Stop Buying Lemons, With SWYM.ai

0:12Ravi Patel: Welcome to Ad Exchanger Talks, the podcast devoted to examining the issues and trends in advertising and marketing technology that matter most to you.

0:24This podcast is sponsored by Snap. Ditch your outdated marketing plan and spend smarter with Snapchat for Business.

0:44I'm Alison Schiff, Editor-in-Chief of AdExchanger, and my guest this week is Ravi Patel, CEO and co-founder of Swim.ai. Ravi's been working on programmatic transparency since his days at Adfin. a startup that helped reveal just how much of an ad dollar actually reaches publishers. Now, with Swim, he's focused on tackling waste before the money is spent by looking beyond domain lists and shaping the supply path around what drives real outcomes. We'll dig into why a trusted domain doesn't guarantee a worthwhile impression, the challenge of spotting AI-generated ad clutter, is AI Slop the next generation of MFA, and how advertisers can unlock performance gains without sacrificing reach by cutting duplicated low-quality inventory.

1:29But before we get into it, I'd like to take a moment to tell you about a really cool event we've got coming up next year. Join us in New York City at the Time Center on March 30th and 31st for CTV World, which will bring together more than 500 senior advertisers, distributors, and programmers to talk about all things CTV. If you want to master the new TV economy, then you need to be in this room. We'll see you there. Hey, Ravi, welcome to the podcast. Hello, good to hang out today, Allison. The last time you were on was in 2019, and the world has changed, I mean, drastically. Yes, it has. A lot has gone on between then and now, I'll tell you that.

2:14So we're going to talk about all of the new stuff and we're also going to go back in time and talk about your ad fin days just a little bit. But before we do, I love to start out this way because I get the best and weirdest answers from people. What's one thing about you that not a lot of other people already know that I couldn't find if I just Googled you or perused your LinkedIn? in? One thing about me that not a lot of people know, I would say that I have been a citizen of three countries. It's still a citizen of two. So yeah, I've been around. Well, you got to list them. Which other countries?

2:52So obviously I was born and raised in India. So there, and then Canada, and then America. So, you know, just, it's just an apartment above a great party at the end. it's a funny way to describe it um so yes let's turn back the clocks a minute and go back to ad fin and for anyone who needs a refresher ad fin was your startup launched back in 2012 to benchmark programmatic pricing and show advertisers what was actually happening to their money in the supply chain. It's also what you're very focused on today. But in 2016, you guys teamed up with the ANA and Ubiquity, and you did a study about hidden programmatic fees, published a report the following year, found that only 58 % of every programmatic dollar was reaching the publisher.

3:43So just 58 cents. Lots of cents. We're not reaching them. And we reported at the time about how this study kind of made you like a, I guess, an enemy of agency holding companies, or they saw it that way, like they just didn't want a light shining into these dark corners, which I think you didn't really anticipate. You were thinking, hey, this is going to be, this is a good thing for the industry. And then you were disabused of that notion. And in 2019, Adfin had to shut down. But looking back, what did you learn? What did the team learn? What did you misjudge about how the industry would respond to that kind of transparency?

4:22Like they weren't ready to look in the mirror yet, maybe. Yeah. So I'll add one thing to it, right? Like I think one of the things is Adfin was an incubated company. So its co-founders were just very wealthy people who wanted, the purpose was to bring the Bloomberg terminal idea to AdTech, right? And to say like, hey, we should know the prices of things at any given moment and then be able to use that to create indices off of that we can potentially trade on, whether in the financial markets or in the tech markets. When it came to getting the data, you want as granular data as possible, built a tech around processing trillions of impressions on a daily basis and all that stuff.

5:02I think what was misjudged was how hard it was to get access to that data because there was always this what I don't know could hurt me kind of mentality. A lot of people thought there's like something in the log files that would expose them or anything like that. I think what we learned from that is, though, the fact that only a majority of the money was going to fees, and mostly on the buy side, and then some on the sell side as well, and very little was making it to the publisher. But then the other piece is also like what we didn't put into perspective was the auction dynamics, right? Like, hey, if only 55 cents on the dollar is getting to the publisher, How is this impression still getting bought at the end of the day?

5:45It's an auction, right? It looks for the highest bid. So that also, what we didn't explore through the years is that means that the entire industry has this issue, not necessarily a specific advertiser or specific agency or anything like that. So it's been, I think that's one of the big pieces. The other piece was like, you know, we wanted to create more transparency into the ecosystem. It was very difficult. It was a product that kind of went a little too deep into some of these things. But having said that, through the years and my experience in ad tech, we learned a ton across auction dynamics and how they work.

6:26And it's like, hey, if you start just focusing more upstream in this industry, you can actually control a lot of the stuff that you don't want to be on and try to capture more of the stuff you want to be on. At that time, the tech wasn't there. to kind of do a lot of the stuff that we can do today. And a little bit of it was a lot of DSPs and SSPs were still relatively newer to the market, even if it was a couple of years in and everything. So a lot of the feature functionality, a lot of the capabilities also weren't there. And we're figuring out different things. So the technology does exist today and you're building it.

7:05Adfin's whole thing was auditing the supply chain, but it was after the money was spent. And now you want to help people make smarter decisions before they spend the money. I mean, right off the bat, is the better, I guess, less combative approach to do that? Like how much of it was a lesson from Adfin and how much was just where you saw the opportunity? It was a lot of learnings from Adfin and it was a big lesson from Adfin, right? Right. Adfin, what they gave was insights back to brands and agencies of how to improve. But those insights were so, could be so granular and you couldn't just say, hey, just, you know, stop, you know, pull this lever in the DSP and it's fixed.

7:50It doesn't work like that, right? Because of just even today, and we could talk about this of how DSPs and their QPS threshold versus SSPs and what they have access to. So what you needed was millions of different decisions needing to be made. And they changed based on the interaction of the KPI or whatever you kind of get back as a result. And that would be dynamic. No trader could do it. So what we kind of focused on is, okay, how do we get inside the pipes to help kind of shape it and then help decision on it at the same time? So I think a lot of people know this by now, but maybe they don't.

8:29what SWIM stands for stop wasting your money, S-W-Y-M, which is very clever. Nice job. And yeah, I mean, the pitch, like we, like you just said, although you didn't cite the number, something like$20 billion a year, maybe it's more, I don't know, tens of billions of dollars are getting wasted in programmatic. And you guys have this metaphor that you use that advertisers are out there buying lemons, like low performing inventory, but they could be buying cherries, like tasty inventory that actually hits their KPIs. But how do you define waste? Because I know that's kind of like a weird question because waste is waste, right?

9:16But there are a lot of companies out there that claim to be providing a service, but they're not actually helping do anything. They're adding to the waste, but people who use them think they're getting value from it. It gets quite mushy very quickly, I think, when you're trying to define exactly what it is. Yeah. I mean, waste is, it's two things at once, right? There's a binary component to waste, which is just black and white, right? Like this is just bad. And then there's a subjective, not even that. That's the subjective piece, right? Like what is bad inventory to one brand is different to another brand, right?

9:55Some brands will use paid signals such as how many ads on a page, how it loads, what do I have access to? Where is the placement? Other brands will use things like, where am I getting the actual conversion? Is it yielding dollars? Am I getting sales in? And then the binary stuff is what I would call like the fraud piece, right? Historically, what we've done with fraud is we just said, oh, this publisher domain.com has, you know, 3 % fraud means let's block it from the entire, you know, let's just put it on our blacklist. And I think you got to get more granular than that. It's, you know, at the waist in that scenario is a bot, is a non-human, and you need to block that non-human.

10:37And so there's waste like that. And then the subjective piece is, you know, there was also a waste piece around why am I showing up on brand safety? So one of the biggest things, my experience at IAS, you know, early on, it was different. We created these post-bid tags that would sit, you know, that would fire when you're, when you bought the impression and they would tell you about what actually happened. And then there would be pre-bid segments that try to avoid what they can in real time. But it was a little bit, you know, it wasn't as defined and it wasn't as controlled. but it's like one of those things it's like okay so is it wasteful if i'm using pre-bid segments and a post-bid tag and then i keep paying for both over and over again or is it better to pay for it once uh and keep monitoring it but then you know be able to not use the pre-bid segments on the same thing that i just thought that you said was fine you know so there's different colors of waste and that's that so our whole premise is uh the other piece is in 2011 i remember this the martech lumoscape was 150 200 ish companies today it's over 16 000 companies right what if you look it up um i think uh that's the stat i had seen that's give or take you know a couple grand but still it's huge the amount of companies right sure yeah it's still big like uh and And there's so much out there in terms of like everyone kind of figuring out where they position themselves, where they get into the ecosystem.

12:06And we made programmatic to be supposed to be more simple. It's supposed to be more automated and it's more scalable. It's definitely more scalable. It's definitely not more simple and slightly automated. Right. And so our whole premise was, OK, how do we go back to it and fix the pipes instead of reinventing new pipes? The pipes are already there. It can scale very, very quickly. It's just we've got to fix what goes from the sell side to the buy side and streamline it all. So that's a nice segue into a column that you wrote for us, for AdExchanger, back in the spring about how buyers are just way too focused on domains.

12:46they build block lists, approved publisher lists, so-called premium lists. And it's all based on this idea that if you only buy on sites that you trust, then you're in good shape. So yeah, the pipes are there, but the way we put things through it, it's messy. It's silly. It doesn't work. So your argument is that a domain is just a container. It's not like inventory. And what really matters is the inventory, which is what is inside of the container. And then even when a buyer approves a publisher, they don't get to see every bid request that comes from that publisher site. It just depends how the SSP packages things up and how the DSP filters and throttles.

13:27So a lot of it just never even reaches the buyer. When a buyer puts a publisher on their approved list, What are they actually getting? And would they even know what they're missing? Would they even know that they're buying in a wasteful way? Well, they won't know until they've spent the money, right? That's the whole beauty of programmatic. Beauty or lack thereof. It's a nice way to say it, right? But it goes back to the grocery store example. Everyone in our industry that's setting up a campaign is going into a grocery store with the money in their pocket. Caveat there is you cannot walk out with that money in your pocket.

14:07You put a budget into a DSP and it's supposed to be spent today. I guess, well, by midnight, they're going to spend the budget. Maybe the campaign did amazing. But what campaign has ever done amazing and kept improving KPI beyond what you set for it, right? So at the end of the day, and the issue stems from two pieces, right? One is DSPs and SSPs historically talk at each other and not to each other. And this is not their own fault. It's just how our systems were set up. You know, SSPs were built to set up to basically their clients were publishers and DSPs clients were advertisers and agencies.

14:40And because of that, there's no feedback loop that goes in between them. So SSPs only know you bought this thing yesterday. It means you want it. So if you bought a lemon yesterday, you're getting a lemon today. Like you're going to get access to bid on it today. And then the other piece is all DSPs are QPS throttled. So even the biggest ones throttle it because it's extremely expensive to allow every single line item to be able to auction on every single bid request that hits. So you see, they're seeing about 40 % on the larger side and then smaller DSPs are seeing probably around 20-25 % of the entire supply ecosystem.

15:15Granted, there's a lot of duplication and stuff that's added into that. But at the end of the day, if you're only getting access to, let's say, the domain that you send. You added a publisher domain there. You're only seeing 20 to 40 percent of the bid requests from that publisher domain. And it's also what the DSP decides to put, give you access to of that 20 to 40 percent. There's still a lot more out there on the on the sell side available. And so it's like, how do you kind of close that gap? And how do you start thinking from the demand side only to saying, first, I got to go control my supply ecosystem and then I got to bring that back into the demand side?

15:52So I was doing a lot of reading, even though we know each other before this podcast to get stuff that I can ask you. And I read a whole press release that Pubmatic put out about decision fabric, which is this containerization thing to let partners plug their algorithms directly into Pubmatic's exchange. So they can evaluate every bid request as it comes in. And you guys swim. You're one of the launch partners. And in the release, you said something like the buy side has been optimizing against a filtered view of supply for a decade. But like you were just saying, I mean, DSPs purposely throttle how many bid requests they look at to keep their cloud costs down.

16:35So filtered view, I mean, it's not a technical limitation. It's a business decision. And what can be done if this is just the reality of the situation? They do have to keep their cloud costs down. It's very pricey. Yeah, I mean, one of the things I think you're starting to see is because the idea of containerization, right, there's RTD, which is what DSPs connect to, right, that's basically saying, hey, send a bid request to my cloud, and then we'll run the auction mechanics inside, do an auction and then send you back a response, you have about 100 milliseconds to respond, give or take. and then in the case of containerization and what some of these SSPs are doing is just saying, hey, we already have the infrastructure.

17:18We're just going to give you basically, and I'm simplifying this a little bit, but we're going to give you a server. You sign into it and we just minimize our egress costs and saved this ecosystem a lot of money because now you decide out of everything we have access to what you want and where we should shape it to essentially. So one of the things I fear about containerization is a lot of people say containerization, but are they truly doing, you know, bid by bid decisioning? Or are they just saying, hey, instead of, you know, here, instead of just having a domain list, they're like, okay, here's, you know, a page level list and that's it, right?

17:54And, or here's an audience that doesn't change, right? At the end of the day. So yes, DSPs are doing this now by some, you've seen some of the newer DSPs starting to say they're containerizing into the sell side, but then that goes, it's like, why are you going to sell another DSP at the end of the day? Why not go direct into supply and kind of bid on it too? But yeah, I mean, it's just, it's very compute heavy. It's very expensive to be able to auction everything and everyone, even if a DSP saw a hundred percent of the QPS, it doesn't mean your line item that you're running has access to auction on all a hundred percent of it.

18:30And that's one of the biggest things is, you know, they still have to limit what you can auction on because they just wouldn't be able to get through that 100 millisecond timeframe unless you start co-locating these things even closer and within the same data centers. Well, what is SWIM's approach? You guys do a form of supply shaping, don't you? Yes. So we do a form of supply shaping. We started off with the RTD side. And again, we ran into early on when we started, we were experimenting it. We came up with our own algos that kind of helped encode it so that our response times were fast enough.

19:04But when you have egress, it eats into a lot of the latency. So we worked on encoding algos and compression algos for how to respond and things like that. But we've containerized. And the reason is we want to be inside the infrastructure of the SSPs. And so all our models are constantly changing at all times. The reason that we decided to do it at a later stage of containerization is because these containers were static early on because SSPs were still figuring it out. And they're figuring out how to make it more dynamic. So early on, say, you know, even last year with one of the SSPs that started coming out with it, it's like, you can only, you know, our models are changing every day, all day in a way.

19:45But if you're only able to change it once a week, you kind of lost the whole premise of sell side shaping or traffic shaping in a way. So yes, we're containerized. Another nuance to containerization, because, you know, everyone wants to say the container has this, how much access do you have? So Swim's taken a stance that we're going to pay money to get access to 100 % of QPS across all our sell site partners. And that's exactly what we do, right? So the amount of data and the amount of bandwidth that we see is great. But typically, even containerization companies that are containerized see a subset of all the QPS.

20:23So just because you're containerized doesn't mean you get access the entire supply side, you're still throttled. How expensive is that for you? I wish not to disclose. It's not cheap. Let's just say that. But we've been, you know, one of the things that goes back to it is we focused. So one of my background is also a lot on database optimization. So, you know, even at Adfin, we built our own database from scratch. We optimized it. It competed against the likes of Snowflake and others, too. and that's why I was able to process so much data. We're also good at encoding algorithms, right? So in a simplistic view, think of a bid request or let's say a long paragraph hitting you, right?

21:07Coming to you. That has a lot of, that could be a decent sized file, right? But if you're getting a similar pieces and words of those files over and over again, what if you can suddenly turn that into, encode it into just three letters that mean that entire sentence because you've seen it so many times? Right. Similar to, as they say, like a DNA, you know, our DNA, if you were to spell it out, you know, or actually even the number pi. Right. If you were to keep going with it, it's if you translate it to letters, it said that it can write out every single name sentence ever spoken if you go long enough.

21:44Right. So we encode these things into smaller pieces so we can respond quicker. So we spent a lot of time kind of focused and doing that. And that's made it a lot more economical for us, at least, and the way we do things. It's like shorthand. It's Pittman's shorthand, but for what you do. Exactly. And we enjoy doing that. Like we enjoy the deep tech pieces. And, you know, these are the, I think these are the pieces that as we fix the ecosystem and as more partners also come and do the similar things, I think our ecosystem itself becomes more efficient. One of the other things is like, you know, when you're doing sell-side decisioning, you limit the amount of people in Illumiscape that can enter it then, right?

22:29Because you're controlling that bid basically from the publisher to the SSP. And then once it hits the SSP, you're controlling it all the way to the DSP. So you know exactly what goes on inside in that entire path. Well, fixing the ecosystem from your lips to God's ears, there's a lot of work to do. We're going to take a quick break. And when we're back, we'll talk more about how SSPs are evolving, curation, all of that good stuff. So stick with us.

23:08All right, we're back and let's talk about curation. So SSPs, agencies, data companies, sort of all the rage, they're packaging supply for buyers. Some of it's really smart, but I think plenty of it is basically just a repackaged PMP with a markup. And like we were talking about before the break, you have an approach called supply shaping. but isn't optimizing bids based on campaign performance like exactly what the dsp is supposed to do like why does this logic need to move upstream into the supply layer and like what is it that dsps are missing yeah so no it's what the dsps are supposed to do and it's what they do do right they do optimize the bid in terms of how much to bid and which bid to make eligible based on the DSP targeting criteria you've set, right?

23:59Whether it's certain audience and they're trying to manage an audience signal or it's a certain viewability stance. However, the issue has been what the DSP makes available to bid on. One, we talked earlier, right, about how DSPs are QPS throttled. That's one piece. The second piece is of the 20 or 30%, let's say 30 % that gets to the DSP, only 1 % or so gets eligible for your campaign to bid on. So it's like, what's the right pieces in those pieces? It might be more than 1%. In certain cases, it depends on the size of the campaign. There's a few other factors that go into it. So the issue is, though, let's say the DSP bids on something and there were more bids available on the sell side, except since you only bid on it five times and you won once, it's telling the SSP that, hey, you're not too interested in this kind of bid request, right?

24:55At the end of the day. And those are the kinds of bid requests that have a, you know, there are more cherries for you than there are lemons. But the SSP never sends it because they don't have that feedback loop of whether did it actually deliver an outcome? And it's only relying on whether you bought it or not. So the reason you need the world of sell-side activation and supply, we call it supply shaping and not necessarily curation. And it's what you said about curation, right? Curation in our world was mostly like, hey, add some audiences or target specific lists. And those lists largely stay the same.

25:26Not much really changes, right? You basically ask for an audience, you package it, you send a deal idea out, add a markup, and it was never built to be dynamic. Nobody changes the audience definition halfway through a flight or inside of what goes in that audience list. Supply shaping changes the decision on every single bid. So what it does is when a bid hits us inside the SSB infrastructure, we don't care about what the audience label is or anything like that. We care whether it has any propensity to perform based on an advertiser's actual outcome signals that we've normalized and brought in real time as well.

25:59I don't care that user one, two, three, four, five, it lives in New York City, loves to snowboard. I just care whether that potential bid request has, can lead to this client to a sale at the end of the day or whatever the outcome is. So we follow the outcome. Did that impression lead to a conversion or a sale? And can this bid help do it again? Right. at the end of the day. But when you can go, and a simple analogy is, when you can go fish in the ocean, you have a lot more access to find more fishes than fishing in a lake at the end of the day. And that's why this idea of supply shaping is extremely important because it's also not necessarily, in my opinion, like completely decisioning.

26:38That's what the DSPs do. They decision whether you should bid on it or not. It's shaping because we're just telling the bid request to go a certain path, get it prioritized in a DSP auction for a certain campaign or line item, and then let the DSP do it. And then there is other aspects. You can add custom algos on the buy side, which will then help you do bid decisioning on it, which Swim also has a suite of products around. But there's two different pieces to that as well. I hear everything you're saying, but why hasn't this industry moved past domain lists yet? Like, what is this ongoing obsession with lists?

27:14that's because a lot of dsps can only do things at the domain level they don't do things at more granular levels it's just you know they just wouldn't be able to kind of fit it in so you got to remember when an auction happens right what the dsp does is it looks up the line item and it looks up okay where you know what filters does this line item have to make it eligible or ineligible to bid on this bid request now if you have a domain of i'm going to simplify it to 100 domains, easy, look up, right? Very quickly. And they have, again, I said 100 milliseconds across, let's say, tens of thousands of line items.

27:51But it goes at a page level. Now, suddenly those 100 domains is more like 10 ,000 lines, right? So imagine the amount of data that just increases. So now you have 10 ,000 pages that's got to look up against. It takes a little bit more time, more compute, or you get faster servers or more servers, right? Just exponentially increases everything. So if you could take some of that targeting, those pieces that you're trying to go after, pull it to the sell side and do it within that 10 millisecond timeframe that you have, and then come to the DSP side and then figure out what you actually want to bid on it because you have already pre-vetted most of it.

28:24You could still have one or two more vetting factors on the DSP, right, to further trail it down. I have to ask you, Ravi, does your family know what you do for a living? Like when you have to explain what your job is? I don't think so. I think, Like, yeah, when I say advertising, my dad just assumes it's like Google search probably. I never really sat down and explained what programmatic and things are to them. How would you if you had to? I would need a whiteboard to show them exactly what and how it happens. But, you know, it's a little bit easier. It can be easier, but I just think then they would ask more questions that I don't think they would understand.

29:08And I was like, okay, then how do you decide, you know, how does this request go in real time? And like the idea of like all of this stuff in our ecosystem works within a couple hundred milliseconds, right? A publisher decides within the first second whether the page loaded or not, how many requests to send out to which SSB. And then within another 200, 300 milliseconds, you have an impression that's loaded. Daddy works on the internet. I think that's enough. Yeah, that's it. I wanted to talk a little bit about the new wrinkle for dealing with quality control on the internet, which is AI-generated slop sites.

29:47We've written about it a little bit recently, recipe sites in particular, but there's AI slop all over the place. Slop is also a tricky word to define. Maybe in the eye of the beholder, I don't know. perhaps someone would find some value in some forms of it but some people are calling it the new mfa um despite the the nuances um the industry has put a lot of effort into trying to keep ad dollars away from low quality sites from made for advertising sites and now ai can just spin up junk sites mfa sites included but in bulk overnight faster than anyone can even add them to a list if domain blocking can't keep up with that.

30:33Like, what is the alternative? I guess it's everything we've been talking about, but this seems like an arms race. Yeah, I mean, it's crazy how fast you can kind of just tell whether you use Claude or ChatGPT or whatever. Say, hey, find every blog article on this page because I like it, and I have this domain that I want to re-release in my, you know, rewritten, essentially. And that's a lot of what's going to start happening. So one is, I think a lot of those, you got to first get traffic to those sites and buying bot traffic. And that's a whole other story, right? Like we've been used to blocking bots at the domain level versus at the user agent level.

31:12But as the industry gets better and better, and we've done work with Georgia Pacific and, you know, Foo Analytics was their fraud partner that partnered with this where we took, you know, we got the actual user agents and we just blocked them from the bid stream. So I think that will help. So if they're buying bot fraud, you can help remove those AI slop sites without having to scrape everything and figure it out. The other piece is, I think, understanding what is actually ad-eligible and what's new. And that's why we love the idea of seeing everything in our supply ecosystem. Because when something is new to us, we actually have other tech that, well, yes, based on my experience at IES and OpenSlate and things, we actually go get all the page signal data, all the page quality data.

31:53We have something called the swim pH, right? Which tells us exactly where the placement is. What is it near? What's the text to word ratio? Does this seem like it's a human that wrote it and things like that? So that way, if it's something new, we can actually vet it as well to see if it should be eligible inside the bid stream or not. And then, you know, it's not ours to decide whether we like a certain impression. Listen, you can have an AI generated site that's generating legitimate traffic. And guess what? Legitimate humans are enjoying it. Like recipe sites are crazy. I think the best thing is take a recipe site, put fewer ads on it, and just a simpler experience.

32:29And I bet you, you could do pretty well on it. And there's not, you know, just because they created a recipe off of AI, it's not, I don't know if it's necessarily bad, as long as legitimate people are going to it, and it's resulting in outcomes. And that's what it's going to come back to. So like, I think if people stop buying on upper funnel vanity metrics and mid funnel vanity metrics and move towards more lower funnel attributed metrics, even if it's not perfect attribution, and I know attribution has its flaws and measurement has its flaws, you could still get so much better and you can move the needle so much more and deciding if this is working for you or not.

33:06I mean, before we had these AI slot sites, I mean, we have a lot of apps that are all clickbaits and we still keep buying these apps. Why? Fantastic question. While you were talking, I was thinking, well, nothing bad will happen unless one of the ingredients in the recipe is glue or something, or the measurements are totally off. Speaking of bad measurements, and it's like, put in seven pounds of flour, you end up with bricks or something. But yeah, no, I hear what you're saying. And I think there are companies out there making the argument that we need to let the market decide. And if there's a human that has decided to click on this thing and there could be some value, then there's no reason why it can't be a vessel for advertising so long as the advertising is respectful and makes sense and there's no fraud.

33:56Yeah, that's the big thing. It's like, it's not fraud. And, you know, brand safety is subjective. So every client uses a verification provider or every large client at least use as a verification provider of their choice to determine brand safety, brand suitability. If it passes those pieces and it's legitimately giving you certain kinds of conversions or outcomes, then it should be fine, even if it's AI generated. Maybe it's someone using AI just to, because they have 100 recipes they want to talk about, but they only have time for one to do it themselves, right? If they can do 100 at once, it's better.

34:28But again, it's going to evolve, and it could get either really, really bad, where a lot of this turns into that. But I also see a movement where people are being better about what kind of placements you end up on, right? What are the size of the placements? Where is in the user viewport and all that thing? I mean, good. It's all to the good. I want to talk about Georgia Pacific, which you brought up because you just won an Ad Exchanger Award for some of your work with them. Congratulations in September. I think everyone knows what Georgia Pacific is. If they don't, they definitely know their brands.

35:03I mean, they're a humongous paper and napkins and tissue company. If you've used vanity fair napkins at like a dinner party, or if you've used a public bathroom, like you have used a Georgia Pacific TP dispenser, like no doubt. But yeah, so they drastically cut their SSP count and they started scoring inventory before the bid. So they slashed their SSP roster by, correct me if I'm wrong, it was like 80 % or something. They went from three partners to just like a handful, like five or six partners. And my question is more about the pushback that you sometimes get. Like when brands aggressively put filters on supply upstream, the argument against it is that they risk choking off their reach.

35:51Georgia Pacific didn't have any pacing issues though, like their viewability went up, their CPMs dropped. But they just proved that the majority of SSPs are selling just duplicated noise. I mean, is that the takeaway here? Like not put too fine a point on it, but I feel like the proof is in the numbers. From that angle, yes. I mean, if you look at a lot of the bid requests that come in, and that's one of the other pieces that we really focus on is deduplicating bid requests. So that way, when you see a bid request through SSP1, you're not going to see it through SSP2 because we already know that we're prioritizing the auction and pushing it into the DSP.

36:30So a lot, if you even look at Ads TXT, go to any big website that has ads and do forward slash Ads TXT and you'll see how many SSPs are direct sellers. Every direct seller gets probably north of 85 % of all bid requests on that site, right? So some is as high as every single SSP will see that bid request. There's no throttling whatsoever into the SSPs. And because of that, all those SSPs more than likely send it to the same demand sources. And so you're not choking off access to anything. By actually working with a more consolidated partner set, you get access to do a lot more because now you can send in more diversity.

37:12Right. Even if you had, you know, even with Georgia Pacific, when they went from their number down to a handful of them, it worked. It worked better because then they can focus on where is this bid request more efficient and which path is it coming down more efficiently? And how can I take advantage of that? Right. And so that they can deliver better and more efficiencies. They can bid better. It's a first price auction now. It's on a second price auction at the end of the day. And also it's not that the highest price always wins in a header bid. It's based on they have their own yield management functions around like pacing and making sure they get enough from each supply path, etc.

37:50So yeah, you're not choking it off. And also that's the other piece why this industry is moving to more of, hey, do more things on the supply side. So you pre-filter out exactly what you don't want. And then the highway lane, to make it in simplistic terms, that gets filled, is filled with stuff you actually want. That lane is still going to get filled because there's enough supply out there. The question is, what do you want? Unless you're targeting like three users in New York City, you know, there are constraints where you do run into pacing issues, but largely you wouldn't have them. To simplify it even further, it feels like less is more or can be anyway, if you approach it correctly.

38:28Yeah. And it's a lot easier dealing with a handful of partners on the supply side because they also want to, you know, a lot of SSPs also want to do good work, right, with their brand and agency partners. This gives them a little bit more to kind of lean in and engage with as opposed to like, you know, 30 random partners. And Georgia Pacific, I read something about how they're holding off on autonomous buying agents until the waste is cleaned out of the system because like what's the point of having an agent automate waste faster? Do you agree with that? Like if we don't fix the supply chain first that we're just setting ourselves up for, this is like a contradiction in terms, but just like efficient waste.

39:11Like if we don't do all of that stuff upstream, then all of the shiny, exciting stuff that people love putting press releases out about, like that's just not really going to work. Yeah, I agree. I mean, agents will buy whatever they've been told to buy, right? in terms of it. Like what an agent is, it's just an automation at the end of the day. I mean, we put a fancy label on it and say what you will about it, right? And I'll gladly go grab a drink with someone and talk about it at length if they want. But it's an automation. It's saying that, hey, I want my campaign to be set up this way. Can you just get it done for me with a new campaign name and a few other things?

39:48Yes, over time, it'll learn a little bit more as it builds up its knowledge base and repository and all that stuff. but if what it's learned off of is your historic data which has been buying waste it's going to keep buying waste it's kind of like hey do you want to restock the shelves with lemons over and over again without doing anything so yes I think you know just handing the keys over to an agent to just start spending money is probably worse than having some human kind of oversee it and make sure it's set up properly because you know an agent will want to since it's an automation play instead of setting up one line items, they're probably going to want to set up a hundred line items.

40:25And then a human would typically set up a handful of line items, but they have more control over it because they know when something's working or not working at the end of the day. I'm so in the mood for cherries right now, by the way. I'm going to close this podcast down. I don't think they're in season right now. I know, but I live in New York. I'm sure I can get cherries somewhere. There won't be good cherries. Maybe there'll be lemon cherries, but... Frozen cherries. Frozen cherries, there you go. So one more bring us home. Everyone out there, not everyone, but a lot of people out there are writing obituaries for the open web right now between AI search summaries, eating traffic, signal loss, wall gardens growing and growing.

41:06That's not even a new story. The moves that Apple is making. It's just like a little bloody out there. I will say shout out to you guys for a second time because you won another ad exchanger award for this work that you did with Butler Till for Geloso Beverage Group. They were running campaigns for Club Tales, which is this ready to drink canned cocktail thing. And they were dealing with like massive bid duplication where the exact same ad slot was showing up across dozens of different PMPs and resellers simultaneously. it was just duplicated and messy. But by filtering out that noise before bidding, surprise, surprise, they managed to cut their CPAs by like a large amount, just under like 50%, which is not too shabby.

41:54If cleaning up the plumbing can still drive that kind of performance, I mean, does that offer a realistic survival blueprint for publishers or are the macro headwinds and I listed only a few of them facing the open web just like too strong to stop um I think you know I go back to the idea of did linear tv die and uh and uh you know give me give me a second on this one it's dying it's just but is it but didn't we say this like 15 years ago that it was dying my mom canceled cable she's 76 and I mean if Shirley Schiff is canceling cable they have troubles. Yes, but almost 15, 16 years after we thought it was already dead, right?

42:43I'm just saying, I don't think open web, like I think, yes, you'll see fewer monetization on placements that would have never worked and we just bought wasted placements, essentially, wasted spend. I think the places where there is better inventory, better placements, better interaction and better outcomes, you'll see it actually do better. I think if anything, publishers will be able to yield better CPMs. Same money in their pockets, it'll seem more expensive, but it'll deliver better outcomes at the end of the day. I think we've just, this world of just going down to 50 cent CPMs and dollar CPMs has turned into this MFA environment.

43:24And I think we're going to go, we should potentially go the other way because publishers also don't want to see this drying up. Now, where open web people think about is dying is like, do we read enough on the web overall? And are people going to engage or are they just going to use LLMs to answer all their questions in life? That's to be seen. I am more of the glass half full that's saying that I think we've over-indexed on LLMs and the usage of it. It's great. I use it every day to figure out my grocery list and make sure it keeps me honest with what I want to cook this week. But at the end of the day, I think it's because it's cheap and subsidized now.

44:00And I'm curious what happens over the next 12 to 24 months as these companies go public and it becomes unsubsidized and they need to show growth. Am I willing to pay$200 a month for a subscription? Probably not because I am not that busy in life that I couldn't open up an Excel file and create my own grocery list. Add cherries to that, frozen cherries. And hey, anyone who wants to buy Ravi a drink and debate the definition of an agent, they should reach out, right? Yeah, reach out. I mean, you know, we're all amongst friends in this industry. So LinkedIn, our website, Father Allison, and then she'll get you to me.

44:41I mean, frenemies, if not friends. So this podcast is sponsored by Snap. Ditch your outdated marketing plan and spend smarter with Snapchat for business.