Latest / The Yield Doctor with James Deaker / The 900 Optimizations Your Ad Ops Team Never Gets To | Frans Vermeulen, Swivel

Transcript

The 900 Optimizations Your Ad Ops Team Never Gets To | Frans Vermeulen, Swivel · The Yield Doctor with James Deaker

0:00
I am James Deika. I am the yield doctor.
0:03
And today's discussion is part of a
0:05
project that I'm embarking on to talk to
0:07
companies that are driving change in the
0:09
new AI landscape for advertising. And
0:13
specifically, I'm talking to those who
0:15
are using it to drive new solutions for
0:18
publishers and adtech providers. And I'm
0:21
very transparent about the fact that I'm
0:22
doing this to make sure that we have a
0:25
better take on all the interesting and
0:27
different companies out there. partly
0:29
for my own benefit, but I also believe
0:31
that there are a lot of people out there
0:33
who have looked at slides of the
0:35
Lumiscape or some other material and
0:37
they don't really know what those
0:39
companies do or maybe they know what
0:40
that company did 5 years ago. So today
0:43
my guest is the president of Swivel,
0:46
France Vermuan. And France has been in
0:50
the digital advertising space for 25
0:52
years. And he has worked for companies
0:54
>> 25 years
0:57
>> worked at companies like Doubleclick,
0:59
Freewheel, Comcast and TransUnion. And
1:02
the specific motivation for today was
1:04
that I was talking to a former colleague
1:06
uh that I was catching up with and I
1:08
asked him about what companies are doing
1:10
interesting things and he said to me you
1:12
know France he is at swivel they are
1:15
doing a lot of yield management and you
1:17
need to understand what they are doing.
1:19
And the final aside is that France is
1:21
married to Joanna Foil who I used to
1:24
work with at WP and she has been a
1:27
previous guest on the yield doctor
1:29
podcast. I mean, listen, I James, I am I
1:32
am very used to playing second fiddle to
1:35
to Joanna at this point. So, I'm I'm
1:37
happy to continue the trend uh in in
1:40
playing this sort of secondary role.
1:42
>> I enjoyed the conversation with her and
1:44
I'm going to enjoy the conversation with
1:46
you too, France, and I'm so glad that we
1:48
made this happen.
1:51
>> Tell me. Yeah, pleasure to be here.
1:52
James,
1:53
>> what what is the origin story of Swivel?
1:56
How did it get started? How did you end
1:58
up there?
1:59
>> Yeah, I I mean it's it's kind of an
2:01
interesting story. Um, you know, we we
2:04
the founding team, myself, Joe Hirs from
2:07
Springerve, uh, Matt Dearborn from
2:09
Spring Serve, and Rich Lynn, all come
2:12
out of the the Spring Serve um, uh, uh,
2:16
company, which which traded a couple
2:18
years ago to to Magnite. Uh, was
2:21
fortunate enough to to meet the the, uh,
2:24
the Spring Surf team. I think it was
2:26
like 2018 or you know somewhere around
2:29
there 2017 2018 you know pre-acquisition
2:33
by magnite uh I joined joined the board
2:36
with with another friend of mine um and
2:39
you know when I first met the spring
2:40
serve guys just having come out of you
2:42
know being from freewheel from from day
2:44
one and then all the way through through
2:46
that acquisition which which is also
2:48
dating myself back you know by Comcast
2:51
in 2014 and then working in adtech M&A
2:54
on behalf of Comcast acquiring a couple
2:56
other businesses in and around Free Will
2:58
to kind of beef up that publisher ad
3:00
tech stack. You know, between myself and
3:02
and and the guys from Springerve, um you
3:05
know, when I met them, I kind of looked
3:07
at the Spring platform. I'm like, "Wow,
3:09
this is kind of like a baby next
3:11
generation version of Freewheel." I
3:13
joked, you know, and as we got into
3:15
working with them on the product and
3:17
kind of where to take the business, um,
3:19
you know, along the way, taking an
3:21
investment from from Magnite and then
3:23
the eventual sale of Spring Serve to
3:25
Magnite. Um, you know, that's kind of
3:27
the origin of of how we how we met each
3:30
other. Um, you know, a couple years
3:32
pass, you know, we were we were
3:34
fortunate on my end to sell a business
3:36
to uh TransUnion, as you mentioned, a
3:39
business called Churoptic, which we
3:40
sold, I think, in 200 uh 20 during the
3:45
pandemic, which was a weird time to try
3:47
to sell a business over Zoom, but we did
3:49
it. Um, you know, but both Joe and and
3:52
myself and the rest of the team, you
3:53
know, after these deals, you you
3:55
typically have to stay at the acquiring
3:56
company for a couple of years. You know,
3:59
in our case, we were both kind of coming
4:00
out of that that window of of I joke a
4:04
little bit of of forced servitude. Um,
4:07
candidly, I really enjoyed my time at at
4:09
TransUnion after the acquisition. But,
4:12
you know, at the end of the day, I think
4:13
we we are entrepreneurs at heart. You
4:16
know, like you mentioned, uh, we've been
4:18
at this for for a long time, and I kind
4:20
of know good when I see it. Um, and so,
4:23
you know, I got a call from Joe and he
4:25
was like, I got this, you know, Joe's
4:26
from Alabama. Uh, I got this call and he
4:29
was like, "I got this great new idea.
4:31
We're gonna automate ad operations." I'm
4:33
like, "Well, h how are we going to do
4:35
that?" He's like, "Well, why don't you
4:36
just kind of make an investment?" And,
4:38
you know, as we were sort of completing
4:40
that conversation, he called back. He's
4:42
like, "I got a better idea for you. You
4:44
know, why don't you come on come on in
4:46
as as a as a co-founder and president
4:49
and help us run this thing?" And, you
4:51
know, that that was uh almost three
4:53
years ago to to the date uh here here in
4:55
October of 2026. So about 3 years ago,
4:58
we we all sort of got together and uh
5:00
there was a very very very early stage
5:03
product at that point. Uh we had a
5:05
couple of interesting anchor clients. Um
5:08
and uh it was just a it was a great
5:10
idea. It was exactly what I wanted to be
5:13
doing with a group of people that, you
5:16
know, I knew that I I liked cuz we had
5:18
spent a lot of time at Spring Serve
5:20
together working and thinking about, you
5:22
know, stuff. So it's it's a group I can
5:23
get along with. It's a group that kind
5:26
of really knows their stuff. Um, I
5:28
really liked the problem uh and the
5:30
opportunity to to try to automate a lot
5:33
of the ad operations process in
5:35
particular on the publisher side where
5:37
you know we've just spent so so so many
5:39
years you know even back to to
5:41
doubleclick pre Google you know I was
5:43
working on the the dart for publishers
5:46
product which is now Google ad manager.
5:49
So, you know, for me, having spent, you
5:51
know, a couple years in the data and
5:53
identity space at at True Optic and then
5:55
and then TransUnion, I I was kind of
5:57
psyched to get back to my roots and, you
5:59
know, back in ad server land and and and
6:01
nerdy yield topics like this one um and
6:05
the systems that you need um and many of
6:07
them exist obviously and have been
6:09
around for a long time. And and we did
6:11
kind of uncover this this interesting
6:13
opportunity in and around, you know,
6:15
automation. That's that's the hopefully
6:17
the the quick version of the you know of
6:19
of of the origin story at least how how
6:21
I got involved.
6:22
>> That that's great and I'll I'll press
6:24
you on a little bit more what the
6:25
company does. But one thing that's funny
6:26
is you mention how you know the problems
6:29
never really get solved. I I talked to
6:31
someone uh a couple weeks ago who had
6:34
not really been in the space for about
6:36
10 or 15 years and they said they went
6:38
away, they've come back and the problems
6:40
are just the same problems they were the
6:43
first time around. It's the exact same
6:45
set of problems and it's, you know, the
6:47
exact same set of publisher ad platforms
6:50
that have been around for, you know, 10
6:52
10 or or more 20 20 years at this point.
6:56
You know, nothing had like really
6:58
fundamentally changed in publisher ad
7:00
server land. Um and and um you know
7:03
hence you know some some of these
7:06
platforms are really good at certain
7:07
specific things and you know Springerve
7:09
was different than freewheel and and
7:11
solved a slightly different challenge
7:12
which was let's call it a more
7:14
programmatically oriented ad server at
7:16
at this at the time at springerve. So
7:19
you know there had been small evolutions
7:20
but they had typically been platforms
7:22
that you know are good at like a
7:24
particular thing you know and that makes
7:26
sense because you know as you're
7:27
building platforms the focus is really
7:29
important. Um but but again there there
7:31
had been no like major shifts in like
7:34
let's call it a a you know a larger
7:36
enterprise class publisher. You know
7:38
it's even more complicated in the in the
7:40
video and and TV space. You know free
7:43
wheels 15 years old plus you know at
7:45
this point as well. There just hadn't
7:47
been something kind of materially
7:49
changing and hence we saw saw the
7:51
opportunity and and we've been chasing
7:53
it and obviously I assume we're going to
7:54
get into that a little. So we we will
7:57
and so I ask pretty much everybody the
7:59
same question which is explain what your
8:01
company does uh as if you're explaining
8:03
to somebody who'd been in the industry
8:05
for a number of years but other than
8:06
knowing the name of your company didn't
8:08
know what you you did and and I guess
8:10
the context here uh given I said people
8:13
who have been in the industry they
8:15
probably know freewheel but obviously
8:17
your background is very heavily around a
8:19
lot of the CTV ad servers so as you're
8:22
explain explaining what you do how much
8:24
of that is CTV specific speific the
8:26
answer may be everything I don't know
8:28
but if you could just talk about the
8:29
problems that you're out there solving.
8:31
>> Yeah I mean let's talk about the sort of
8:33
the fundamental problem that we're
8:35
trying to solve and you know think and
8:37
you know we can start with TV and a TV
8:39
centric or a streaming centric you know
8:42
broadcaster and you know I've been
8:43
spending a lot of time in Europe and I'm
8:44
actually heading over tomorrow to go
8:46
talk to even more broadcasters. It's
8:48
it's the same challenge on on both sides
8:50
of the pond, you know, and globally. The
8:53
the opportunity to to solve for what
8:55
we're solving though is bigger than, you
8:57
know, just TV and streaming. It's just
8:58
where we started. Um, you know, like
9:00
like any company with limited resources,
9:02
we we started with what we knew. Okay.
9:05
And sort of what we know is if if you're
9:07
a very large broadcaster, you know,
9:09
think uh NBC and and and folks like
9:12
Univision and WBD now merging with
9:15
Paramount, you know, there's a a couple
9:17
of interesting dynamics. Um you know,
9:20
the the sort of economic dynamic has
9:22
been around for a long time, which is as
9:25
I grow my media business as a publisher,
9:27
right? Just think about like gross ad
9:30
sales dollars traded. I don't care if
9:32
it's direct or or programmatic. you
9:34
know, for every 50 or hundred million
9:37
dollars in spend that you process as a
9:38
publisher um or an enterprise media
9:41
class company like this, you know, for
9:43
every 50 or 100 million bucks that you
9:45
add, you know, there's always been like
9:47
a linear relationship between uh that
9:50
growth in spend and the number of people
9:52
that I need to do things like
9:56
respond to RFPs, traffic orders, traffic
10:00
programmatic deals, you know, manage
10:03
sort of pacing and delivery for
10:05
guaranteed deals, yield optimize,
10:08
programmatic deals, you know, so like
10:11
you keep sort of having to throw bodies
10:14
at the problem and and no one has really
10:16
been able to scale um this challenge,
10:20
right? Maybe 3 to 5% of every odd dollar
10:22
kind of goes to the people that it takes
10:24
to operate these systems. And a large
10:27
part of what they're doing is super
10:29
repetitive, right? even in, you know,
10:31
you're the yield doctor, so we'll we'll
10:32
focus a little more on on the yield
10:34
stuff up front. And that's where we
10:35
started the business is automating
10:37
various yield practices. Um, but like at
10:41
the end of the day, we're trying to
10:42
automate as much of this as possible.
10:45
And ultimately give like one ad operator
10:48
the power of 15 to 20 people's worth of
10:51
work. And we can get into a bunch of
10:53
examples of of kind of like what that
10:55
looks like. So that's sort of like
10:57
business dynamic or or sort of the core
10:59
problem that we're solving for and it's
11:02
complicated by the complexity and the
11:05
number of systems that it takes to run
11:07
an enterprise class media business. I
11:11
lean a lot on on Univision. We've we've
11:13
published a case study with them. You
11:15
know, they have a very complicated set
11:17
of systems that they use. I won't name
11:20
all of them to protect the innocent. Um,
11:23
but let's just kind of talk through a
11:25
couple of them. There's a there's an ad
11:27
server. There's a primary ad server.
11:29
There's a different ad server that they
11:31
use for a different use case. There's a
11:34
video programmatic or a streaming
11:36
programmatic ad serving platform. They
11:38
use a different platform to buy media.
11:41
Think about like a reach extension use
11:43
case. They have a different ad server
11:45
for audio. They have two different order
11:47
systems. Some handle the traditional
11:49
linear side. another order system
11:51
handles kind of like the digital side of
11:53
what we might call order management or
11:56
MS type systems and like they barely
11:59
talk to each other and that's because of
12:02
like that stack fragmentation back to
12:04
our original point where like there
12:06
haven't been any major innovations um I
12:08
think in this world so not only are you
12:11
throwing more people at it but you're
12:12
dealing with multiple systems and those
12:14
systems all kind of move at their own
12:15
pace in terms of you know becoming more
12:18
advanced or more automated within
12:20
themselves.
12:21
But the opportunity to kind of
12:23
orchestrate and automate between like a
12:25
large and complicated set of existing
12:27
systems, you know, is another business
12:29
dynamic that like is really important to
12:31
how we think about the world.
12:32
>> So, so let me push you on that cuz one
12:34
of my general questions is where do you
12:36
sit in the ad stack? You obviously you
12:39
you've said previously that you're not
12:41
an ad server. I think
12:43
>> we're never doing that again. I think
12:44
you've said that you're not an SSP, but
12:47
as you also mentioned that many
12:49
customers already have a very
12:51
complicated fragmented stack and the
12:54
last thing they want is another
12:55
application sitting in there. So do you
12:59
replace other applications or do you
13:01
integrate and provide sort of the glue
13:04
across different applications and if so
13:07
do people log into you instead of
13:10
logging into some of the other
13:11
applications?
13:12
>> The answer is yes. Um, I will never
13:15
build another ad server again. Uh, I can
13:18
say that with confidence at age 52. I
13:21
will never build another forecasting
13:22
system again. That was the bane of my
13:25
existence for for many years and is a
13:27
very difficult math problem and there's
13:29
plenty of companies out there trying to
13:30
solve just that specific problem. Um,
13:33
but at the end of the day, the way we
13:35
wanted to design this and if you think
13:36
about the evolution of these various ad
13:39
platforms that I've mentioned, you know,
13:41
there there have been these APIs, not to
13:44
get too nerdy and technical, but I I
13:46
hope that's what the audience is is
13:48
looking for a little bit in this
13:49
conversation. You know, the APIs have
13:51
sort of gotten better. I can pull data
13:54
from some of these systems. They're very
13:56
different, by the way. like some some
13:59
systems are easier to pull data out of,
14:01
other ones are quite a bit harder or
14:03
something like a reporting API doesn't
14:05
exist. You know, you think about the
14:08
other side of the fence, which is well,
14:10
if a human's going to log into an ad ad
14:12
platform and take some kind of action,
14:15
it almost like doesn't matter what the
14:17
action is or the button that they push
14:20
or there's a slider bar in the UI of the
14:22
ad platform that they want to manipulate
14:25
under different business conditions as
14:27
dictated by, you know, even just kind of
14:30
basic ad server data is what people look
14:32
at. Um but the APIs have improved a lot
14:35
over the years and based on that as well
14:38
as all the stuff that we talked about
14:40
adops is expensive and again these
14:42
people are using you know multiple
14:44
systems for for very specific use cases.
14:46
We're really just an an orchestration
14:49
layer that sits on top of an existing
14:52
ad server or MS or multiple ad platforms
14:55
like we've mentioned. Um and we create
14:58
rules that a human, you know, would have
15:01
taken under certain conditions and we
15:03
start by automating that. If this
15:06
happens in freewheel, then go turn this
15:09
thing up in a in a different ad server,
15:11
be it, you know, Spring Server or some
15:13
other ad platform. And by doing that,
15:15
you're not ripping and replacing any
15:16
systems. I think that's kind of like
15:18
critical thing number one. you know an
15:21
ad server migration having done I cannot
15:23
count how many of them over the years is
15:26
like a huge exercise right and now I've
15:29
got it's further complicated by you know
15:32
you know enterprise companies that have
15:34
multiple platforms for multiple use
15:35
cases you rip out one and it kind of has
15:38
these like ripple effects throughout the
15:40
business both on sort of the data
15:42
aggregation side you know as as as well
15:44
as sort of the features and
15:45
functionalities of each of those ad
15:47
platforms you know but the hope for us
15:48
is that the APIs are there, you know, we
15:52
we kind of are only limited by the APIs
15:55
of the underlying ad platforms that we
15:57
support. If there's a, you know, what
15:59
was the the old Apple thing? There's
16:01
like an app for that, you know, we need
16:03
there to be an API for that. If there's
16:04
an API for that, there's a lot we can do
16:07
as this sort of rules layer or
16:09
orchestration layer, you know, without
16:11
having to replace any any given system.
16:14
And and that's the hope and that's what
16:16
we found, you know, over the last 3
16:17
years. Okay, we are now more than 15
16:21
minutes into our conversation and you
16:22
haven't said AI which is honestly
16:24
refreshing given most of these
16:26
discussions that I do. So instead of
16:29
asking how are you using AI, let me ask
16:31
a slightly different question which is
16:33
where as you build out where are you
16:36
investing in innovating the most?
16:39
>> Sure. There's a a couple of answers to
16:41
that. You know if we think about the
16:43
origin of swivel let's call it three
16:45
years a lot of what we were doing was uh
16:48
built on the spring serve platform
16:50
initially we built the spring serve
16:52
platform it happens to have extremely
16:54
robust both read and write APIs and we
16:58
knew that we could create this rules
17:00
engine that ver that worked very neatly
17:02
in in a complimentary way to automate
17:05
what a human would be doing. You know,
17:08
we sort of philosophically talk a little
17:09
bit about this is like our number one
17:11
goal with clients is to replicate what a
17:12
human would do. That's step one. Step
17:15
two is we want to or step two, you know,
17:17
is we want to scale it, right? So we you
17:20
making a couple go into a change log
17:21
with an ad platform. Maybe a team is
17:23
making a couple hundred optimizations a
17:25
day. That's probably a lot. You know,
17:27
maybe there's a team of, you know, 30 40
17:30
people kind of looking at data, figuring
17:32
out what's optimized that day, and then
17:33
actually going in and making the
17:35
changes. You know the step two is
17:37
scaling that. So we go from a couple
17:39
hundred changes a day to a virtu
17:41
virtuous a virtually unlimited set of
17:44
actions that could be taken per day. But
17:46
those rules have to be set by humans. So
17:49
back to sort of the the AI of it all.
17:51
What we started by doing was really like
17:53
an automation layer. You know, of
17:55
course, there was AI in our development
17:57
relatively early on, you know, but using
18:00
natural language to impact a change or
18:03
create a rule wasn't inherent in the
18:06
system. So, we used to refer to this as,
18:08
you know, an a set of automation tools.
18:10
You're almost thinking about it like a
18:12
deterministic script that can be created
18:14
to replicate and then scale and then
18:16
ultimately step three improve human
18:19
behavior. with the benefit of hindsight
18:22
and a new CTO here at Swivel uh uh you
18:25
know at the beginning of this year, a
18:27
friend of mine from from Freewheel. Um
18:29
now it's a much more sort of native AI
18:32
story. But the key part is like you
18:34
can't have an LLM running a muk in a
18:38
multi-billion dollar, you know, TV and
18:41
streaming business or or even a more
18:42
traditional kind of web type publisher
18:45
or traditional, you know, print web type
18:46
publishers. think that the Hondes and
18:48
and the hearse of the world like the
18:50
worst thing that we could have done
18:52
early on is kind of walk into a senior
18:54
leader and be like, "Oh, we've got this
18:55
AI thing that's going to make all these
18:57
yield decisions and it's going to change
18:59
your price wars and it's going to change
19:00
your pacing." That would never have
19:02
flown. Um, you know, AI can make
19:05
mistakes, you know, and over the years
19:07
the system has become much more AI and
19:10
natural languagecentric where now you
19:12
can just use natural language to kind of
19:14
create a rule. that rule lives in a very
19:17
different layer of the system which is
19:19
the part that actually you know impacts
19:21
or make changes or actions in an ad
19:23
server. Um and that rule that that sort
19:26
of layer of the system is is a very
19:28
deterministic layer. So, wherever we can
19:30
use AI, you know, in natural language to
19:33
say traffic this campaign or hey,
19:36
whenever the trade desk is bidding more
19:38
aggressively from midnight to 4:00 a.m.
19:40
incre, you know, increase my price floor
19:43
by hour by 5% per hour until the data
19:46
says to turn it back down. That's just a
19:49
natural language operation which
19:51
effectively loads like a deterministic
19:53
script, right? And that's the thing
19:55
that's actually looking for those kind
19:57
of if then business rules and actually
20:00
making a change. But only when the human
20:03
strategy has been set. So human history
20:07
>> but the the the the rule set is is
20:11
driven and created by natural language.
20:13
It has to be approved by a human. But
20:16
then it just goes right and it goes in
20:18
perpetuity until you change it, until
20:20
you edit it, until you say, "Hey, you
20:22
know, our our our yield strategy in Q4
20:26
where we're we're pretty full sold out
20:28
uh or we have high sell through is very
20:30
different than, you know, the day after
20:31
Christmas or or January 1st when
20:33
everything kind of falls off a cliff. So
20:35
then it just becomes this like ongoing
20:37
evolution and responding to supply and
20:39
demand dynamics, you know, in and around
20:41
the yield strategy. that humans drive
20:43
that and you know
20:44
>> and G give me a sense for the actions
20:47
and how they're being replaced by your
20:49
system. Is it is it one to one meaning
20:52
an action that was previously taken by
20:53
an adops person is now done by the
20:55
system or are you taking additional
20:57
actions that weren't previously
21:00
able to be done by humans? Both both are
21:03
true, right? And and and again like to
21:06
build trust with these large publishers,
21:09
you know, we start small, right? you we
21:11
we kind of use this this sort of a an
21:13
analogy of like well you know let's say
21:16
I'm a yield manager at a at a big
21:17
enterprise publisher and you know I run
21:19
a report and there are a thousand things
21:21
that I could optimize today my team only
21:24
has the time to optimize or make a 100
21:27
changes today okay based on the data so
21:31
what is the human team going to do by
21:33
nature they're going to focus on you
21:35
know an optimization that drives $500
21:38
incremental a day or $1,000 of
21:40
incremental a day and that I'm going to
21:44
do those you know 100 changes out of the
21:46
thousand things that I could do and then
21:48
I'm going to pull the same report
21:49
tomorrow and I'm going to say hey is
21:50
this is this actually driving you know a
21:52
better result be it revenue or yield or
21:54
margin or whatever the objective is you
21:56
know at the moment but when you're
21:58
automating and creating rules for all
22:00
thousand things that you could have done
22:03
which a human team could never have done
22:05
you know now we're making a thousand
22:07
optimizations a day just to keep the
22:09
math simple cuz I'm terrible math, which
22:11
doesn't make sense for being on the
22:13
yield doctor podcast. Um, but from a
22:15
strategy perspective, right, like if I'm
22:17
if I'm optimizing all 900 things that a
22:19
human team didn't get to, and each of
22:22
those, you know, below the top 100 were
22:24
worth $50 or $20 or $5 or 25. When you
22:28
make tens of thousands of those optim
22:31
optimizations per day, the money adds
22:33
up. The incrementality adds up.
22:35
>> So, how do you how do you prove that
22:37
out? Cuz one of the questions I was
22:39
going to ask you actually is is your
22:41
pitch to um publishers more on the cost
22:46
reduction side or on the revenue uplift
22:49
side? But I'm going to take it from the
22:51
what you just said that there's a
22:53
revenue uplift. How do you prove that
22:55
out? Because I'll tell you as somebody
22:56
who's been in this industry a long time.
22:59
It's very easy to tell a a good yield
23:01
management story, but then actually
23:04
measuring and proving out the impact can
23:07
be harder. So, what's your experience
23:08
with that?
23:09
>> Yeah, I mean, listen, we we started the
23:12
company with this assumption that it
23:13
would be all about efficiency. Um, but
23:17
you know, it's a challenging media
23:18
environment, challenging economic
23:20
environment even even 3 years ago. A lot
23:23
of these companies are emerging. There's
23:24
a lot of pressure to do more with less.
23:27
So when we can come in and say well
23:30
listen for every you could think about a
23:32
yield strategy that a human sets as like
23:34
a yield agent. I might have 50 yield
23:37
agents live at any given moment which
23:40
are looking for those you know nuggets
23:42
in the data that says hey now is time to
23:44
turn a price floor up or a price floor
23:46
down or you know um kind of offer more
23:51
bid requests under these conditions but
23:53
not those conditions. So, you know, the
23:57
way that we structure all of these yield
24:00
agents is each one has a little like
24:03
hold back or control group. Um, the
24:06
methodology for that can vary a little
24:08
bit depending on the yield strategy.
24:10
Some yield strategies or yield agents
24:12
are much more focused like think about
24:14
price warning for example. you know,
24:16
that's a very explicit revenue focused
24:20
um an incremental revenue focused
24:21
strategy simply by holding back um you
24:25
know, a 10% hold out or doing some AB
24:28
testings. It it varies a little bit, but
24:30
every yield agent has their own little
24:32
control and we can sort of isolate as
24:34
many of the variables as possible to
24:35
show that this strategy is working. You
24:38
know, maybe we started it on the bottom
24:40
20% of my inventory. let's increase the
24:43
coverage of that automation up to 50%.
24:46
And thus the incrementality the gross
24:48
incrementality basically you know
24:50
continues to to grow as the coverage of
24:52
the automation or the coverage of the
24:54
inventory increases. So you know we we
24:57
it's it's a traditional kind of like
24:59
hold back group. The nuances again vary
25:02
a little bit depending on kind of what
25:03
the strategy is and and kind of what the
25:05
timing of it is.
25:08
>> Changing directions a little bit. I want
25:10
to talk about data
25:12
>> and particularly we all know especially
25:15
for a large publisher how fragmented
25:18
their systems can be. There are leakage
25:20
issues, there are latency issues. When
25:24
you go into a new client and you're
25:26
looking under the hood, um what is the
25:30
sort of type of issues that are most
25:33
problematic or or holding them back from
25:35
taking the most advantage of your
25:37
system?
25:39
in particular with a publisher who's
25:41
running different ad servers for
25:43
different purposes. Like I mentioned
25:45
early on, like the API capabilities tend
25:49
to vary platform by platform. Um, you
25:52
know, I I I pick on Spring Serve and
25:54
Magnite only because they're good,
25:56
right? The data is relatively real time.
25:59
Anything that you can do in the Spring
26:00
Serve UI, there's an API or an app for
26:03
that. Um and but that's different than
26:06
say you know freewheel or you know
26:09
publica or even Google ad manager. So
26:13
one of the interesting sort of um you
26:16
know I don't benefit is the wrong word
26:18
um but but sort of dynamics that we
26:21
found is because we understand natively
26:24
how we can access all these systems by
26:27
uh what was originally a lot of API
26:29
integrations now it's moving towards
26:31
kind of more of an MCP based you know
26:34
set of more AIcentric protocols
26:36
including ADCP and and Tony's IB
26:38
framework and just you know basic MCP
26:41
integrations and protocols. You know, we
26:44
actually understand that in freewheel if
26:46
you want to change a price for that
26:48
field is called price. In a different ad
26:52
platform, it could be CPM. In a
26:54
different platform, it could be eCPM. In
26:56
a different platform, it could be called
26:57
oranges. Right? So we've we've kind of
27:00
assembled this Rosetta stone across the
27:03
platforms that we support where we
27:05
understand not only hey how do we kind
27:08
of normalize that a price floor in
27:10
general in freewheel means this field
27:13
and it means a different field in a
27:15
different ad platform. But as we pull
27:17
data from these systems, we also are
27:19
able to normalize that. Well, whenever
27:20
we get data from platform 2, we know
27:22
it's four hours delayed. And if it's
27:25
four hours delayed, it changes the
27:28
strategy that you want to follow a
27:29
little bit. And so, you know, that's
27:31
just kind of the nature of how we work
27:33
and some of the dynamics that, you know,
27:35
we have to deal with.
27:37
Um staying on the philosophical in terms
27:40
of where our industry is going, one of
27:42
the uh pain points is that relative to
27:45
life, let's say 15 years ago when there
27:48
was a direct sales team selling to an
27:50
agency or an advertiser directly, it was
27:53
a simple world and life has got more and
27:55
more complex along the way. And you are
27:58
obviously trying to help simplify the
28:00
world of adops operators or at least I
28:03
assume you are. Where where does this
28:06
end up? Does the publisher world end up
28:09
being more in control? Cuz I know for a
28:12
long time now, frankly, it's felt a
28:14
little bit out of control. What do you
28:16
see the next few years look like?
28:17
>> I mean, one of the first things that we
28:19
heard from clients in particular around,
28:21
let's call it, you know, programmatic
28:23
yield optimization and the automation
28:25
thereof is like they felt like we were
28:27
giving a little bit of power back to the
28:29
cell side. And I think that is an
28:32
interesting perspective, right? Like if
28:35
you think about in in programmatic and
28:37
just think about the buy side in general
28:39
concentrated you know amongst you know
28:41
four very very large agencies and then
28:43
another couple of very large
28:44
independents and then another 2,000
28:45
agencies right the power more often than
28:48
not has sort of lit and access to data
28:51
and you know information asymmetry I
28:53
want to be able to cherrypick audiences
28:55
out of my DSP and I'll pay more for this
28:57
audience but I want to tell the
28:58
publisher like all that stuff's been
29:00
going on and I think that's what you're
29:02
referring to for for years now you know
29:04
but by sort of enabling an unlimited set
29:06
of optimizations and automations
29:08
powering those you know it does give a
29:10
lot of power back to to the the
29:12
publisher um and and that's just a
29:15
dynamic and a control dynamic that I
29:17
think you know behooves us and behooves
29:19
our our publisher partners and frankly
29:21
it also you know is is beneficial to to
29:24
the ad platforms that that are being
29:25
used kind of under the hood as well.
29:28
you've you've been in the space a really
29:30
long time and I hope you have a
29:32
contrarian position. So I I'll ask you
29:35
uh what is something that you believe
29:38
that is different than the majority of
29:40
people either in adtech or AI
29:42
development of what's going to happen?
29:45
Man, that's a tough one, James. Um, you
29:48
know, what's a contrarian position that
29:51
um might uh might be sort of more
29:55
common? I you know,
29:58
at this point, I would say Swivel is is
30:00
an is an AI native automation platform,
30:04
right? We we've we've been able to
30:06
automate a lot of stuff, not just yield
30:08
and programmatic and pacing. Um much
30:11
more interestingly or or as
30:12
interestingly at least we we've been you
30:14
know investing in in ADCP
30:17
um the concept of like agentto agent
30:20
trading uh is still in its infancy and
30:24
this is a classic sort you know this
30:25
stuff it's a classic chicken and egg
30:27
problem reminds me a lot of when when
30:29
SSPs and DSPs first kind of came into
30:31
being let's say in the in the right
30:33
media type days um you know I think
30:37
there's a sort of a fallacy on two
30:40
fronts. One, you know, is that AI is
30:42
going to like solve all these problems
30:44
and and it's not like, you know, another
30:47
one I I kind of heard a lot was like,
30:49
ah, man, with natural language, like the
30:51
UI is dead. Um, I kind of believe that
30:55
for a little while, you know, while
30:57
while working on on this business, and
31:00
the reality is that that's it's not
31:02
going to be the case. It's it's a it's
31:03
hybrid, right? The answer, as usual, is
31:06
a little bit of both, right? You still
31:08
need a structured piece of software when
31:11
you're running a multi-billion dollar
31:13
media business with very specific rules
31:16
that need to be followed and this buyer
31:18
has these rules and this buyer has those
31:20
rules. You know, those rules kind of
31:22
live in like a lot of different places
31:23
and and I kind of think there's like a
31:25
false belief that, you know, AI is going
31:27
to solve all of that and like you're not
31:29
going to need to like log into a UI
31:31
anymore. I I don't believe that as much
31:33
anymore. You know what we have found is
31:35
is by putting those two, you know, a
31:37
natural language interface and a smartly
31:40
designed piece of software that need to
31:42
be sort of designed developed by our
31:45
product folks and our engineering folks
31:47
very closely in parallel. Um, you know,
31:50
that that's kind of like the the answer.
31:52
There is no, you know, sort of natural
31:55
language interface that's going to like
31:56
solve all these problems for these very
31:58
very kind of large, you know, enterprise
32:00
class media companies. Yeah, that's
32:03
probably one. Um, you know, for years,
32:07
if we think about the traditional way
32:08
that advertising is is transacted, you
32:11
know, direct sold or human to human
32:13
sold, there's this sort of, I guess,
32:17
viewpoint that that can largely go away,
32:19
but that didn't go away with
32:21
programmatic. You still need people out
32:22
there building relationships and doing
32:24
deals. In the direct sold world, we've
32:26
been trying to automate forever. I don't
32:29
think it's gotten very far at all, but
32:32
you can automate responding to RFPs. And
32:35
you can do that across any media type or
32:37
media channel that you have. Most of our
32:40
customers still have moderately large or
32:43
billions of dollars of traditional
32:44
linear TV. They sell podcast and audio
32:47
inventory on top of all the the
32:49
streaming stuff. So, you know, will
32:53
sales go away? That's also a fallacy.
32:56
Will age to Asian trading take over the
32:58
world? Probably not. But can it add a
33:00
lot of efficiency to the process? Sure.
33:03
Can you give a sales team the power to
33:05
respond to an a complicated
33:06
multi-channel RFP in an hour? We've
33:09
recently shown a number of our larger
33:11
customers that we can do things like
33:13
that. And again, that's an AIcentric
33:16
hybrid SAS plus AI piece of software.
33:18
Um, those are I don't know, those are a
33:20
couple things that that kind of come to
33:22
mind in the stuff that that we deal with
33:23
every day here.
33:25
One last question before we could come
33:27
to a closing. You've been in the CTV
33:30
space now for a while and it's not an
33:32
area I know very well, but I was talking
33:34
to somebody who is in it and was talking
33:36
me through some of the dynamics have
33:38
changed in the last, let's say, four or
33:40
five years where I think the supply used
33:43
to be very limited, the price is very
33:46
high. Amazon came in and released a lot
33:48
more inventory, created a bit of a glut.
33:51
There's a question about access to
33:53
YouTube. Where do we stand now in terms
33:57
of the liquidity of the whole CTV
34:00
marketplace as you see it and and where
34:02
do you see it going?
34:04
>> Uh another good question. I you know I
34:08
think your framing of that question is
34:10
is is correct. Very perceived as limited
34:13
high value hightouch direct sold you
34:16
know more complicated sponsorships and
34:18
ad products. A good bit of that has gone
34:20
programmatic, but the CTV sort of
34:22
inventory ecosystem
34:25
will never look like the web display
34:29
when I was at doubleclick infinite
34:32
supply of, you know, web visits and now
34:34
app visits um and kind of basic you know
34:37
display type ad units. We will we will
34:40
never get to that point in TV. You know,
34:43
I would argue you're also correct that
34:47
even in programmatically traded CTV
34:50
supply, there's been a lot of pressure
34:52
on price. Um, which is actually where
34:54
our tools have added a lot of value when
34:56
there's pressure on price being really
34:59
microscopic almost like highfrequency
35:01
trading of these various like yield
35:03
operations that that we automate, you
35:05
know, into the tens of thousands of or
35:07
more of actions per day. um like I it's
35:12
still a bit of a constrained market, you
35:14
know, I I still think it's it's a a
35:17
relatively premium experience. Um
35:22
I also, you know, and and that just has
35:24
impact on price and the way that things
35:26
are are traded. So, yeah, there's a lot
35:28
more supply that's out there. Yeah,
35:30
there have been pre there's been
35:31
pressure on on on sort of effective
35:33
rates that that publishers can earn. Um
35:35
but but again, you know, now now there
35:37
are tools to kind of help with some of
35:38
that stuff. You know, long-term, you
35:41
know, I'm I'm pretty myopically focused
35:43
on on our customers and you know, this
35:45
business, you know, linear is declining
35:48
at a much faster rate. Will streaming
35:51
inventory replace that? Not for a couple
35:54
of years. I mean, gosh, I remember when
35:56
when Free Will was was sold to Comcast
35:59
in 2014,
36:01
you know, all the Wall Street Journal
36:03
and New York Times business articles
36:05
were about TV is dead and, you know,
36:08
it's dying and the upfront is dead and
36:10
like, God, I get so tired of of hearing
36:12
that stuff. You know, of course, linear
36:14
is in decline. Um, but but it's these
36:17
are still massive pieces of business.
36:19
You know, people need to be able to plan
36:20
and optimize and and and execute these
36:23
these multi- channelannel campaigns for
36:26
kind of many years probably, you know,
36:28
still still to come.
36:30
>> France, thank you so much for taking the
36:32
time to talk today. Where can people
36:35
learn more about you and about Swivel?
36:37
>> Sure. I mean, I'm just at Fr with an S
36:40
this time, fans. Swivel.ai.
36:43
Um, I'm always available. You can always
36:45
reach out to the the Swivel. ITE um or
36:49
you can just ping me on on on LinkedIn.
36:52
So
36:53
>> thank you so much for joining me. Find
36:56
us. Thank you James for having us.
36:57
>> Thank you. I am James Deer. I am the
37:00
yield doctor.