The Estimated Action Rate Explained

Quick Answer

The Estimated Action Rate (eAR) is Meta's per-impression probability prediction that a specific user will take your optimised action if shown your ad. It's the multiplier on your bid in the Total Value formula. eAR is computed by a deep learning ranker that ingests user features, ad features, and context features, and outputs a calibrated probability between 0 and 1.

The Mechanism Explained

The eAR model is technically a multi-task neural network with shared lower layers and task-specific heads. The shared layers learn general user-ad interaction patterns. The task-specific heads predict different actions: link click, video view, lead, purchase, install, and so on. When you set an optimisation event, you select which head's output gets used in your auction.

Inputs to the model include (publicly disclosed in Meta's engineering posts):

1. User features — demographics, interest embeddings, recent actions, app usage, social graph signals, device type
2. Ad features — creative embedding (image and text are both encoded), advertiser fingerprint, ad category
3. Context features — time of day, placement, scroll position, current session length
4. Cross features — interactions like (user × creative category), (user × time of day)

The model is trained on labelled clickstreams from past auctions. Crucially, Meta uses implicit calibration — the raw output is post-processed so that an eAR of 0.05 actually corresponds to a 5% historical conversion rate in similar contexts. This calibration is what makes eAR comparable across different ad sets and bid strategies.

In 2024-2025, Meta migrated the eAR backbone to the Andromeda architecture (sparse mixture-of-experts), which lets the model specialise per vertical without ballooning compute.

Practical Implication

You can't see eAR but you can influence it. The biggest levers are creative quality (the image embedding contributes a lot), conversion event quality (clean labels train better tasks heads), and consistency over time (the model trusts your account more once it has stable history). High eAR means you can bid less to win the same auction.

Real Numbers

• eAR model has billions of parameters across the multi-task heads
• A 10% eAR lift typically reduces effective CPM by 8-9% in steady state
• Meta retrains the eAR backbone on a weekly cadence, with online updates more frequently

FAQs

Q: Is eAR the same for everyone seeing my ad?
No — it's per-user, per-impression, per-context.

Q: Can I see my eAR?
No. The closest proxy is your relevance diagnostics.

Q: Does eAR include landing page quality?
Indirectly — Meta has a separate post-click quality model whose output feeds the ranker.

Q: How does eAR handle new ads with no history?
Cold-start from creative embedding + advertiser priors + similar ads' performance.

Q: Can creative alone push eAR up?
Yes — image and copy embeddings significantly influence the prediction.

Pix-Vu

Image quality is one of the largest single inputs to eAR. Pix-Vu lifts your creative quality without you needing a photographer or studio — try it at https://pix-vu.com.