Wasted video ad impressions, lost revenues & how to fix video ad tech deficiencies [insight]
Video is an ever increasingly important revenue line for publishers, with video making up 18% of digital display investment in Europe equating to over $3 billion spend (according to IAB Europe) and $15 billion in the US (according to eMarketer). But making money from video isn’t easy. Yes, you can plug in some code and access demand from DSPs, but as you dig deeper into video ad ops you become increasingly aware of the reality of the inefficiencies of video ad tech stacks - which inevitably mean lower revenues (and lost revenues).
Many of these inefficiencies are caused by video ad tech infrastructure deficiencies. Hackers.Media sat down with Vishak Nag Ashoka Director of Growth Engineering at Polymorph to discuss the issue of wasted video ad impressions and ideas of how publishers can solve the problem . . .
Q: What’s your view of what’s happening with lost video ad impressions?
A: As the video adtech ecosystem has grown, so has the problem with lost ad impressions. The core of the problem is down to video adtech infrastructure based around VPAID tag performing ad auctions on the client side within the publisher video player, therefore leading to delayed video loading speeds and ads timing out - ultimately resulting in lost ad impressions.
Q: What are the main causes of lost video ad impressions?
The main causes of lost video ad impressions boil down to four main reasons:
> Video Ad Network latency : demand partners taking too many seconds to respond
> VPAID Errors : Creative failing to load due to Flash version mismatch, bad creative etc
> VAST Redirects : Multiple level of VAST redirects creating unnecessary complexity
> Multiple Auctions : demand partners running a second, third, or even fourth client-side auction after VPAID tag is loaded into video player
Q: What’s the background to VPAID & why’s this happening?
A: The VPAID standard was originally invented by the IAB as a way to add interactive capability to video advertising content; this was a great idea at the time which was widely adapted. The only issue is that the VPAID specification has been used in a variety of ways that its originators never intended. For examle, VPAID is used to conduct client side auctions and to measure video viewability. These are tasks t at are not explicitly defined by the IAB VPAID standard.
Around 2011 / 12, the approach of using VPAID for performing client side auction took off. More vendors starting running the auctions and taking the concept further: if they couldn’t fill the video impression they’d pass it off another company that would run another auction. The problem created here being that vendors would layer multiple auctions to the point where you are waiting and waiting to watch your video because your advertising vendor is running a second, third, or even fourth client-side auction, each with its own set of bidders. Some video players are sophisticated enough to understand there’s no creative or there is a VPAID error, thus the player will call on a different ad source in the waterfall.
All of this back and forth adds latency, which is the number one reason for a user to abort their session or close the video player. Because of this, the publisher usually gives up on the ad impression after the latency has crossed a certain time-based threshold and ends up showing the content instead, therefore wasting a valuable ad impression.
Q: What are the options for solving this problem?
A: The solutions are out there already for for the video monetization industry to take steps collectively to create a Better Video Internet. These need to be changed at the core layer of video technology and need to be adopted by all the ad networks, publishers, video players to make the ecosystem more reasonable and transparent (away from VPAID auction black box). We are speaking of long term here (new standards to be built, accepted and adapted), and this will not stop Ad networks from using the current VPAID solution and publishers cannot move away from VPAID as they need the revenue.
The top recommendations out there to save wasted impressions and improve user experience are roughly three-fold:
1) Move the auctions to server side - RTB, Server-to-server video programmatic connections
- Majority of the video ad networks mandate publishers to use their programmatic VPAID because of the flexibility of client side auctioning they can run.
- Top video ad networks do not support OpenRTB for the same reason
2) Develop new IAB standards for client-side auctions
- This is huge initiative that Video ad industry should take up. Once built, it is years away from being adapted at scale in the wild.
- Example VAST 4.0 was released almost 2+ years ago which decouples interactive element from creative for VPAID but is yet to be adapted by any major Ad network at scale. Networks are still on VAST 2/3
3) Configurable timeouts per partner :
- By setting a timeout for each video ad tag in the waterfall process, if a tag does not provide a result quick enough they need to stop and signal a opportunity to the next. This allows publishers to try more partners before wasting the impression.
Q: Out of these, what would you recommend to do practically?
A: Assigning a priority/position in video waterfall to each video ad network (VAST/VPAID Tag) based on demand fill rate, eCPM is the most widely used setup by publishers to maximize their video programmatic revenue. Few publishers who have sophisticated video adtech also set a custom timeout for each network after which they give up and move on to the next one. This approach is definitely better than just indefinitely waiting for the tag is load or signal an error/no-creative.
This setup is still is not efficient because you try loading all the networks in the order of priority no matter what user, geo, device etc is this impression coming from.
Q: And beyond this, given you specialise in applying AI to ad tech problems, what’s your ideal solution based around Machine Learning (i.e. using the ability to be smarter)
A: If you really think through the problem and consider using Machine Learning to solve the problem, the first factor to consider is that it’s obvious that all demand partners will be different (i.e. each demand partner will have different strengths when it comes to demand from different geo, devices etc.) and therefore not all of them will want to bid for the same user or at the same price.
Using Machine Learning you can capture the bidding pattern of each demand partner and predict the probability of a particular ad tag returning an ad for the current geo, device, user etc. If the probability is lower than the minimum you’ve set, then it will signal to the player to SKIP. sending a request/loading that particular VAST tag.
Beyond learning ‘bid patterns’ the second signal you could train your system to learn from is predicting VAST errors. In a similar way using Machine Learning you can capture the error pattern of each demand partner and predict the probability of a particular Ad tag returning an ad at the current time and SKIP sending a request/loading that particular VAST tag if the prediction engine says so.
By using Machine Learning to create a prediction engine that to identify bid patterns and error patterns will save valuable seconds, meaning that the bid can be given to the next eligible partner who might actually fill with an ad, reducing the wasted impression % and improving user experience (as load speeds will be increased).
Q: Could publishers add ML solutions like the one you’ve just described to their existing ad tech stack?
Yes - what I’ve just described isn’t talking about ripping up a publishers existing ad tech stack, but adding an ‘intelligent layer’ on top - to help the existing stack make smarter decisions and ultimately make more money for the publisher.
In technical terms, this can be done by plugging in an Machine Learning API into video player dataset - enabling it to train in near real-time. The dataset will include both the VAST errors and the Bidding patterns from each of the partners; these two data points will be used as the input “features” and the prediction engine will come up with a fill-probability score as the output, which will then be returned to the video player adtech module. Finally the publisher’s adtech module will then do the simple decisioning of which partners to skip and which partners to try for a video ad opportunity based on the fill-probability score.
Thank you for your time Vishak. If you want to learn any more from Vishak's insights, contact him at GetPolymorph.com