The short version: Google and Meta pursue the people most likely to complete the conversion event you specify. If that event is a form fill, the platforms find people who submit forms, including many who never buy. Event depth, match quality, and conversion value determine whether automated campaigns learn from customers or cheap proxy actions.
Advantage+ and Smart Bidding are often called black boxes. That label hides the input you still control: the definition of a successful customer.
Tell the platforms that a form fill is success, and they will find people likely to fill a form. Some will be tire kickers, students, or competitors because the campaign was never given a revenue outcome.
This is the part of paid media that survived automation. You have less audience and bid control, but you still define what winning means.
Why Do Google and Meta Ads Find the Wrong Customers?
Customer quality under automated bidding depends on three signal decisions. You choose the success event, how accurately it matches to a person, and the conversion value. Those inputs shape who Google and Meta pursue after placements and bid adjustments move inside the model.
Most teams keep tweaking campaign settings while ignoring signal quality. Ask what would change if you doubled the budget tomorrow. If the answer is “more of the same low-quality leads,” the problem sits upstream of the budget.
| Dial | What it tells the model | Default setting most accounts run |
|---|---|---|
| Event depth | What counts as a win | The easiest event to fire |
| Match quality | Who the winner was | Browser pixel, one identifier |
| Conversion value | How much the win was worth | Every conversion worth the same |
Each row is a decision someone made, usually by accepting a default during setup. None of them are permanent.
Which Conversion Event Should You Send?
Send the deepest event that still generates enough weekly volume for the model to learn. Deeper events point it toward revenue, but sparse events starve it of data. Choose the point where depth and volume meet, then revisit it as you scale.
Optimizing on a shallow event is the most common failure we see in audits. A lead form fires 400 times a month, so it feels safe. The model finds the cheapest 400 form fills available, but those people are not necessarily buyers.
| Business type | Usually the wrong event | Usually the right event |
|---|---|---|
| Ecommerce | Add to cart | Purchase, with order value |
| B2B SaaS | Demo form submitted | Qualified opportunity created |
| Subscription app | Free trial started | Trial converted to paid |
| Lead generation | Any form fill | Sales-accepted lead |
When the deep event is too sparse to support learning, do not force it. Optimize on the mid-funnel event, import the deep one for reporting, and watch the ratio between them. Once the deep event clears a stable weekly cadence, move optimization down to it.
The ratio itself is the useful number. If form fills grow 40% and sales-accepted leads stay flat, the model found a cheaper pocket of people who were never going to buy.
How Much Does Match Quality Change What the Model Learns?
Match quality changes more than most teams expect because it tells the model who converted. An event that cannot be tied to a person is a conversion you paid for but cannot learn from. More identifiers give the platform more chances to find a match.
Meta accepts a long list of customer information parameters on server-sent events. According to Meta’s Conversions API parameter documentation, you can attach hashed email, phone, first and last name, date of birth, and location fields. You can also send an external ID of your own, plus unhashed IP address and user agent. Only one of them is strictly required.
The minimum rarely produces the best match. Default setups often stop once the event meets the platform requirement.
Google’s equivalent move is user-provided data on lead conversions. Its guide to managing offline conversions says enhanced conversions for leads uses hashed customer data to improve accuracy and bidding performance. Google matches that data to website leads and signed-in customers who engaged with the ad.
Three checks catch most match-quality gaps:
- Count your identifiers. If your server event sends email only, you are leaving phone, name, location, and external ID unused. Each one is another chance to match.
- Check your hashing and formatting. Meta expects lowercase, trimmed values, phone numbers with country codes, and dates as YYYYMMDD. Malformed input hashes cleanly into a value that matches nothing.
- Confirm deduplication. Pixel and server both firing without a shared event ID inflates your conversion count, which quietly corrupts every cost-per-conversion decision after it.
Not sure what your account is teaching the algorithm?
We audit the signal layer as part of every paid media engagement, starting with which event your bidding optimizes toward.
Book a Free Strategy CallShould You Send Conversion Value or Just Conversion Counts?
Send value when your conversions are worth meaningfully different amounts. Count-based bidding treats a $40 order and a $900 order as identical wins, so the model chases whichever is cheaper to produce. That is usually the small one.
Value-based bidding optimizes toward total conversion value rather than volume. Where order sizes or expected lifetime values spread widely, that difference is the whole game. Where every conversion is worth about the same, value and cost-per-acquisition bidding land in a similar place, and the added plumbing earns little.
You do not need perfect revenue data to start. Begin with a defensible estimate:
- Real revenue when you have it at conversion time, as in ecommerce.
- Predicted value by segment when you do not, such as an average first-year contract value per plan tier.
- Relative weights as a floor, where a qualified opportunity is worth 10 points and a newsletter signup is worth 1.
Even option three stops the model from spending your budget on the cheapest possible action. Refine the numbers later, once the direction is right.
What Is Changing in How You Send This Data?
The pipes are consolidating, and one deadline has passed. Google is moving offline conversion and lead uploads toward its Data Manager API. Its Google Ads API documentation warns that since June 15, 2026, UploadClickConversion requests fail if the developer token has not previously sent offline conversion or enhanced conversion for leads uploads.
The restriction applies to new integrations on the old path. It does not switch off existing uploads. New setups should use the newer path.
This matters most for B2B accounts whose real outcome appears in a CRM weeks after the click.
What Guardrails Keep the Signal Honest?
Feeding the model better data also means feeding it wrong data faster. Four rules hold the line:
- Change one thing at a time. Switching the optimization event and turning on value bidding in the same week leaves you unable to attribute the result to either.
- Expect a learning reset. A new optimization event restarts learning. Budget for 2 to 3 weeks of noisy performance before judging it.
- Never send a value you cannot defend. An inflated placeholder value teaches the model to buy more of whatever carries it.
- Verify the signal against reality. Platform-reported conversions are a model output, not the truth. Check them against the CRM or ledger monthly.
That last point is why signal work and measurement work belong together. Better inputs make the platform more confident, but they still need independent validation. Pair this with incrementality testing to separate a better model from better reporting.
What Should You Do This Week?
Open your bidding settings and write down the answers to three questions. Which event does the campaign optimize toward? How many identifiers travel with that event? What value is attached to it?
The answers reveal which customers Google and Meta are trained to pursue. A shallow event teaches the platforms to find more people who take that action, whether or not they buy.
Most accounts uncover at least one setup default they never chose. Correcting the event is usually the largest available change, and it costs only a learning period. For the wider picture, our paid media audit guide covers the rest of the account. Our breakdown of Advantage+ versus manual campaigns explains where automation earns its keep.
Want to know what kind of customer your account is trained to find? Book a Free Strategy Call and bring your conversion settings.
Like this? Get the next one.
Short emails. New posts as they ship.
