Native Automation: Behind the Algorithms | AWasia 2017

Native Automation: Behind the Algorithms | AWasia 2017


Thank you, if there’s one thing to know about native, it’s that the native network,
they’re the one in control, it might feel like as an advertiser because the native
network gives you so many knobs to tweak and variables to choose from, that you’re
the one in control and you’re deciding which ads go, where and how much to spend, that’s kind of the illusion that you’re under, but the truth is running an ad
campaigns on native, it’s not like driving a car, it’s more like riding in
the back of a bus and on this bus are 38,000 other native advertisers on this
same network, and the network what they’re doing is they’re looking over
the back of that bus, they’re like driving the bus and they’re really
deciding how to allocate the ads, which ads go where and so, we’re gonna look at
that in a little bit more detail and so you’ll be able to understand from the
perspective of an advertiser, you’ll be able to see how these native networks
work and actually all of their optimisations and all of their decisions
are around one simple formula, so we’re gonna look at that formula and how the
native networks work and how they optimise that formula, first and we’re
gonna understand things from the perspective of the native networks, when
you upload an ad technically from the algorithmic perspective, how the native
network optimises your campaigns and how they test new campaigns, how they test
creatives, we’re gonna go into that from the perspective of the network, first
then we’re gonna look at things from your perspective as a native advertiser,
and for some practical information about how to optimise your campaigns on native
and then at the end we have a little secret, a little secret
algorithm that I’m gonna share with you for how to scale on native, okay, so let’s put yourself in the shoes
of the Native Network, you guys understand a little bit about Native
right, these are networks like RevContent, MGID, Taboola Content ad, Outbrain that’s the network’s I’m talking about here and the way they work is
they’ve got these, you know they’ve got one placement on the site, on the
publisher and instead of like display right, on display there’s just like one
creative, they auction off to the highest bidder, the charge on a CPM basis they take a cut, they give it to the publisher, that’s
display right, native they’re charging on a CPC and
they choose which ads go where, so like for this example, I took this example I’m
from Santa Monica California, I was on my Mac so in the US desktop right, US
desktop and I went to Newsweek com and this is the ad that was served to me
personally, right so think about it from Newsweeks – from RevContents
perspective, they had to serve up this ad, how many ads do you think were legal for
them to show, meaning weren’t blacklisted by the advertiser, so the advertiser had
said I want Newsweek com on Revcontent and the publisher had also not
blacklisted, because one of the things that advertisers or some of us forget is,
that publishers can blacklist things too, so publishers have they can block like,
if they see an exact creative that they don’t like, they can block that creative,
they can block categories of creatives, they could say I don’t want any skin
offers, they can also block like, they have a quality, tiers of quality
that they can block, so they say like pg-13 and above like they don’t, so
publishers can do that kinds of thing, but so anyway how many ads do
you think were available for them to show on this place, I would say maybe
like a thousand, maybe they had a thousand creatives, they could show and
they had to choose eight of them, so what would you do if you were the native
network and you had to choose eight out of a thousand on every single impression
right, so try to think about that, if you’re trying to make as much money as
possible, you’re getting paid per click, and you’ve got a thousand different ads,
which ones are you gonna show, so you got to think about that, I mean one algorithm
that I thought of is how about half the time you just show random ads, and then
half the time you collect data and see which ad made you the most money right,
so clicks, you make money whenever some on it, you know if it makes you the most
money, maybe we can just like do this and show that at eight times right, why don’t
the native networks do that, so I have been asking the people behind these
optimisations at the network, the people who designed these algorithms, I’ve been
asking them these questions, I’m trying to understand how the native networks
work, I’m the CEO of Drive which is a native automation platform and I’ve been
asking them how do you guys optimise, how do you guys run your systems and what
they’ve said is that it comes down to one formula, that
they’re trying to maximise on their end when they’re deciding which ads go on
the placements, and this is that formula revenue, equals CTR times CPC, so revenue
and this is the build, you run this like eight times, if there’s a widget with
eight, sometimes there’s eight positions, sometimes there’s fifteen or sixteen
sometimes, there’s just one but they’ll run this like eight times, and they’ll
see like it’s in our example when you had a thousand choices and you have
to choose eight of them, they’ll go through each thousand of those ads on
there in their system and they’ll say for each of those ads, give me a guess of
the CTR, so the CTR on this exact site on this exact ad, so guess that number, which
as we’ll see is actually the hardest part of this equation, so guess the CTR
number and then the CPC they look that up from there you know, whatever the
advertiser is currently bidding and then they do a combination, sometimes it’s a
first price auction sometimes it’s a second price auction but basically the
higher you bid, the more you’ll pay, so they look up the product of CTR and CPC
and basically, how they work is they, okay so they run this equation you know, a
thousand times and they take let’s say the top eight answers that they get and
they look at those top eight and they do a couple other tweaks here in this phase,
so after they get, after they maximise this number they do a couple other
tweaks, so one tweak they do is we can call like a tweak for a variety, so they
actually learned that you know going back to our example, previously where we
had that same a date, they actually learned that they don’t want duplicates, they don’t want to show the same ad, they
don’t want to show the same creative, they even don’t even want to show the
same category twice, so if you go on native, you’ll very rarely see like, two
diet ads on the same widget or two skin ads, usually what they do is they cap it
to one, usually they say I only want one per category, so that gives them kind of
more of a variety, more likely they’re gonna find someone who clicks, so you
know, going back to that, they take the top eight out of thousand and there’s two
diet ads, they’ll they’ll skip that and go to the next one, so that’s one of the
tweaks they do, they also do some stuff like per user, different networks do
different things, here they also sell some of their inventory on a CPM
depending on the network and sometimes they sell things programmatically, so
there’s some other things on the side but essentially what it comes down
to they also have re-targeting but essentially, what it comes down to is
just maximising this number, these you know finding the top eight of these,
every time okay, so you’re just trying to maximise you know, the top eight of these
and the hardest part here is knowing what the CTR is, so trying to guess
because the CPC we just look up right, the CTR we want to know the exact CTR on
this exact site of is creative, so the way they solve that is they use
something called memory, so memory it’s like kind of like a floppy disk right,
like you’re you’re saving this information and what that means is
instead of them experimenting on like every single site, every single creative,
like a thousand created they have tens of thousands of sites and they have
millions of creative, so like even if there’s only like a thousand creatives
that are valid at a time they still don’t want to have to spend
three thousand impressions every single time, because you know, why do that when
you can just look at trends across your data and make approximations and make
guesses, so they try to make guesses, that’s one thing, the other thing is that
often there’s like a creative that never got that much volume, because it’s a
small campaign or it’s a new publisher without that much volume, so they have to
make these generalisations, so like if there’s a new publisher or a small
they can say oh, you know this is a new site, I already know something about news
sites so they organise their data into these hierarchies and the reason for
this is like going back to our bus analogy right there, driving this bus
they’ve got tens of thousands of advertisers, millions of creatives
running at any given time and the more they can do to look at all that data and
divide things into categories and make sense of their data, they do that as much
as they can so for example, like look on the left where it says publisher
category, so like that’s like a news site right or a humour setter, well however
they divide their publishers up and so if you’re running an auto insurance ad
on Newsweek they’ll look at the performance of this exact ad on this
exact placement, but then they’ll also zoom out and they’ll say okay based on
the performance of this exact ad on this exact placement, I know things about the
advertiser category, so I know that you know auto insurance ads do certain, well
don’t do well on this exact placement, I know that this advertiser I
know that auto insurance ads work well on news placements in general, so they
kind of bubble up they make all these generalisations about the advertiser
about the creative category and the other thing is look on the right look on
the right here where it says creative image, so that means they know that
advertisers are like always stealing creatives from each other right, so if
they’ve seen your creative before from another advertiser, they’re gonna again
take that you know, make a generalisation from that other advertiser about your
creative, because that way they save money on testing, and so they use you
know, this hierarchy of information, they use that memory, so if you upload a
creative, this is why a lot of people don’t really realise this but there’s no
such thing as a true run of network campaign on native because on display,
when you run right off network you kind of expect all of the placements
to get a fair look at your campaigns and kind of the more volume, the placement
has them you know, depending on the going rate on
that placement, it’s kind of fair on display right but on native, you’re not
uploading this campaign in isolation you’re uploading your campaign with all
your different creatives and they’re looking and they’re saying okay, I’ve
seen this creative before, this is where it performs well, this is where it
doesn’t perform well, this category of creative I know that this category of
creative works well on new sites, or doesn’t work well on new sites or
whatever it is, so there you’re uploading the creative sim and they’re routing it,
there driving that bus, they’re in control and they’re writing those
creatives to where they know that that type of creative is gonna work the best,
and then once your creatives actually perform, then they’ll use the most
granular data possible, if there’s enough volume on that exact placement but most
of the time they’re using these trends, and they’re using this memory okay, so
now we’re gonna look at things from your perspective as an advertiser, for some
practical information for how you can use this, so the first is creatives, we’ve
talked a little bit about creatives, the most important variable for native for
creatives is CTR the publisher CTR, so this as advertisers, I mean the thing
about publisher CTR is that we’re not used to like going into the native ad
network, and looking at how many impressions based, because we’re just
paying on a CPC, you have to look at how many impressions they served and then
how many clicks it got on there and so usually this isn’t for a lot of people,
this isn’t even something that they consider looking at to optimise by, but
actually like CTR is such a huge part of their equations that a better creative,
with a better CTR obviously, that’s going to get you more volume and cheaper
volume, cheaper clicks but it has this other unintended consequence that you
might not have realised, which is that it helps other creatives in your account,
and it helps other creatives in this category too, so you’re gonna be helping
your competitors a little bit with a high CTR creative, but you’re mostly
gonna be helping yourself and also your campaign and your whole account, so
creatives CTR matters, one other thing to say about creatives is sometimes, and
this happened to me personally, when you launch a campaign and you see
that certain creatives have a higher CTR and certain creators have a lower CTR,
sometimes like the lower CTR creative will still get more volume or you launch
a campaign with 8 creatives and it seems like certain creatives don’t get a fair
shake right, like certain creatives just got a couple hundred impressions than
they give up and the reason for that is this memory thing that we were talking
about, so they’re know they’re the bus driver right, and they’re looking and
they’re saying, oh you got this new creative already, know where this
creative performs, it performs over here, I already know where this one performs
or someone already ran that creative and you know they’re using their memory,
they’re using their data that they already figure it out, okay, so the next
thing is bidding, bidding I wish there were like a single formula, I
could give you for bidding like a single rule of thumb and the truth is what’s
happened for me a lot with bidding is sometimes, you’ll make a change to your
bids and then there will be an outcome to that change but it won’t be because
of anything you did it’ll be because some advertiser is running your same
creative somewhere else, that’s on the same category and you know that’s why bidding is tough, there’s so many exogenous things, but there are some
trends like in isolation, if assuming nothing on the network happens, nothing
outside you know, it’s out of your campaign, happens there’s some general
trends so first higher bid obviously gets you more volume because
you know you’re more likely to be on you know, in that eight, in that top eight
you’re more likely to win, there and also you’re more likely to show up on the
widget, you’re also more likely to be in a higher position, so like the top
left position on the widget is usually the one that gets the highest volume, so
you’re more likely to be on the top left position, but unlike other kinds of
networks, the highest bid does not always give you the highest traffic quality,
actually sometimes the traffic quality is a little bit worse, seems like this
seems to especially happen on mobile, where it’s maybe, there’s more accidental
clicks or something, anyway it’s hard, I mean, you don’t always know what you know, with the best quality that you’re gonna get with bids, so you just have to really
experiment it depends on the category depends on the exact site you just have
to test different bids see what happens, there’s
some people who bid really high, there some people who bid really low, like try
to get in position seven or eight and then just have a lot more margin, so you
have to experiment for bids unfortunately and just like another
media buying, when you lose traffic you bid higher, okay, so the last thing is
BOTS, so BOTS we know that like half of our industry’s fraud and especially on
the lower tier native, networks bots are a big problem but actually BOTS can be
your friend, BOTS can help you or at least not hurt you as much, as you might
have assumed so there’s a couple thing about the thing about bots, so first
about BOTS, it’s kind of easy to detect bots because the way the bots work, is
they click a link on the publishers page, where they click the creative and then
they visit your lander and when they get to your Lander, they either like never
click on, so what you do is you look at the lander click-through rate, because
the bots either never click on any pages on your lander or sometimes
they’ll click on every single link on your lander, so if the CTR is like the
lander CTR is really high or if it’s really low, then you know that this is a
bot placement, so that’s the first thing you don’t have to spend you know, 2x the
offer payout to know that this is a bot placement, you can just send a hundred
visits to it and see what happens, you could have more sophisticated
solutions like some JavaScript thing for your land or whatever but actually CTR
is good enough, so just look at CTR, so that’s the first thing about bots, the
second thing about bots is once you find a bot placement on a network, then you
know that that’s about placement forever, so you can just block that, you can just
never run a bot you know, never run traffic to it again and that’ll
actually save you a lot of money, especially you have a like an auto block
or tool like an optimisation system, because usually the way that bot traffic
works is it spiky, like usually though if you look at bot traffic, it’ll
be kind of like on a low volume site and then on Sunday morning when no one’s
looking at it, it’ll just spend $300, so if you had like an auto blocking rule
setup that said, okay every time something spends it gets
to 100 visits, block it then that would solve that problem, the last thing about
bots is bots are kind of like a regressive tax, so when you have like a
taxation system right like a progressive tax, you tax the rich, more bots are like
taxing the poor more, because it’s the poor advertiser, it’s
the new advertisers who suffer the most from bots on any given network, so if
you’re an advertiser who doesn’t know where the bot placements are, you don’t
have a good blocking system for bots, then bots are gonna hurt the smallest
least sophisticated advertisers the most, so if we got rid of bots you might have
more competition, okay so finally we have the algorithm that we’ve all been
waiting for, the algorithm that ties everything together, the algorithm to
scale any performance campaign on native and here it is it’s the called the
tiered whitelist algorithm, so this is in three phases, the first phase is to find
creatives, the second phase is to find placements and the third phase is to
find bids, so the first phase requires that you know the top placements on any
category on any geo, which you can either get from your own experience or you can
get from a spy tool or whatever, and what you do is you test all your creatives
and landers on those and you test the creatives for CTR and also for
profitability, so you learn figure out your best lander and the output of step
one is creatives, and lander step two is you use the creatives in the lander, from
step one and you run everywhere, you run a run of network campaign on the
network and you block placements that are BOTS and you also block basements
that don’t get any revenue and step three, so step two is the one
where you’re gonna be spending the most money and you should probably
be losing money on step two, and actually step two is something that kind of
accumulate over time, like you’ll learn over time which are the best placements
on any network for you know, for four different verticals,
so step two is just trying to get as many placements as possible,
step three is where you use what we learned about bidding, so you just test
out different bids on every single placement on every single creative and
this is gonna depend on the network, so like on Taboola you can have on Taboola content you’re gonna have custom bids for every domain on Revcontent you
have to make different white lists, but you just have to experiment with
different bids, keep tweaking them and looking at the profitability based on
the bids, so this should tie everything together, thank you guys very much.

One thought to “Native Automation: Behind the Algorithms | AWasia 2017”

Leave a Reply

Your email address will not be published. Required fields are marked *