00:00:04
and so my research deals with online advertising attribution and it's easier
00:00:09
to explain it by giving an example so which most consumers actually know so let's suppose you went on a brand's
00:00:17
website let's a gap and you put into a shopping cart a shirt and you didn't buy
00:00:21
the shirt and you don't continue to surfing the internet what would happen typically that you'll start seeing ads
00:00:26
for the same shirt again and again and again on every website now the question from the advertiser's perspective which
00:00:32
is gaps gaps retail is if in the end the consumer did buy the shirt how do you measure the effectiveness of each one of
00:00:40
those ad exposures the consumer head and this is called attribution problem now what most advertisers do today is they
00:00:48
use something called less touch attribution or last click attribution and the idea is very simple if me is a
00:00:54
consumer and I saw let's say 17 different ads and 17 different websites and then purchase the product the last
00:01:00
ad they saw before purchasing the product is going to get all of the credit for this sale ok and motley me
00:01:05
research is trying to do is you trying to determine is this method efficient or not and what is this impact of this
00:01:11
method on different types of advertisers and different types of publishers in the
00:01:17
online marketplace in my research what I try to say is I try to ask is this last click method
00:01:26
which is used by the majority of advertisers efficient and why do they use this method because it's a bit
00:01:32
counterintuitive why would they look only the last add that the person sees and what I do is I compared two types of
00:01:39
compensation schemes which are used on line one is paying for every exposure of an ad being seen and one is paying just
00:01:45
for every sale like a commission on a sale that the publisher gets from the advertiser so my key takeaway is that
00:01:51
actually paying for exposures is more efficient and more profitable for the advertisers than paying a commission on
00:01:58
sales she's very counterintuitive much people think if you compensate a sales
00:02:02
person you should pick them a commission only when they make sales and not when they show off work but if you compensate
00:02:08
a publisher let's say Facebook and you can pay Facebook for just showing the
00:02:12
ads or actually generating sales actually my key takeaway is that paying for showing an ads is more efficient and
00:02:20
more profitable for the advertiser the reason being is that if you have more than one publisher let's say Facebook
00:02:25
and Google are competing what you will see is that there is going to be free writing when one of them on the other so
00:02:32
because of this class click method are going to compete among one another etc and one of them is going to claim
00:02:37
they're gonna be very very very efficient although maybe to just show the last ads etc so this is one key
00:02:43
takeaway the second one is that the last click method creates this competition between publishers that creates kind of
00:02:52
a race to try and cheat the advertiser out of money which is a bit strange but the best way to explain it is just to
00:02:59
try and think what would you do if I told you that the last add your show gets the commission and you know a
00:03:05
consumer comes to your website and there's a high probability the consumer is going to buy the product and there's
00:03:10
another consumer that comes to the website and you don't know if the consumer is going to buy the product to
00:03:14
which consumer will you show that you will show that the consumer is going to buy anyway and just bomb the consumer
00:03:20
with ads to try to be the last one so the other key takeaway is that when you're using last click or last touch
00:03:26
methods all of the measurements and all of the metrics you're getting are going
00:03:30
to be maximized you're going to get a lot of sales through a lot of channels
00:03:33
and publishers but maybe they didn't cause any of those sales or just happen to be
00:03:38
the last ones so the first conclusion which surprised me what actually that paying for
00:03:46
impressions or exposures is more profitable than paying for commissions because if you ask any advertiser online
00:03:51
they say I want to pay only Commission I went apparently when sales are being generated and also this was my intuition
00:03:57
when I started this research and this was very surprising to find and when I tell this even to managers and
00:04:03
executives today they basically don't believe me although I can show them the
00:04:06
math and show that it just works the other surprising finding was that last click and last touch is not necessarily
00:04:12
a bad method per se under some conditions it increases the profits of the advertisers but these conditions are
00:04:19
quite limited so if you ask advertisers why do you use last click and last touch
00:04:25
typically they say it's just the best method they know so far but I can show
00:04:29
that if it happens that your campaign has these specific conditions it's actually better to use last touch this
00:04:36
conditions are very specific is typically when the first ad has the biggest impact so there's a huge impact
00:04:42
to the first ad and not that not that many to the other ones and consumers don't visit the websites too often then
00:04:48
the last click method is very very efficient one takeaway you can take is that realizing that whatever attribution
00:04:59
method you're going to use over time the publishers are going to learn how to
00:05:02
maximize this metric so even if you told the publishers I'm going to compensate
00:05:07
you according to who runs the fastest someone will be the fastest runner and they're going to be the best maximized
00:05:13
method so the idea is to try and sit down with the publishers and with the company you're using for doing
00:05:19
attribution and try to think is the model that they're using makes sense so does the model make sense to my
00:05:24
consumers and the other thing you need to do is you need to be very carefully design an experiment and say what would
00:05:32
happen when using a new attribution method and what would happen without using it you can just move to a new
00:05:38
method see all of the metrics maximized and say oh I just made more money because probably this isn't correct
00:05:48
so one interesting misperception it's there's these ads color retargeting ads
00:05:53
these are very specific ads that you only start seeing them after you put let's say a product in the shopping bag
00:05:59
it in the industry they believe these ads are very very useful and actually are are very good at reminding the
00:06:06
consumer to come back to let's say gaps website and purchase this shirt that
00:06:10
consumer was pondering to purchase and my research showed that actually most of these retargeting ads are probably very
00:06:19
inefficient and have no impact on the consumers the only thing is that they're
00:06:24
being shown last before purchase because if a consumer went to a website and put
00:06:29
something in the shopping cart there's much higher probability the consumer is
00:06:32
going to buy the product anyway so if you start showing that consumer ads compared to consumers who just didn't
00:06:38
visit the website and they didn't put anything at the shopping cart it would
00:06:41
appear you have a huge lift a huge increase in sales on those consumers you started showing ads too but even if you
00:06:48
don't show them as they will buy anyway that's the point so the misperception is
00:06:53
basically if you don't run an experiment and have a control group and says when
00:06:57
these consumers who visited my website on half of them I do show ads on the others I don't show ads and only then I
00:07:03
measure the difference you're going to think that retargeting is a very very
00:07:08
useful channel and typically it or not that but not that good my method is unique in that I use game
00:07:18
theory and I actually did mostly in analytical and theoretical analysis of this problem and only then I tested it
00:07:24
on the data all of the other research on this topic used a very empirical approach that took a set of data they
00:07:30
ran algorithms on it and they try to figure out which publisher is better than the other one what I try to say is
00:07:37
I try to say what is the market using currently in measurement and compensation what would be the
00:07:42
consequences and can we observe that in data and to that I basically found out that a lot of the empirical applications
00:07:49
and measurements are basically incorrect they assume that the market is efficient
00:07:54
in behaving correctly but if you do the theoretical analysis you will see that you can figure out from the data if the
00:08:00
market is inefficient or efficient and in that case it's not very efficient the
00:08:04
publishers are basically showing ads to consumers that will buy anyway which is called adverse selection or essentially
00:08:10
trying to cheat the publisher while the other empirical applications cannot differentiate between really showing ads
00:08:17
to consumers who are influenced to consumers who just would buy anyway and you kind of stick them with ads to get
00:08:22
the Commission after working on these three shots I started getting a lot of communication
00:08:32
from companies saying hey Ron this is very interesting and there's this approach that you're suggesting but
00:08:37
there's at least 10 other companies that claim their attribution product is better or is the best how do we compare
00:08:44
how can we even tell which one is better etc so one idea i'm trying to push and
00:08:50
this would be probably with a few companies to experiment with is to try to build a system where you generate
00:08:57
advertising data and you try different attribution methods and you see which one just matches this company type of
00:09:03
customer and the reason this is interesting is that most a tribunal companies don't actually publish size a
00:09:09
lot very transparent about the algorithms which are say we have the best algorithm and if you give us data
00:09:14
we can give you the results what we won't really tell you what's going on so
00:09:18
this approach where you kind of test through many of them and kind of run a competition between all of them is going
00:09:24
to hopefully tell us which method works better or not that we can understand this process much much better
00:09:45
you