Ad creative testing: attention to the ad or to the brand?

We grouped examples by their released product-category label and page layout: left, right or full page. Each of the 59 eligible collections contained at least 20 examples. That gives a top-fifth list at least four places before allowing for ties. It does not make the archive representative of every advertiser. ^[2]

For each collection, we made two lists. The first used average looking time across the whole ad. The second used average looking time inside the brand region, such as its name or mark. A tie at the cutoff kept all tied examples, so a list could contain more than 20% of its collection.

The average overlap was 34.5% when starting with the whole-ad list. Looking in the other direction, the whole-ad list covered 33.9% of the brand list. The percentages differ because ties can make the two lists different sizes. Both figures give each collection equal weight.

Giving each collection equal weight prevents the largest archived categories from dominating this summary. It is a deliberate view of the collections, not an estimate of the average ad a customer encounters. The pooled alternative is reported in the method section so the reader can see how the weighting choice changes the result.

This was not uniform across collections. The share of whole-ad selections also selected for brand attention ranged from zero to 66.7%, with a median of 33.3%. The complete distribution is in the download; the average does not describe every category.

Brand looking time is part of whole-ad looking time, not an unrelated activity. Even so, a high total does not specify where inside the ad people looked. We compare absolute seconds for each region. We do not divide brand seconds by whole-ad seconds or claim that the same examples would lead on that different measure.

For a reference review, the useful distinction is between two selection questions. “Which examples held attention somewhere in the ad?” and “Which examples held attention on the brand?” can bring different material into the room. Neither list becomes a general list of the best ads.

We also compared the recorded brand looking time in the two shortlists while holding the number of selected examples fixed. When the cutoff was tied, we shared the remaining places equally across the tied examples. This avoids choosing a favorable tie winner by hand. ^[3]

Averaged equally across the 59 collections, the whole-ad selections had 0.536 seconds of recorded brand attention, compared with 0.822 seconds for the brand selections. The difference was 0.286 seconds; the median collection difference was 0.266 seconds. Individual collection differences ranged from 0.068 to 0.898 seconds.

The direction of that gap is built into the calculation. Selecting the highest brand values from these same records must produce at least as high a brand mean as another selection of the same size. The empirical information is the size and spread of the gap, together with the changed membership. It is not evidence that a new selection strategy would improve a campaign.

A separate evaluation would be needed to learn whether a selection method helps with the team’s actual outcome and audience. For now, the useful question is whether a reference was chosen for the quality the brief asks it to demonstrate. A visually arresting example can still be valuable, as long as its selection reason is clear.

Keep the units in the brief. These are seconds of looking time in a particular viewing task. They are not remembered brand names, purchase intentions, clicks or sales. A team interested in one of those outcomes needs evidence that actually measures it.

Before examining the shortlist results, we fixed two alternative cutoffs. Selecting the top 10% produced an average whole-ad-to-brand overlap of 25.4%. Expanding to the top 30% increased it to 42.2%. The corresponding recorded brand-time gaps were 0.371 and 0.219 seconds. A wider pool brought more shared examples, but the two objectives still did not produce interchangeable lists.

We separately grouped examples by the same released brand label and layout, retaining collections with at least 10 examples. In these 48 collections, containing 685 examples, the top-20% overlap averaged 40.5%, and the brand-time gap averaged 0.150 seconds. Some collections had identical selections. ^[4]

That second grouping is narrower, but it is not a comparison within verified campaigns or companies. We used the source labels as released, including labels that may describe an advertising page rather than a brand owner. Its role is to show how the result changes under another fixed grouping, without silently cleaning the labels to improve the story.

The category collections also differed by layout: whole-ad-to-brand overlap averaged 39.7% for left-page, 30.1% for right-page and 32.9% for full-page collections. These are descriptions of different archived examples, not effects of moving an ad. They do not justify a placement recommendation.

The released looking times are recorded to hundredths of a second. We therefore tested a deliberately small numerical change: allow every whole-ad and brand score to move independently by up to 0.01 seconds, never below zero, and recompute the selection cutoffs. ^[5]

Of the 348 examples selected only for whole-ad attention in the original comparison, 318 remained guaranteed whole-ad-only selections throughout those allowed changes. Of the 359 selected only for brand attention, 306 remained guaranteed brand-only selections. These are counts pooled across the 59 collections, unlike the equal-collection percentages above.

This check addresses closeness to the recorded cutoffs. It does not recreate new viewers or estimate sampling error. The release does not supply the individual viewing records needed to establish statistically resolved rankings. A stable numerical difference in these files is still a historical descriptive difference, not a promise that another audience would rank the ads the same way.

Imagine a small homeware brand commissioning a creator to introduce its new lamp. The team has room for two main references. The numbers in this example are invented to explain the decision; they are not five ads from the archive or a prediction for the lamp campaign.

Suppose the five reference cards have these whole-ad and brand looking times: A, 4.0 and 0.2 seconds; B, 3.5 and 0.8; C, 3.0 and 1.0; D, 2.0 and 0.4; E, 1.0 and 0.1. Choosing the two highest whole-ad values selects A and B. Choosing the two highest brand values selects C and B.

Only B appears on both lists, giving 50% overlap in either direction. The first pair has an average brand time of 0.5 seconds; the second has 0.9 seconds. The 0.4-second difference follows from the invented inputs. It is not an estimated benefit from changing a real campaign.

The two-place limit matters too. A manager could combine both pairs into A, B and C, but that would expand the review rather than resolve the original two-reference decision. Keeping the limit visible makes the tradeoff explicit. The team can either choose the objective or deliberately make room for more examples, instead of treating two different lists as though they answered one question.

The useful discussion is about the assignment. If the references are meant to demonstrate how the brand appears within a composition, the team can inspect B and C for that specific purpose. If A is retained for another reason, the brief should name that reason rather than call A the overall winner.

A creator may still find a technique in A worth adapting. The study does not prescribe a bigger logo or a particular visual style. It supports keeping the selection objective visible, so an interesting composition is not mistaken for evidence about brand attention. For an unmeasured reference, write “selected for its composition; brand attention unknown.”

Start with the decision: “We are selecting reference ads to help us decide ______.” Complete the sentence before assembling the final folder. A reference for composition, a reference for showing the product and a reference with measured brand attention need not be the same image.

Name the evidence: “This reference was selected because ______. The supporting measure or observation is ______.” Record whether the reason comes from measured viewer behavior, a campaign result or the team’s visual judgment. Keep those descriptions separate; a popular saved post does not acquire eye-tracking evidence because it is in the folder.

State the comparison: “The candidate pool is ______; we will retain ______ examples; ties will be handled by ______.” Agree on comparable format and context before ranking. The archive’s 20% rule is our research choice, not a recommended campaign budget or universal shortlist size.

Explain each retained example: “Borrow ______; do not assume ______.” This is where the creator gets usable direction. For the fictional lamp brief, a note might ask the creator to study how the product and brand fit into the composition, while leaving the campaign’s actual attention and sales results unclaimed.

Leave unknowns visible: “We do not have evidence for ______.” If an example comes from a public ad library or a creator’s portfolio, its design may be visible while audience attention is not. Record the available source and the reason for choosing it. This keeps an inspiring reference useful without giving it an invented measurement.

Assign the next check: “______ will review ______ before ______.” If the team will run an audience test, define the outcome and test conditions separately. If it will only review the design, describe it as a design review. Keep the reference-selection evidence attached to the brief so a later handoff does not turn a narrow observation into an unsupported performance claim.

The source is the pinned AdGaze-3500 release. It contains 3,531 ad-and-context examples with finite, nonnegative whole-ad and brand looking-time means. The primary analysis retains 2,601 examples in 59 category-and-layout collections with at least 20 examples each. We did not treat the unselected 930 examples as missing scores; they fell outside the fixed collection-size rule. ^{[1] [2]}

The viewing task used static magazine ads with surrounding editorial material on desktop screens in Dutch research facilities. Participants browsed pages at their own pace. The exact collection dates are not supplied; part of the underlying material had appeared in research published in 2004. The 2026 dataset release therefore should not be read as a study of 2026 audiences or Meta advertising. ^[6]

Within each collection of n examples, the cutoff rank is the ceiling of the selected fraction multiplied by n. Membership comparisons include every tie at that rank. We calculate both directional overlaps within each collection, then average the ratios with equal collection weights. Pooling the primary lists instead gives 207 shared selections out of 555 whole-ad selections, or 37.3%, and 207 out of 566 brand selections, or 36.6%. Those answer a different weighting question from the headline.

For the fixed-size brand-time comparison, examples above the cutoff receive full weight, while the remaining places are shared equally among cutoff ties. We also calculate the smallest and largest gaps attainable by choosing among those tied examples. Averaged across the primary collections, that range is 0.285 to 0.288 seconds. It describes tie choices, not a confidence interval. ^[3]

Earlier research already distinguishes attention to an ad from attention to its brand. The AdGazer paper also reports their relationship and studies attention prediction and context assignment. Our contribution is the specified reference-shortlist comparison using the released observations, including directional overlap, fixed-size recorded differences and numerical cutoff checks. We make no claim to have discovered the general distinction. ^[7]

The source provides aggregate viewing means rather than viewer-level records or uncertainty for each example. We therefore report descriptive results without confidence intervals or significance tests. The downloads retain all predefined group results, source hashes, grouping rules and calculation instructions, including the fixed comparisons that did not become headline findings. ^[8]