Glance

How to use an attention heatmap for landing pages

Use an attention heatmap for landing pages to check whether the headline, image and primary action compete for early attention. Start with one screenshot and one question. Treat the result as a design hypothesis, then test comprehension and behavior with real visitors.

Open Glance with a screenshot you can analyze. No Glance account is required.

1. Define the decision before opening the map

Write one sentence: “Does the decorative image compete with the signup action?” This gives the review a purpose. Without a question, a bright overlay can become decoration in a presentation. Pick the viewport your audience actually uses and capture the fully loaded page in that state. Keep the original screenshot for comparison.

Do separate checks for desktop and mobile layouts. A narrow viewport can move the action below the fold or change which image dominates. Do not resize one screenshot into another device shape and assume it represents that layout. Avoid putting confidential customer information into files used for a demonstration.

2. Inspect the image, then the prediction

Before generating a map, identify the headline, primary action and strongest competing element. This makes your expectation explicit. Run the image through a supported model and record its name. Compare the overlay with the original image rather than judging the colors alone. A heatmap shows a distribution of predicted attention, not a recording of anyone’s eyes.

If your workflow supports areas of interest, draw consistent boundaries around the elements you want to compare. Predicted attention share depends on the region you draw: a larger region can include more map mass. Keep those boundaries comparable across versions and do not present the share as a click probability. Glance’s science page explains its formulas and limits.

3. Change one visible element

Try reducing the competing image’s contrast, simplifying surrounding decoration or making the action easier to distinguish. Keep the rest of the page fixed so you can explain the change. Save the original and revised files with the same crop and dimensions. Re-run the same model with the same settings.

For example, compare a hero with a large decorative badge against one with a quieter badge. Ask whether the predicted distribution now supports your intended hierarchy. This is a worked review procedure, not evidence that either version converts better. A viewer may notice a button and still reject its offer or misunderstand its label.

4. Check what the map cannot tell you

A saliency model does not know your visitor’s budget, intent or familiarity with the brand. A screenshot cannot represent keyboard navigation, form validation, hover states or a slow network. Check those directly. Keep text readable and controls accessible even if a map suggests a subtler treatment would attract less attention.

Research also distinguishes visual importance in graphic design from generic saliency. Fosco and colleagues’ graphic-design study examines importance across design types. The UEyes workshop derivative provides interface-domain context. These sources support asking whether a model fits the material; they do not validate an individual sales page.

5. Take one hypothesis to a real test

Ask participants what the page offers and what they would do next. Watch whether they can complete the intended task. If you run an experiment, define the outcome and analysis before looking at results. Use observed behavior to decide whether the visual change helped. A map can identify a question cheaply; it cannot supply that answer by itself.

Keep the screenshot pair, model identity and your original question beside the result. For benchmark claims, Bylinskii and colleagues explain why AUC, CC and NSS measure different properties. Read the accuracy guide before treating any headline percentage as a forecast for your visitors.

Next review: try the manual squint test on the same image and compare what each method notices.

Frequently asked questions

Can a heatmap predict conversion?

No. Predicted visual attention does not establish clicks, comprehension or purchase behavior.

Which screenshot should I use?

Capture the intended viewport and loaded state. Review mobile and desktop layouts separately.

What does area attention share mean?

It is the predicted map mass inside your selected region. Boundaries affect the result; it is not click probability.

Should I change several elements?

Start with one element so you can explain which change affected the predicted distribution.

How do I validate a revision?

Observe task completion with intended users or run an experiment with a predefined outcome.

Choose your next check

Try in browserDownload