← Marketecture: Get Smart. Fast.
Marketecture: Get Smart. Fast.
Playwire with Jayson Dubin: Human Intelligence vs Machine Learning in AdTech at Marketecture Live
00:21:19
Loading summary
Transcript44 lines
- [00:00]
A
Hey everyone, it's Ari here. I want to let you know about our upcoming Market Live conference in New York on March 10th and 11th. Our live events last year were smashing successes with sold out standing room only crowds, amazing speakers and the best content you'll get in any setting in the advertising business. This year we've expanded to two days and over a thousand attendees, so it's the must attend event for the doers and thinkers in our business. You're going to learn something at this event. The speaker lineup has just been announced and it's really strong and we're just getting started. So we announced Sophia Kolushi, the CMO of Molson Coors Neil Vogel, the CEO of People Joanna o', Connell, the Chief Intelligence Officer at Omnicom Jeremiah Owang, the General Partner at Blitzscaling Ventures, he's an expert in AI and Lance Armstrong, the General Partner at Next Ventures.
- [00:51]
B
Get your tickets now.
- [00:52]
A
Early bird ends soon, so your tickets are available at market, that's markitecturelive.com and we have special deals for brands, agencies and publishers while tickets last, so we're going to sell out. So you want to get your tickets. It's a two day event so plan ahead. But it's in New York, nice and easy to get to and we're looking forward to seeing you there.
- [01:12]
C
This podcast is brought to you by audiohook, the leading independent audio dsp. Audio Hook has direct publisher integrations into all major podcast and streaming radio platforms, providing 40% more inventory than what could be accessed in omnichannel DSPs. What's more, audiobook has full transcripts on more than 90% of all podcast inventory, enabling it advanced contextual targeting and brand suitability. Audio Hook is so confident that in addition to CPM buys, they offer the industry's only pay for performance option where brands can scale audio and podcasting with peace of mind, knowing they are only paying for outcomes. Visit audiohook.com to learn more. That's audiohook.com.
- [02:03]
D
My name is Jason Dubin. I'm the CEO and founder of Playwire, a company I started, I can't wait to say this almost 18 years ago, making me one of the. Actually, you know what, is there any company here that's over 18 years in business in EdTech? Anyone? 1. Anyone else? 2 All right, quite a bit. All right, never mind. So I've been around for a long time and I've done this for, you know, 18 years and today I'm here to talk about, you know, human intelligence versus machine learning and how Playwire adopts both of these into our principles and how we run our business. And I'm going to go over a few slides and show you some of the results we've got by marrying both of these. So who is Playwire? Real quick, in case you don't know, we're a technology and services company that helps publishers either make more money from their user base or run their businesses more efficiently. Or in layman's terms, we help them lower their cost and make their money. Real simple. So we do this for about 1000 different websites and we're now doing managed service for app as well. And the only way to really, for us to manage it with a small subset of team, about 150 people, is really to embrace automation and AI. That's the only way we can absolutely manage this massive amount of inventory and traffic. And we work with thousands of sites and each one of them is, I guess, snowflake. And they all have different requirements. So there's a false choice. I think there's a false dichotomy going on. Publishers feel they need to go either one way or the other. So it's either full automation or a full human involvement. But the reality is you need both of them to make a decision, really, and figure out these complex problems. So what's my decision framework or the company's decision framework? Well, AI is perfect for repetitive pattern based decisions. So, you know, basically things that happen at such a rapid frequency, a human couldn't keep up with that. And then you need human intelligence for strategic contextual decisions. Right. So there's a lot of smart people in the room. And we all know that ad tech is a gray industry. Not everything is black and white. And a lot of you have to understand who the players are on the field, what chess pieces are on the board. And by taking all that information, you can make strategic decisions. And, you know, data just couldn't do that for you.
- [04:07]
B
Right.
- [04:07]
D
Data could tell you there's a problem and they can tell you something's not working. But what they can't tell you is what's the answer to that problem? That's where contextual thinking comes in. So the first thing I'm gonna talk about right now is, oh, sorry, went the wrong way, is where AI absolutely wins. And that's traffic shaping. And I wanna put a definition around this so we all understand what traffic shaping is. I'm gonna steal a term from Chris Kane, I think I saw him earlier, which is feed DSPs what they. Well, you know what, let me Strike that, feed DSPs and SSPs. I'm not really sure which is which anymore. Cause they kind of both do both feed them what they eat. So we know Magnite likes to eat 70% viewable inventory, 300 y 250 on gaming content. We're going to feed them more of that. And by feeding them more of that, they're going to get better signal fidelity and they're going to end up buying more of our inventory. And that's how the flywheel works. And the inverse side, we know they don't like traffic from Uzbekistan or Korea or whatever will stop sending them traffic from there. And the results are pretty astounding. So two cohorts, right, we had a control group with no traffic shaping and one with traffic shaping. And the cohort that had no traffic shaping rose 9% on the RPS. RPS is revenue per session. So that's our North Star. So when we look at everything, everything's based on a revenue per session because it kind of gets rid of the noise around traffic fluctuations and things like that. And on the experiment where we did have AI, we saw a 21% increase in RPS. So it's a pretty staggering delta, right? 12% on a company of our size growing 12% on RPA is pretty impactful. I think it would be on anyone, on anyone's business. So increased revenue, increased efficiency. But there's also a secondary effect to this, which is what we found was that we actually, since by reducing the amount of bids per bid request per session by 17%, they actually the pages loaded faster and therefore they got more traffic. So there was a secondary effect to that. The second area where AI absolutely wins is price flooring. So we came up with this notion of something called the price floor controller, right? So for complete transparency, we use GAM to run our ad stack. And we know it's pretty limited on their rules, how many rules you can have and where they're pretty much in their rules ancient, how many rules and levers you can pull. So we said, how can we take this out of GAM and leverage AI for that? So we found nine dimensions that made sense. So think day of week, hour a day, ad size, refresh count. So about nine of them. Not going over all of them right now and you start to multiply that. So you've got nine different dimensions times every ad. So we do it on every single ad, on every single request. And again, pretty astounding numbers that starts to compound and compound and compound. We average 1.2 million price floors per site, per day on every single site we work with. And again, you see a 20% uplift in RPS. So it starts to add up, right? We've got 20%, I'm sorry, RPM, not RPS RPM. We had a 20% lift on RPM on the page view and the delta on the AR, the traffic shaping. Oh, God, it did again. Jesus. I'm sorry. So every great story starts an origin story, right? So how did I get here? And mine starts with just honestly talking to the industry, talking to individuals like you. The DSPs, the SSPs, the agencies and I got industry feedback. And at the time we started this initiative, which is about three years ago, the notion was quantity over quality. More bids, more traffic, more ads on page. The more we put out there, the more the SSPs are going to ingest and then you're going to make more money. But by listening to the feedback of the agencies, again, which AI or the data wasn't supporting, all people were saying, we're moving from quantity to quality. So I came back and I came up with this initiative called qpt, stands for Quality Performance Transparency. So all the inventory had to be quality, it had to be performant and had to be transparent to every one of our partners. And ultimately what we're saying is we needed to trust every bid coming from Playwire and we want to make sure that we're in your, you know, in your good graces. So the first thing we did, and I know a lot of people are not going to love the thing I'm about to say is we took all the SSPs and we looked at them and we said, which ones have are accretive to us and which ones are not. So we went from 26 SSPs down to 16 SSPs and we're probably gonna go down to 13 SSPs by the end of the year. The next thing we did was ask all the remaining SSPs to remove all the reseller lines because we noticed again, talking to all the industry executives that they cared about SPO or supply path optimization. So we again, we noticed, we did a lot of data mining and a lot of, you know, a B testing, and we saw that the lines that were resellers just weren't adding any value to us and to our publishers that we represent. And the last thing we do is clean up our content, right? So we got rid of anyone who sniffed at MFA or any or any content that we just did deemed unacceptable. And I'm going to let the results speak for themselves, which are pretty astounding so there's three lines on this I'm going to look at which are one are the vertical lines which are requests. The orange line is viewability and the green line is cpm. So this was a culmination of two years of work. So this includes the traffic shaping, the bid shaping, the price floor controller and the QPT initiative cleaning everything up. And we reduce bids requests by 61%. Right. So the SSPs and DSPs love that because it's less on their servers. Revenue went up by 58%. CPM soared to 168%. And by the way, this is on a single large site that is about five to seven million dollars a year. So just to give everyone context and that there's full transparency, viewability went up 107%. And again, just like with traffic shaping, there was this butterfly effect where due to the less load page on this, on the page we were able to increase traffic by 8% with better lighthouse scores and things like that. So my final task after we did all these things two years ago, I said, well how do we productize this? How do we make this into a platform we can use? And then also listening to the industry, I was hearing that hey, we would love to use your product but we're never going to outsource to a third party like you. We have to manage that ourselves. So I said, well, how can we productize this to give it to a make a platform that enterprise publishers could use. So today we've partnered up with marketexture. I'm here to announce that Playwire has given all this technology, the bit shaping the traffic shakes and the flooring and all the things I just showed you up there, into one unified platform made for enterprise level publishers called ramp, stands for Revenue Amplification Management Platform. It's the first of its kind and really what we want to give everyone is full control and full vis visibility. A rules based management control system so you can control what you want to do on what page. It's your platform. You should decide that in the event that you want, you don't want to do it manually, you can actually use our AI and just turn that on and you decide whether to take the wheel. So this is all powered by real time analytics that updated every five minutes directly from the SSPs. Right? We're not relying on GAM anymore because the SSPs are the ones that feed us and they're the ones that pay us. So why are we measuring GAM at this point? So that's the platform. Thank you for letting me speak. I know Ari's going to come up here and basically ask me a few questions. And if anybody's interested in talking, just scan the QR code or email me. Thank you. Sorry. Thanks, Ari. Thanks, Jason. I tried to keep it in seven minutes and you did a fantastic job. I didn't do a good job.
- [11:21]
B
No, you did. Don't be hard on yourself, man.
- [11:23]
D
I'm always hard on myself.
- [11:25]
B
So let's talk calculus.
- [11:28]
D
Oh, God, no, please don't.
- [11:29]
B
No.
- [11:30]
D
So I failed math.
- [11:31]
B
It's interesting. You're talking about publishers here and the publishers are getting kind of whacked by AI, Right? I know it's a little off subject, but you're sort of representing their interests a little bit here. So what's it feel like right now to be a publisher in the face of AI?
- [11:48]
D
Listen, I'm not going to talk about, like, what everyone's talked about, which is. Listen. On average, we're seeing about a 15% decrease in SEO traffic to LR publishers. So across a broad spectrum, some are getting hit as bad as 40%. Right. But on average, we're seeing about 15%, I would say, you know, from a publisher's perspective, the things that are really. I think there's two things. One is complexity. Right. Think about all the things that we're being forced to do, whether it's identity, gdcpr, ccpa, just all the complexity comes to that and then taking that analytics. But I think there's a bigger, broader problem, which is honestly. So I was at pre bid last week and everything I heard from them was transparency, transparency, transparency. I think the, the supply side has done a really good job of making everything very transparent to the buy side. They know what they're buying, when they're buying, how they're buying. I know there's the transaction ID that, you know, there's a big fuffle last week about that or last month. But from our perspective, there's little transparency on the buy side to the, the sell side, supply side, buy side, to the sell.
- [12:54]
B
You don't, you don't necessarily know who the buyer is, except for, you know, a domain or something like that.
- [13:00]
D
Yeah, I mean, I think one, there's two things. One, we don't know. There's not much transparency from when the DSP makes the bid all the way through the SSPs and all the intermediaries in, in there. So, you know, why is, why don't we have gross bid value passed all the way so we can see who's taking what out of the, you know, of the midstream. Right. And then on top of that, we spend how much time and effort tracking down malicious ads? Mobile redirects, heavy ad intervention, you know, and honestly, from our side, we spend hours and or months of our time just tracking these things down. We go to the SSP and say, who is buying its inventory? Where is it coming from? No one can give us an answer. We spend all our time, I spend hundreds of thousands of dollars working with third parties just trying to determine and, or how to stop these malicious ads. It's a huge problem.
- [13:49]
B
Right. And you think that problem's at the DSP level, but shouldn't it be at the SSP level too? Shouldn't they be the one if they're passing a bid into you, into your publishers?
- [13:58]
D
I think it's all the way through the bid stream. Right. So I think it starts at the DSP level. Right, because you have DSPs buying into DSPs or seats on the DSPS buying in, and you really don't know where it's coming from. And then there's like Inception. It's like a DSP within a DSP and it's almost impossible to track down. Like, why isn't there one universal ID per buyer that's passed companies like us, or any publisher for that matter, or SSP that lets us know who the buyer is? I don't. It seems pretty.
- [14:25]
B
Isn't there buyers that. Jason, is that the proposal that I.
- [14:28]
D
Don'T think is, you know, that would be nice if, if it was, if it was adopted, but it really, you know, it takes a village to get that done. And I think the supply side, you know, I think we need to get together and try to figure out how to solve this very complex problem. It really shouldn't be a problem, in my opinion. I mean, on our side, we're given all the information. How come it's not reciprocal?
- [14:47]
B
So back to your presentation. You were talking about using AI for certain things but not for other things. What. How did you come to choose it in one way or another? Is it just like the cardinality? You have a million rules, so obviously AI should do that. And other things are more human. Touch.
- [15:03]
D
Yeah. I mean, listen, everything starts with human intervention, right? It all starts with, you know, someone having an idea and we're sitting, sitting around like a think tank and saying, okay, what are the dimensions that are going to affect change or outcomes for that matter? Right. So it's a great, great topic, outcomes. And we'll sit there and then we'll start with like a bunch of testing. Right. And get efficacy around that. And we'll roll it out on one site and five sites and 10 sites. And once we kind of see, hey, this works, we then roll it out across the entire network. That also takes time. I know everyone says, oh, AI will solve problems. I can't tell you how many companies approach me saying, oh, we have outsourced AI. We've spent two years making these models for both the price floor controller and, you know, the traffic shaping and bid shaping and timeout shape, whatever you want to call that, timeouts. And it's not as easy as everyone makes it seem, honestly. I mean, I know AI is great, but it's not this magical thing you just sprinkle on that just. I wish it was. Trust me. I asked my team all the time. I had a procus here, like, why is this working? What's taking so long? And the answer is, these are complex things and you ingest tons of data and have really smart people making very salient decisions and then rolling that out and testing the efficacy of it. It's not so easy.
- [16:17]
B
Yeah, I think you're possibly creating problems if you're using the AI without knowing what you're looking for.
- [16:24]
D
Oh, 100%. And on top of that, we're also working with probably a lot of people in this room and, you know, we don't know what we don't know. Like, as the JavaScript. Is the JavaScript going to break the tags? Will there be latency in getting your data? When do you send your data? How is your data put together? Will it go easily into our, you know, our data warehouse and that we've put into Snowflake? I mean, these are, again, you know, when you're dealing with billions and billions of requests and you have thousands of publishers depending on you, you don't have room for margin of error. Our clients aren't going to care that, oh, well, the JavaScript broke and the ads were down for an hour or two. Well, that's not acceptable.
- [16:58]
B
So you imagine you just. Just kind of threw that in somewhere, that you're using AI for timeout management, timeout optimization. Tell us about that. That sounds interesting.
- [17:06]
D
Yeah. So again, looking at, you know, we run numerous tests all the time. Right. We're seeing which things, again, will produce some kind of outcome. And we know, you know, timeouts per SSP was actually something that drove results.
- [17:19]
B
So the amount of time the publisher is giving the SSP to respond is what it's about, right?
- [17:23]
D
Yeah. Correct. Okay. So by doing that we actually reflect a few percentage points change. And again, it's not like one huge thing that makes a difference. It's multiple things that make a big difference.
- [17:33]
A
Right?
- [17:36]
B
So another example I brought up this morning very briefly was this idea of AI generated content. You know, publishers can create content using AI. It feels like that is a danger area where you could actually be reducing the quality of the content, right. Have you seen any of your publishers start doing that in a white hat way?
- [17:55]
D
I guess I think all publishers, you know, I think we need to find a way to embrace AI and it's not going anywhere. I don't think by saying we're not going to do AI for our content is the right answer. And I don't know what the right answer is, by the way, but I know we need to figure out how to do it better. What I have seen though is publishers that would go take a template of someone else's website that looks almost identical to the other 500 baking sites that are out there, or 5,000 and have almost identical content. I think what you're seeing is Google being smart about it and saying this is not adding any value. And so it's no different than anything else in the world, right? If you have a company or you're making content and you peel back the layer of that onion and there's really no value there, at some point it's going to be exposed, right? So even us as my company, right? If, if we're not adding, if we're just, we'll call a middleman, right? Hell, they're just a middleman. If we're not adding value, if we're not making more revenue for our publishers, if we're not lowering their cost to do business, then we don't deserve to be in business. And I think the same thing applies to content. If you're not making great quality content that people want to inject. Listen, like you said, Sora is making pretty cool content, right? And people are watching it. People are watching it. So if you can figure out how to embrace that, if you can figure out what that the mechanics are to make that work, then I think, you know, everyone would, would lean into that. But what I have seen is just people just. And again, it was a really great business model for a long time for these publishers because it was almost like a franchise where, hey, we'll give you the template, we'll give you the content, you put it up there, it will be picked up by SEO, you'll get a good rps and you know, we'll make, you'll make money. And then once you know, they started saying, hey, is this quality content? That's when the wheels fell off the trail.
- [19:38]
B
And AI kind of flattens that because if you're getting a single AI response from Google, Gemini or, or whomever the sources are tiny little links at the bottom, you have a lot fewer clicks. Doesn't that kind of. Do you think that will reward lower quality copycat content versus Canonical, you know, the original content?
- [20:03]
D
I don't really have enough data to support that, but I can, I can tell you where we are seeing. I think, listen, we get a lot of referral traffic for honestly from the chatgpts and Right.
- [20:13]
B
It's so it's rising. It's rising, right?
- [20:14]
D
Yeah. So we're, we're actually seeing a positive result. I think us as a community or whoever's making content needs to figure out how can I rank in the LLMs instead of fighting it? Because they're not going anywhere. Right. I don't think anyone, I mean, unless you're, I guess, dot Dash or someone else, you're not going to get them to pay you for the rights. At least I don't think so.
- [20:32]
B
I was going to bring that up next. Yeah.
- [20:34]
D
Okay.
- [20:34]
B
So are you doing anything in that area, dot Dash we should all notice changed this name.
- [20:38]
D
It's now people.
- [20:39]
B
Yeah. So, yeah. So are you saying that people has an advantage just because they're so big they could ask for money or because their content is more canonical or both?
- [20:48]
D
I, I, It's a very contentious statement, I would say, but I think we, it's a very hypocritical industry in that the bigger you are, the more you can come in you, more, more you can get away with. I remember when Twitch first came out, we did a lot in gaming and we would run a campaign and you know, God forbid it ran on and that's, you know, not safe for work content or something like a curse word. They would literally I want to make good. You know, we want cancel the campaign. Whatever they felt like doing at the time because we had to take it. And I would say, hey, well, well, Twitch is doing. Oh well, they're twitch or it's YouTube, it doesn't matter. So if you have scale, you get away with a lot more than, you know, someone who does not have scale. And you know, as we're getting bigger, we're having better conversations higher up in the food chain and we're helping our publishers and publishers need companies like us to really, I wouldn't say, you know, like, I guess we're like the Lorax. We speak to the trees, you know?
- [21:40]
B
All right. The Lorax speaks for the trees. All right, everyone, give a big hand to Jason Duvan.
- [21:58]
C
Thank you for listening to the Market podcast. New episodes come out every Friday, and an insightful vendor interview is published each Monday. You can subscribe to our library of hundreds of executive interviews at marketecture tv. You can also sign up for free for our weekly newsletter with my original strategic insights on the week's news at News Market tv. And if you're feeling social, we operate a vibrant Slack community that you can apply to join@adtechgod.com.