If you run value-based campaigns on Meta or Google, you're training their bidder every day, whether you think of it that way or not. Every conversion value you send tells it which kind of person is worth paying for.
Most advertisers never check what they're teaching it, and for a lot of subscription and ecommerce businesses that's exactly where the money leaks. This post is about your own account: how to check whether your conversion value is misleading the bidder, and what to send instead if it is.
Meta and Google know a lot about the people they can reach: their device, when they're online, what they've bought, how they behave around ads. Google lists some of the signals Smart Bidding uses in every auction, including device, location, time of day, remarketing lists, language, browser, operating system and the search query (Google Ads Help). What neither platform knows is what any of those people are worth to you. The value you send on each conversion is the only place that comes from.
So the model learns from it. It looks at the people who came with high values, finds others who resemble them, and bids more for those people and less for everyone else. In practice, the conversion value you send decides how much you pay for each kind of customer.
The bidder learns from what happens in the first few days. On Meta, value optimization uses conversions inside the attribution window, which by default closes seven days after the click. So the value most advertisers send is a trial or a first purchase.
Say three customers each pay $8 on day one. By day 90, one has refunded, one never came back, and one has paid $96 and is still subscribed. On day one they all look like $8 to the bidder, so it will pay the same to get more of each. You pay full price for the refunder and nothing extra for the customer who turned out to be worth twelve times as much.
Across a whole account that costs you in two places. You overpay for people who buy fast and leave, because they look great in the first week. And you lose auctions for the customers you actually want, because nothing in your data tells the bidder they're worth more than anyone else. That's how day-7 ROAS can look fine while the same cohort disappoints by month three.
You can test this with data you already have. Three checks, from five minutes to an afternoon.
Send a value that reflects what each customer will be worth, early enough for the bidder to use it. That means predicting long-term value per user from their first hours or days of behavior and passing that prediction as the conversion value.
Meta now supports this directly with pLTV value optimization.
If you build the prediction yourself, watch two things. Keep the range of values tight at first, because an accurate prediction with a very wide range will overbid on the top end. And start low, because platforms happily accept a value you raise but rarely let you walk one back down. Signal Engineering Is Right. Now Comes the Hard Part. covers the rest of what goes wrong in production.
When the value you send is right, the bidder stops paying for refunders and starts winning the customers who stay, without any change to your campaigns or creative.
If you want to know how much of this applies to you, start with check number two. It's an afternoon's work, and it'll tell you whether this is worth fixing.
