How Conversion Value Optimisation Works Explained

Quick Answer

Value optimisation replaces eAR with Estimated Action Value (eAV) — a prediction of how much revenue a user will generate, not just whether they'll convert. The auction formula becomes Bid × eAV × pConversion + Quality. The model is trained on conversion events that include a value parameter, and it learns to bid higher on users predicted to spend more.

The Mechanism Explained

Standard optimisation predicts probability of conversion. Value optimisation adds a second prediction: conditional value given conversion. The full estimate is:

Expected Value = P(conversion) × E[value | conversion]

The model needs three things to do this well:

1. Conversion events tagged with values — passed via pixel value parameter or CAPI custom_data.value
2. Variance in those values — if every conversion is $50, value optimisation adds nothing
3. A volume floor — Meta needs roughly 30-50 value events per week per ad set to train a stable value head

When configured correctly, the model biases delivery towards high-value users. In practice this looks like Meta serving your ad to a smaller audience but at higher average order value. Spend efficiency (ROAS) usually goes up while volume goes down, especially for ecommerce with wide AOV distributions.

The Highest Value bid strategy uses eAV directly as the bid signal. Minimum ROAS adds a constraint: only enter auctions where the predicted ROAS exceeds your floor.

A subtle mechanism: Meta uses a log-transformed value model so a $500 conversion isn't 10x weighted vs a $50 conversion — it's roughly 2.3x. This prevents one whale from dominating the model.

Practical Implication

Value optimisation shines when your AOV variance is wide (e.g. SaaS with multiple plans, ecommerce with bundles). It hurts when AOV is uniform — you're adding noise to the model for no signal. Verify variance before switching: pull your last 90 days of conversions, compute the coefficient of variation. If CV < 0.4, stick with cost-based optimisation.

Real Numbers

• Value optimisation needs ~30-50 value events/week to stabilise
• Median ROAS lift from value vs cost optimisation: 15-30% when AOV variance is high
• Value events with missing or zero value are dropped from training
• Log transform applied at value > $10

FAQs

Q: Do I need server-side conversions for value optimisation?
Strongly recommended. Pixel-only signals lose too many events post-iOS 14.

Q: What if my AOV is fixed?
Use cost-based optimisation — value adds no information.

Q: Can I set a minimum ROAS during learning phase?
Yes, but the model often misses the floor in learning. Expect underdelivery.

Q: Does subscription value count?
Only the first-purchase value unless you're passing predicted LTV.

Q: Is value optimisation available for all events?
Only events that accept a value parameter — Purchase, Subscribe, custom.

Pix-Vu

If you're running value-based bidding, your creative needs to attract higher-AOV buyers — and that often means premium-looking imagery. Pix-Vu transforms basic product shots into clean, premium-feeling visuals at https://pix-vu.com.