Beyond Viewability: The Promises and Pitfalls of Attention Metrics

For years, digital media buying lived by a shockingly low bar called "viewability." Under standard industry guidelines, if 50% of an ad appeared on a screen for just one second, it was logged as a "viewable impression."

A viewable impression was a deeply flawed metric that rewarded cheap, invisible ad placements and creating inefficiencies in media budgeting.

Now we have Attention Metrics. (See here for a detailed explanation: eMarketer Attention Metrics Ecosystem Report). Powered by a blend of dwell time tracking, page interaction modeling, scroll velocity, and opt-in eye-tracking panels, attention models promise to evaluate whether a human actually gave your ad mental surface area and for how long.  Seems like a stretch to think we are suddenly able to measure what goes on inside someone’s head, but I digress…

The industry’s rush toward attention data makes intuitive sense, but to get real value out of these tools, it’s critical to enter into this type of measurement with your eyes wide open to what the data actually represents – and what it may leave out.

Outside the Lab: How Attention is Measured Today

Eye-tracking studies have existed for decades, but modern attention vendors aren't just dragging people into physical labs. Today, attention metrics are calculated at scale through two primary methods:

• Active Interaction & Dwell Time: Measuring real-time page physics – how long an ad unit remains in the active viewport, scroll speed, touch/mouse movement, and user interaction duration.

• Predictive Eye-Gaze Models: Machine learning algorithms trained on panel data to predict visual focus based on ad placement, screen real estate, and page clutter.

While combining dwell time with predictive modeling creates a vastly superior quality filter compared to legacy viewability, scaling these models doesn't automatically mean they can or ever will predict revenue.

The Nuance of Dwell Time

Modern attention models have made great progress in filtering out passive screen idle time by recognizing when a page is in a background tab or when user interaction stops completely.

However, dwell time alone is a noisy indicator. A high dwell time score tells you an ad was on an active screen for four seconds, but it cannot reveal user intent:

• Did the user pause for four seconds because they were captivated by the brand?

• Or did they pause because someone was calling their name, the headline was confusing, the video was jarring, or they were looking for the "X" button to close an overlay?

The Core Limitation: Measuring the "What," Missing the "Why"

Whether a vendor relies on dwell time algorithms, interaction proxies, or eye-tracking models, every attention framework shares the same fundamental limitation: they measure the "what," but remain blind to the "why."

An attention model deployed at scale can calculate that a user engaged with or focused on a placement for 3.2 seconds, and vendors will eagerly point to correlations showing that those 3.2 seconds led to higher conversion rates or elevated AOV.

But correlation with an outcome still fails to explain what it was about the content that actually worked. Did the consumer buy because the visual hook grabbed them, because the offer was irresistible, or simply because they were already in a buying mood and needed a few seconds to read the fine print?

Attention and dwell metrics can flag which assets hold an audience's gaze, but they cannot diagnose the psychological trigger behind the conversion. Ultimately, a high attention score simply suggests that your creative had a higher statistical probability of being seen compared to standard viewability. Machine learning models can react to behavioral duration, but whether that attention was positive, memorable, or persuasive remains entirely unknown – and measuring that requires a different set of tools.

Vendor Bias and the "Outcome Gap"

If you read reports published by attention vendors, for example, this one, you’ll see claims of massive conversion lifts,  like significant increases in lower-funnel conversions when filtering for high Attention Units (AU).

While eliminating bottom-of-the-barrel, invisible ad inventory certainly improves media delivery efficiency, we must separate media quality from purchase motivation:

• Attention measures exposure quality: It ensures your ad isn't hidden at the bottom of an unread page.

• Attention does NOT create demand: You can secure 5 seconds of active dwell time, but if your price point is wrong, your offer is weak, or your targeting is off, that attention yields zero incremental sales.

The Playbook: Use It Upstream, Not as an ROI Proxy

Attention metrics aren't a replacement for revenue attribution or causal validation; they are an upstream filtering tool.

Smart growth teams don't report attention scores to their CFO to prove ROI. Instead, they use attention and dwell data quietly in media buying to stop paying for low-quality ad units—while relying on Causal modeling downstream to prove if that attention actually generated new dollars.

Here are 3 concrete actions you might consider:

1. Step 1: Audit Your Current Inventory (Upstream Clean-Up) Use attention metrics purely as a filter to cut out the bottom 20% of your programmatic placements – the "junk inventory" that technically passes 50%/1-second viewability rules but offers zero real dwell time.

2. Step 2: Run Creative Diagnostic Tests (Midstream Optimization) Use panel-based eye-tracking and interaction data early in a campaign to test creative layouts.  This answers specific execution questions like "Are viewers looking at our product shot or getting distracted by the background graphic?"

3. Step 3: Validate with Incrementality downstream (The CFO Test) Never hand an Attention Unit (AU) score to your CFO to justify budget. Feed your high-attention ad groups into a Causal/Incrementality experiment to verify whether high-attention placements actually drive lower CAC or true incremental revenue.

Are you treating attention scores as an upstream inventory filter, or are you expecting them to solve your attribution challenges? Let's discuss in the comments.