How Glanzz scores attention

When you submit an image or URL, the engine builds a saliency map — a per-pixel estimate of how likely each spot is to draw the eye in the first few seconds. At its core is a neural attention model trained on real eye-tracking of 62 people looking at 1,980 websites, apps and posters (the UEyes study, credited below). How much classic visual cues — text & CTAs, faces, contrast and layout — should add on top was measured on eye-tracking rather than set by hand: adding most of them made predictions no better or worse, so only faces and contrast keep a small extra weight. Every score and overlay is derived from that one map.

It runs on our own servers — seconds for a single screen, longer for a long page, no per-image cost — so the numbers are consistent and repeatable for the same input. When an engine update changes how scores are calculated, averages, percentiles and monitor alerts only compare analyses from the same engine generation.

We test Glanzz against real eye-tracking on screens it has never seen: the 108 screens the UEyes study set aside for testing (27 each of websites, desktop apps, mobile apps and posters). They were used neither to train the model nor to tune the blend (they do set the Impact Score's scale, below). The score is CC, the correlation between Glanzz's predicted heatmap and the heatmap of where people actually looked in their first 3 seconds: 1 is the same pattern, 0 is unrelated.

For specialists, at 3 s: CC 0.695 · KLD 0.643 · SIM 0.585 · NSS 1.638 · AUC-Judd 0.844 (MIT saliency benchmark definitions). The previous engine's model may have seen some of these test screens while it was trained, so the real improvement is larger than the table shows.

What this covers: single screens, viewed freely in a lab. Long pages use the same model screen by screen, plus a scroll curve from separate research (below). The test set has no ads or packaging — posters are the closest — and a good average still leaves individual designs where the prediction is off. That is what the People test is for.

A full-page capture is analysed the way people see it: one screen at a time. The page is cut into overlapping screen-sized slices — a desktop screen, or a phone screen for mobile captures — each slice is analysed at full resolution, and the slices are joined back into one map (up to ten screens; a longer page shows its first ten).

The model expects people to look at the top of whatever screen it is shown. That is right for the first screen, but further down people see each part of the page at many scroll positions, so below the first screen that within-screen bias is taken out: the same content gets about the same attention wherever a slice happens to start.

Attention then fades down the page: the first screen keeps its full weight, attention drops at the fold and keeps falling screen by screen. The curve is set so that a page of equally interesting screens puts about 57% of attention in the first screen and 74% in the first two, in line with Nielsen Norman Group's eye-tracking research on scrolling; your page's actual content moves those numbers either way.

For long pages, Focus, Clarity and the other scores describe the first screen — what every visitor sees — so they stay comparable with single images. The heatmap, the elements and the look order cover the whole page.

The headline Impact Score (0–100)is a weighted blend of five parts, and every part is scored against real designs: Glanzz's predictions for the 108 UEyes test designs. On each, 50 is the typical design, 20 and 80 mark the bottom and top tenth, and scores curve towards 0 and 100 beyond that without reaching them. The blend is anchored the same way, so 50 is a typical design and 70+ is about their top fifth. Every map is measured at one standard size, so a design scores the same at any upload size.

The weights are the first formula's, scaled by how reliably Glanzz predicts each part: measured on Glanzz's map and on the real eye-tracking of the same 108 designs, the parts agree as the last column shows (rank correlation, 1 = same order). The Impact Score itself agrees with the score real eye-tracking would give at 0.61, measured on the same 108 designs that set the scale, so on new designs it may agree a little less. Nothing hidden — these are the exact weights:

Verdict bands: Strong 70+ (about the top fifth of those designs), Solid 50–69 (better than a typical design), Mixed 30–49, Weak under 30. The test designs are websites, apps and posters — no ads or packaging — so for those the scale is the closest available. Where you see a percentile in-app, it's computed against your own workspace history of the same asset type. Analyses saved before 14 October 2026 were rescored from their stored attention maps (those made before 2 October by the previous engine against that engine's own predictions for the same 108 designs); videos from before then keep the first formula, which ran low (most designs scored 25–40).

A high Impact Score is not automatically right for every design. An ad that sells needs the eye on its button and its offer; a page that informs needs the headline read first and attention spread over the parts that matter; a sign-up page needs its form found; a brand ad needs its logo seen. Pick the design's goal (Sell, Inform, Sign up or Brand) and the Goal Score (0–100) checks exactly that. It leads the report, with the Impact Score underneath. Ads and emails start as Sell; for a website you pick the goal; you can change it on any report.

How the elements are found. The engine outlines every element on the scored screen (the first screen of a long page). An AI vision model then labels what each one is: button, price or offer, product, headline, navigation, logo, form or face. It also adds a key element the outlines missed. You can correct any label or mark an element by hand on the report, and the score updates at once.

What is measured and what is chosen.Measured, from the same attention map as every other score: each element's share of attention, its density (its share divided by its share of the area, so 1 is average for its size) and which of the first six predicted looks reaches it. Chosen by us: which elements count for each goal, the weights and the cut offs below. They are fixed rules, not yet calibrated against real results; once a goal has enough analyses we will re-anchor them.

Sell and Sign up also lose up to 15 points when one element that is not part of the goal holds more than half of all attention. Bands as on the Impact Score: Strong 70+, Solid 50–69, Mixed 30–49, Weak under 30. The Goal Score says whether attention reaches what the goal needs; it does not predict clicks or sales.

An ad is never seen alone. In contextdraws your creative into the placement where it will run: the Instagram and Facebook feeds, Stories, Reels, TikTok, LinkedIn, YouTube, news sites on desktop and phone, Google Shopping, an e-shop grid, a store shelf or an opened email. Each screen is drawn at its real size (phones at 390×844 points, twice over; desktops at 1440×900) with real text, icons and stand-in pictures that have the colour and detail of real posts, products and ads, and the creative is fitted the way the platform fits it: feeds crop anything taller than 4:5 or wider than 1.91:1, display banners have fixed sizes, Stories fill a 9:16 screen. For the real surroundings, upload your own screenshot and mark the ad's spot.

What is measured.The engine predicts attention for the whole screen. From that map: the share of the screen's attention on your creative, the share of the screen it covers, the first of the six predicted looks that lands on it, and where the rest goes (the rest of your ad, other posts and products, the page's content, the app's buttons).

The stand-out score (0 to 100).A big feed post always gets more of the screen than a small banner, so the raw share can't be compared between placements. The score places your creative between two anchors measured in the same spot of the same screen with the same engine: 0 is what a plain grey box gets there, 100 is what Glanzz's bold reference ad gets there (a dark background, a big yellow headline, a product and an orange button), measured at the creative's own shape. 70 and over stands out, 40 to 69 holds its own, under 40 gets lost. Where the two anchors are less than 2 points apart (a phone's bottom banner gets almost no attention whatever it shows) there is no score; the share still compares creatives in the same spot. The reference ad and the bands are our choices.

Full-screen placements.In Stories, Reels and TikTok the ad is the whole screen and gets the attention by default; what matters is what the app puts on top of it. Glanzz also analyses the creative on its own and measures how much of its attention falls under the app's header, caption and buttons: under 8% is clear, 8 to 18% partly covered, more is key parts covered.

Limits: the surroundings are drawn stand-ins, and real feeds (with faces and photos) can compete harder; the prediction is for the first seconds on one screen and does not model scrolling.

On top of the raw scores, Glanzz turns the same attention map into the calls you actually make:

• Where attention lands — each detected element ranked by its share of attention, plus a density multiplier (×1 = an average pixel, ×5 = it pulls five times its weight). Elements are named by an AI vision model ("Primary CTA button", "Hero headline") — a convenience that can occasionally be wrong; the shares always come from the attention map.
• Order of attention — the predicted gaze sequence (peak → peak with inhibition-of-return), mapped onto your elements, so you can see if the CTA is seen early or late.
• Above the fold — for tall pages, the share of total attention that lands in the first screenful (a desktop or a phone screen, depending on the capture).

Attention works differently per medium, so your asset-type choice scopes your benchmarks and tailors the AI recommendations to that medium: thumb-stopping power for advertising, hero/CTA/nav scanning for websites, the top fold and a single CTA for email, shelf standout for packaging. It only nudges the attention map itself (the weight of faces and contrast): hand-set per-medium cue weights made predictions less accurate on eye-tracking data, so the measured model does the work for every medium. When a website or email upload is taller than a screen, Glanzz asks whether it was captured on desktop or mobile and analyses it as a long page of that screen.

Any analysis can be checked against real people. Test with people creates a public link to a short test — no sign-up. Participants see the original image blurred for 10 seconds and click (or tap) wherever they would look; each click brings a small circle into focus. This is the BubbleView method (Kim et al., 2017): to see detail people have to click, so their clicks follow their attention, and studies found its click maps closely approximate eye-tracking fixation maps. It's a validated proxy, not eye-tracking.

The clicks become a heatmap on the People tab (each click spread as a Gaussian about 3.5% of the image width), next to Glanzz's prediction — from all clicks, or from each person's first three, which is closer to the first look. Agreement with Glanzz is the Pearson correlation (CC) between that map and Glanzz's predicted attention map, compared on a small grid: 1 is the same pattern, 0 unrelated. Under 0.3 reads as weak, 0.3–0.5 moderate, 0.5–0.7 good, 0.7 and up strong.

The figure appears from 8 people — below that, a few clicks swing it too much — and the more people (and the closer they are to your audience), the more it can be trusted. Participants' answers hold only click positions and timing, screen size and device type: no names, e-mails or IP addresses. You can try the test yourself in preview (nothing is saved), and clear the answers — practice runs, say — at any time; the link stays the same and everyone can answer again.

The 5-second memory testis a second, separate link. People see the creative for 5 seconds; then it disappears, they pick the brand from four names (the right one and three the team chose) or "I don't know", and say in a few words what it was about. The report shows how many picked the right brand and lists what they typed, as written; with fewer than 10 people it is a small sample. How hard the question is depends on the other three names, so use real brands from the same category, and send the two tests to different people (someone who did the click test has already seen the image). Its answers hold only the name picked, the words typed, timings, screen size and device type.

The attention model is trained on UEyes (Jiang, Leiva, Tavakoli, Houssel, Kylmälä & Oulasvirta, Proceedings of CHI 2023), an eye-tracking dataset of 1,980 user-interface screenshots viewed by 62 people, used under the Creative Commons Attribution 4.0 licence (dataset). The People test uses the BubbleView method (Kim et al., 2017).

It predicts attention, not comprehension, emotion, brand recall, or whether someone will convert. It doesn't read your copy's meaning or your offer. Use it to make sure the right things get seen, then let messaging, proof and price do the rest.