How to Read an Attention Heatmap and Know When to Trust It

An attention heatmap on an ad is a model's estimate of where viewers' eyes will land in the first seconds, drawn in warm colors (red, yellow) for likely spots and cool colors for unlikely ones. The estimate comes from a model trained on eye-tracking studies, where cameras record where people look. Nobody looked at your ad to produce it.

Models of this kind pick up on contrast, faces, large text and isolation against a plain background. A heatmap can inform five questions about a static creative, the same five our free ad attention heatmap tool addresses:

• What gets noticed first?
• Is the message visible?
• Is the face helping?
• Can people find the call to action (CTA)?
• Is the brand visible?

It is the wrong tool for six other questions: clicks, conversions, return on ad spend (ROAS), brand recall, message comprehension and purchase intent. Those need behavior, and a prediction of gaze cannot supply it. In machine learning the same term means something different, covered in the last section before the FAQ.

Only once a figure names its metric and its test images. Of four vendor pages read in October 2026, three name no metric and the fourth names correlation without naming a dataset. The claims run from about 90 to over 95 percent, and without a metric none can be compared with anything, including a trivial baseline.

We publish this page and also offer a free attention heatmap tool, so we are part of this market. The table covers four other providers' own claims.

Wording quoted from each vendor's public page, read October 2026. Neurons also describes holding back part of its own eye-tracking database to validate its model. That is a standard method, and it shows how well the model predicts more data like its training set.

The metric that matters is correlation. The correlation coefficient (CC) compares a model's heat map with the heat map of real eye fixations, the points where eyes pause, pixel by pixel. A score of 1 means identical and 0 means unrelated. The MIT/Tübingen Saliency Benchmark is a public leaderboard that scores models on saliency (how much a part of an image stands out) against recorded eye movements on natural photos. Its best listed model scores 0.88, and its gold-standard model, built from the viewers' own fixations, scores 0.98. A claim of 90%+ correlation is higher than the best model listed there. Without a named dataset and metric it cannot be compared, and it should not be taken as evidence about ads.

Ads and screens are harder than photos. The UIGaze study ran nine models over 1,980 screenshots across webpages, desktop screens, mobile screens and posters. Agreement varied by screen type and improved with longer viewing, and the models matched exploratory looking better than the first fixation. These were general models, not purpose-built attention tools, so the result shows how hard the task is, not how any one product scores. Our inference: of the five questions above, "what gets noticed first" is the weakest, so read the overall spread of heat instead of the first-glance ranking.

Center bias is viewers' tendency to look near the middle of an image. A map built from that tendency alone scores a correlation of 0.45 on the MIT/Tübingen benchmark, well below the best models but far from zero. Any element placed centrally looks hot whether or not your design earned it.

That baseline knows nothing about the image it scores. It is fitted from where people looked in other images. So the real value of a heatmap sits where it departs from the center: a headline pulled upward, an offer badge off to one side, a face. A hot product in the middle of the frame is weak evidence by itself.

Our recommendation is to judge every must-see element against what a center-only map would have given it. A center-only map is a heatmap built from center bias alone. The check below does exactly that.

Divide each must-see element's share of the heat by the share a center-only map would give it. A ratio of 1.5 or more passes, below 1.0 fails, and anything between is unclear. We chose 1.5 by judgment, not by calibrating on real ads: it means the element gets half again as much heat as its position alone explains. Ratios near 1.0 give no evidence either way.

Heat share is the fraction of the map's total heat that falls inside an element's box. The center-only share is the same measure on a smooth bell-shaped blob centered on the image, with a spread of 25% of the width and height. That blob is a simple stand-in for the fitted baseline in the benchmark, so treat ratios as a guide, not a lab result.

import sys, json
import numpy as np
from PIL import Image

heat = np.asarray(Image.open(sys.argv[1]).convert("L"), float)
heat -= heat.min()
heat /= heat.sum()
H, W = heat.shape
y, x = np.mgrid[0:H, 0:W]
center = np.exp(-(((x - W / 2) / (.25 * W)) ** 2 + ((y - H / 2) / (.25 * H)) ** 2) / 2)
center /= center.sum()

boxes = json.load(open(sys.argv[2]))
print("element    area%  heat%  center%  ratio")
for name, (x0, y0, x1, y1) in boxes.items():
    area = (x1 - x0) * (y1 - y0) / heat.size
    h = heat[y0:y1, x0:x1].sum()
    c = center[y0:y1, x0:x1].sum()
    print(f"{name:9} {area:6.1%} {h:6.1%} {c:7.1%} {h / c:6.2f}")

A worked example from our own model. We ran the script on a made-up 1,080 by 1,350 pixel layout (the 4:5 feed ratio) and a made-up map: a center-only blob plus three extra hot spots. It illustrates the arithmetic and is not data from real ads. The four boxes cover 23.0%, 11.4%, 2.9% and 4.0% of the frame.

On a center-only map, the product holds 47.7% of the heat in 23.0% of the frame, about twice its area, with no help from the design. In the toy map the offer badge passes at 2.08 and the headline is borderline at 1.42. The product's red glow is mostly position, at 1.04. The CTA fails at 0.57, because it gets less heat than a center-only map would have given it.

Fix position and prominence first, then rerun the center-bias ratio. A failed element usually needs to be larger, higher in contrast, or moved away from a competing hot spot. Changing one thing per rerun shows which change worked.

A decoration here means any element that is not must-see, such as a background shape or a secondary logo. A heatmap cannot score offer clarity, proof or hook strength. For those, use the creative audit scorecard.

No. A change in predicted attention shows that a model's output moved after you edited the design. An A/B test shows what real people did with each version. Only the second says anything about clicks, conversions or sales.

A heatmap sits at the start of a chain: seen, understood, clicked, bought. An ad can pass the check and still fail because the offer was the problem. For video, hook signals measured from real viewers answer a different question than predicted gaze.

Our recommendation is to use the heatmap as a filter before spend. Drop variants where a must-see element fails, then test the survivors against a fair benchmark; how to choose a creative test control ad covers that step.

That is how we approach it at Deepsolv. The free heatmap tool handles the pre-flight read. The Deepsolv creative intelligence platform handles what comes after: it combines ad-performance data, competitor activity, customer signals and past ad learnings into ranked weekly creative test plans and execution-ready briefs for teams running Meta ads. Pricing is shared in a personalized demo with our founding team, so book a Deepsolv demo and ask.

No. In machine learning an attention heatmap is a grid of weights showing how strongly one token (a word piece) or image patch draws on another inside one layer (one stage of the network) and one head (one of several attention calculations run side by side in that layer) of a transformer, a neural network built on attention (Vaswani et al., 2017). It shows how information flows inside a model, not where a person looks.

Each row is one query (the token asking) and each column one key (the token being looked at). A softmax, the step that turns scores into weights, makes every row sum to one. A bright cell means a large weight in that head and layer, not that the token caused the prediction. Experiments on text classification found attention weights frequently uncorrelated with gradient-based measures of feature importance (a way of testing which inputs changed the output most). They also found very different attention patterns that give the same prediction (Jain and Wallace, 2019). Use these maps to inspect and debug a model, and confirm any claim about cause with another method.