TL;DR
This paper investigates how to optimize the scheduling of online display ads to maximize user memory and advertiser value. They found that showing two shorter ads back-to-back is more effective for memory recall than a single longer ad, especially when the ad slots are short. This suggests a shift from impression-based pricing to time-based models could benefit both advertisers and publishers.
What Problem Does This Paper Address?
Online display advertising is a massive industry, with advertisers aiming to boost brand recognition, awareness, and sales. Traditionally, the effectiveness of these ads has been measured by memory metrics, and ads are sold based on CPM (cost per thousand impressions). However, previous research has shown that the duration an ad is visible (exposure time) significantly impacts whether a user remembers it, with the initial seconds being most crucial. This raises a key question for advertisers and publishers: how should time-based display ads be scheduled to maximize their impact on memory and, consequently, their effectiveness? The paper specifically tackles the problem of optimizing ad scheduling within a given time slot to achieve greater total recollection.
Methods
The researchers designed an experiment to compare different ad scheduling strategies. Participants were shown an online news article with display ads in the sidebar. They were randomly assigned to one of four "time treatments," which varied the duration and number of ads shown. The core comparison was between showing a single ad for a longer duration (e.g., 20 seconds) versus showing two shorter ads sequentially for half that duration each (e.g., two 10-second ads).
To measure ad effectiveness, they used three memory metrics:
- Unaided Recall: Participants were asked to list any advertisers they remembered seeing.
- Textual Recognition: Participants were asked specific questions about seeing certain ads (e.g., "Did you see a Netflix ad?").
- Visual Recognition: Participants were shown ad images and asked if they recognized them.
They also incorporated "lure ads" – visually similar ads to those shown – to check for false positives and ensure participants were remembering the specific advertiser, not just a general visual characteristic.
Datasets
The study used a sample of 1,100 participants from Amazon's Mechanical Turk. These participants were screened for approval ratings and to ensure they participated only once. The ads used in the experiment were from actual Yahoo! news articles and included advertisements for Netflix, Jeep, Avis, and American Express. No new datasets were introduced; the study utilized existing online advertising formats and participant pools.
Key Insights and Ideas
- The Power of Sequential Exposure: Showing two shorter ads back-to-back (e.g., 10 seconds each) leads to significantly higher total memory recall than showing a single ad for twice the duration (20 seconds). This suggests that the opportunity for a user to notice and process information from multiple distinct ads, even if brief, is more impactful than a single, longer exposure.
- Duration Matters, But Not Indefinitely: While longer exposure times generally increase memory, the benefit tapers off. Crucially, the advantage of sequential ads diminishes as the duration of each individual ad increases. This implies there's an optimal "sweet spot" for ad duration within a time slot.
- Onset Time is Critical: The time between a page loading and an ad appearing (onset time) significantly affects memory. Ads with longer onset times are less likely to be remembered. This reinforces the idea that the second ad in a sequential pairing suffers from a later onset time, but the benefit of having two ads still outweighs this disadvantage for shorter durations.
- Implications for Pricing Models: The findings strongly suggest that current CPM (cost per impression) models are suboptimal. A time-based pricing model, where advertisers pay for the duration their ad is displayed, could better reflect actual effectiveness and create a win-win for advertisers (better memory) and publishers (more ad slots, potentially higher revenue).
Key Results
- For short ad durations (10 seconds), the sum of memory rates for two sequential ads (AB + BA) was significantly higher than for a single longer ad (AA + BB) across all three memory metrics (visual recognition, text recognition, and recall). For example, visual recognition summed to 0.58 for the two short ads versus 0.41 for the single longer ad.
- This advantage of sequential ads disappeared when the ad durations were longer (20 seconds). In this case, the sum of memory for two 20-second ads was comparable to a single 40-second ad.
- The benefits of shorter, sequential ads held true even when analyzing individual advertisers, suggesting the effect is not just an average but applies at the advertiser level.
- False alarm rates (incorrectly remembering an ad) were low and unaffected by the visual similarity of "lure ads," indicating participants were remembering specific advertisers.
Limitations and Future Work
- Limited Duration Range: The experiment focused on specific durations (10+10 vs. 20, and 20+20 vs. 40). The paper notes that the effect of sequential ads likely diminishes beyond the second ad due to increasing onset times, but further research could explore optimal scheduling for more than two ads or different duration combinations.
- Generalizability of "Lure Ads": While lure ads helped confirm specific ad recall, the paper acknowledges that the "lure" effect itself was not a primary focus. Future work could delve deeper into how visual similarity impacts memory and false recall.
- Real-World Complexity: The experiment was conducted in a controlled environment. Future research could explore how these findings translate to the dynamic and often interruptive nature of real-world web browsing, including user behavior like ad blocking or rapid page navigation.
- Publisher Benefits: While the paper suggests publishers benefit from more ad slots, a deeper analysis of how time-based pricing models would impact publisher revenue and ad inventory management would be valuable.
Summary updated