Dynamic floor optimisation & how to optimise revenue across exchanges more effectively [interview]

Dynamic floor optimisation & how to optimise revenue across exchanges more effectively [interview]

With revenue optimisation options for publishers becoming increasingly limited, and publisher ad ops and revenue teams being squeezed for resource and time, the idea of looking to Artificial Intelligence to scale smarts and identify revenue optimisation opportunities which the human naked eye / brain might not spot has become ever appealing.

Hackers.Media sat down with Satish Polisetti CEO of Polymorph to discuss current publisher ad stack model failings and how publishers could optimise revenue further by applying dynamic floor optimisation more smartly.

Q: Give an overview of the failings of existing publisher stack set-up and potential revenue opportunities.

A:  The definition of who is an exchange and who is a bidder is changing. The acronym of ‘SSP' is slowly fading away, as to a publisher you are either someone who brings them money or someone who brings technology. If you are bringing money, it doesn't matter what you call yourself. DSPs, SSP, ad-network, bidder are all the same to a publisher. 
Publishers have trusted that their SSPs would do the best job of managing ad-networks and RTB bidders. And indeed SSPs did do the best job until it didn’t matter / their job became limited in value.  As the number of ad networks in the world decreased, the true value add of exchanges reduced. The job of sending a bid request out, getting bids and picking a winner has now become a commodity - largely as exchanges lack exposure to publishers’ 1st party data, meaning that there’s only limited levers exchanges can play with to improve the yield of a publisher.
In contrast, the buy side has been getting smarter about determining the value of the user. Whilst the buy side has got smarter, publisher ad server tech hasn’t bothered / been incentivised to get smarter.  For example, Google’s incentive for DFP is to ensure inventory is available for Google AdX. Exchanges were easy to try with and easy to get rid off. 
Adservers / exchanges allow you to set floor prices. But setting them at placement or geo level is no where as close to the sophistication the buy side is taking decisions today. The sell side should be as smart as the buy side. Adservers / exchanges allow you to set waterfalls, but these are limited in their smartness as they do not auto-correct themselves - but simply make a decision on which bidder should flow through and which should not, with limited data input (e.g. sharing of bid density) from the advertiser side.
Every ad server or exchange in the market say that they maximize your yield,  but if you consider the above - either they aren't incentivized to do so or they aren't equipped with the data to do so. 
As one CBSi exec said in this article that publishers need to get more sophisticated as advertisers are today.
Satish Polisetti, PolyMorph CEO

Satish Polisetti, PolyMorph CEO

Q: Explain in more detail the idea of dynamic floor optimisation between exchanges.

A: Header bidding in the web world is centred around letting publishers conduct auctions with multiple exchanges at the same time - which is a great innovation. Each exchange has a different set of DSPs, meaning a different set of advertisers, in turn meaning every user of the publisher is valuable differently to different advertiser/DSP/exchange.  And when the buy side is taking decisions on whom to buy at what value; it is the job of the publisher to figure what is the worth of that user for each exchange. 
This is a great problem for Machine Learning / AI to solve i.e. solving the challenge of price discovery across various exchanges with the aim of determining what the inventory’s true worth is. In a way this is about freedom for publishers / empowering publishers vs depending on exchanges to do the job - as in reality exchanges can only / are only incentivised to price optimise within their exchange. 
And this is something that is increasingly important as the number of manual or non-tech options to optimise revenue dry up.

Q: OK - so this idea makes a lot of sense.  Surely other ad tech vendors are doing this?

A: In the past, yes, by all means - Exchanges might have tried doing this within their own exchange. In the new world, it doesn't matter - because an exchange is just a pseudo-bidder who is passing the highest bid to the publisher header wrapper. Whereas the best scenario is that price discovery should happen across exchanges. 
Secondly, in the past some exchanges  tried to artificially inflate by introducing fake bids too. That is not the legit way to do things. 
Thirdly, in the past exchanges did implement a certain level of dynamic floors, but it this was only based on superficial data like geo, device vs using more granular level attributes. 
Finally, this basic approach was enough at this point as the DSPs themselves were only buying based on these superficial data points of geo and device. Now DSPs / the buy side need to prove ever increasingly important KPIs, meaning that DSPs have moved away from buying cheap inventory to buying inventory that performs. 

Q: Interesting.  So then what’s required to actually implement this?

A: It’s relatively easy to get dynamic floor optimisation running.  First off you need to add JavaScript to your page, which collects all the bids from the various actors on the page.  Using the data collected from this JS, a purpose built Machine Learning (ML) algorithm then runs in the background to start creating prediction models; to enhance the prediction model further, the publisher has the option to include audience segment data.  The ML model references hundreds of parameters to further fine tune itself, and based on the learnings, the exchanges are then given various levels of floors, and then further iterated.
From a transparency perspective, publishers are provided deep analytics based on the learnings gathered, giving them the ability to take actions with those analytics.  

Q: Results.  What kind of results / optimisation lift might a publisher see?

A: The revenue lift can be anywhere from 5% to 40%.  However, it’s not all about straight revenue uplift.  Additionally, value is returned in cost savings, as using ML means that robots and machines do the heavy lifting, meaning that publishers can save money and make their ad-ops and ad-tech engineers spend time on different areas.

Photo courtesy of Ishikawa Ken (via Flickr)

Sorrell leaves WPP, Zuckerberg grilled by Congress & Netflix reports stellar growth: this week in media

Sorrell leaves WPP, Zuckerberg grilled by Congress & Netflix reports stellar growth: this week in media

Spotify IPO, Facebook closes 3rd party data deals & MTG splits into 2: this week in media

Spotify IPO, Facebook closes 3rd party data deals & MTG splits into 2: this week in media