is under thirty days, because a seed that young encodes the acquisition campaign that produced it rather than the player you want more of. Negative seeds do more work than most teams expect: Chargebacks, bonus abuse flags, closed accounts and self-excluded players go in as exclusions in every case, and as negative training signal wherever the platform accepts one. In practice the negative seed is what separates a model that finds depositors from a model that finds people who deposit once and leave you the payment fees. Identity and match rate, before any promises: Brazilian sign-up flows are phone-first and the email on file is frequently disposable, so an email-only CRM export can match at roughly half the rate of an export carrying hashed phone as well. We run the match test first and quote model quality afterwards, never the other way round. The lawful basis for the transfer under the LGPD is settled in writing before a file moves, and we will ask to see it rather than assume your DPO has looked at the destination. Where the model should live: Google retired similar audiences in favour of optimised targeting in 2023, so a lookalike in DV360 is a bidding behaviour you influence rather than an object you can size, split or inspect. The Trade Desk gives you an audience you can hold in your hand and cut. That difference decides platform more often than any accuracy claim either vendor makes. Sometimes the honest answer is that the platform's own model is adequate and the budget belongs in supply path work instead. We have given that answer before and we will give it to you before invoicing a build rather than after. Validation: Geo holdout or PSA holdout, agreed before launch, with the split written down. No holdout, no lift claim. Not a softer claim with caveats, none. If your team cannot hold back a market for the measurement window, say so at the start and we will scope this as a build with no lift number attached, which is an honest thing to buy and a common one. What is not included: CDP or warehouse implementation. CRM hygiene beyond the export specification we write for you. Creative, media buying and landing pages. We do not hand over model code and we claim no IP in an audience built inside your seat. It is yours, it stays in your seat, and you can end the engagement and keep it working. Who this is not for: Advertisers whose only first-party event is a page view. There is nothing to model and we will say so on the call rather than sell you a thinner version of this. Teams that need the model judged in two weeks. A deposit-value model needs at least one full deposit cycle, which in the Brazilian igaming accounts we have measured has run thirty to forty-five days, and reading it earlier tells you about your creative rather than your model.