0:03

Hello everyone, I'm Eric Golden and

0:05

welcome to Making Markets. This show

0:07

explores the psychology and structure that

0:09

make up markets all over the

0:11

world. Each week we speak to

0:13

experts about a different market so

0:15

you can see what actually happens

0:18

when money changes hands. From mainstream

0:20

stock and bond markets to esoteric

0:22

niches like vineyards, antique art and

0:24

crypto, we explain the transactions that

0:26

underpin our economy. Making Markets is

0:28

part of the Colossus family of

0:30

podcasts and you can find all

0:32

our podcasts including edited transcripts, show notes,

0:35

and other resources at Join Colossus.com. Eric

0:37

Golden is the CEO of Canopy Capital.

0:39

All opinions expressed by Eric and podcast

0:41

guests are solely their own opinions

0:43

and do not necessarily reflect the

0:45

views of Canopy Capital. This podcast

0:47

is for informational purposes and should

0:49

not be construed or relied on

0:51

as investment, legal, or tax advice.

0:53

Clients of Canopy Capital may maintain

0:55

positions in the securities discussed in

0:57

the securities discussed in this podcast.

0:59

including positions that are contrary to

1:01

the opinions offered. My guest this

1:03

week is Eric Soufer. In

1:06

the words of Ben Thompson

1:08

from Stratecery, Eric is arguably

1:10

the best writer and commentator

1:12

on digital advertising. And today he

1:14

gives us a master class on the

1:17

subject. We start by talking about

1:19

Apple's app tracking transparency ATT

1:22

framework, Metis Recovery, and

1:24

what the current ecosystem looks

1:26

like. We also discussed Twitter, now

1:28

X's offerings, advertising on

1:30

the open web, common misconceptions

1:32

about targeted advertising, and

1:35

more. Please enjoy this

1:38

conversation with Eric Supert.

1:40

Eric. Eric, in preparation for

1:42

this interview, I don't think I

1:44

realized how little I knew about

1:46

the digital ad space. I think

1:49

I had some awful assumptions of

1:51

how it worked, so I'm excited

1:53

to have you on today. I

1:55

thought an interesting spot to begin

1:57

would be how digital advertising

1:59

advertising worked before the ATT

2:01

was implemented by Apple, the

2:04

idea that privacy and transparency

2:06

was an idea, and then

2:08

maybe what happened and how

2:10

that all came to be. Yeah,

2:13

sure. And so first of all,

2:15

Eric, thanks for having me. Always

2:17

nice to be speaking with another

2:19

Eric. So I think digital advertising

2:21

has a kind of long and

2:23

nuanced and winding and fascinating history,

2:25

but to my mind, there's two.

2:28

historical paradigms that are really important to

2:30

think about. And we just started the

2:32

second historical paradigm. It's basically three years

2:34

old. And it's a new set of

2:37

operating norms and a new framework for

2:39

thinking about digital advertising performance and reaching

2:41

customers and using digital ads to scale

2:43

products. And so if you sort of

2:45

think about that demarcation line as the

2:47

introduction of Apple's app tracking transparency policy,

2:50

which I think that's where you can

2:52

draw a bright red line, that's July

2:54

2021. What that did, and I'll talk

2:56

in broad terms, then we give on

2:58

back to some more detail if you

3:00

want, what that did was it just

3:03

changed the operating environment for advertisers and

3:05

for ad platforms in the historical paradigm

3:07

before ATT, data flowed freely. There were

3:09

very few restrictions on it. Now there

3:11

were some. I talked about those faint

3:14

red lines that could be marketed, different

3:16

historical paradigms. You could think about those

3:18

as like the different periods of the

3:20

dinosaur age versus just dinosaurs existed and

3:22

dinosaurs died. But there were some limitations

3:24

that were imposed on the flow of

3:27

data between advertising platforms and advertisers, but

3:29

they weren't substantial or they workarounds to

3:31

them or they didn't represent a majority

3:33

of users or of data. But with

3:35

ATT, it really attacked the great ponderance

3:38

of data that mattered for advertisers, which

3:40

was iOS users. But in the previous

3:42

historical paradigm, data flowed freely. A company

3:44

like Facebook could grow to this massive

3:46

scale and size. by being a hub.

3:48

It was a hub for data where

3:51

the advertisers were the spokes. And I've

3:53

called this like the hub and spoke

3:55

model. And the idea was face book

3:57

is a centralized publisher. It says, if

3:59

you want to buy my inventory, you've

4:01

got to come to me. You can't

4:04

go to some third party and buy

4:06

my inventory. You have to come and

4:08

buy it directly from me. I'm a

4:10

walled garden. That's the definition of a

4:12

walled garden. It is my garden. It

4:15

has walls around it. My garden is

4:17

my garden is walls around it. My

4:19

garden is my garden is my garden

4:21

is my garden, my ad inventory. You've

4:23

got to buy my inventory. someone clicks

4:25

on this ad that you bought for

4:28

me what happens after that and that

4:30

was actually a really mutually beneficial dynamic

4:32

for both the advertiser and Facebook so

4:34

the advertiser said sure I'm happy to

4:36

send you more information it's going to

4:39

the ad click why because you're going

4:41

to take that information and you're going

4:43

to use it for the benefit of

4:45

my campaign and that information being they

4:47

bought something they added something to their

4:49

car they registered you had information that's

4:52

all valuable signal now you're going to

4:54

take that information you're going to ingest

4:56

it And you're going to say, okay,

4:58

this was a quality user. They bought

5:00

something. It's a quality user. Well, obviously

5:03

that user's probably not going to be

5:05

advertised to again by that advertiser. So

5:07

why would the advertiser do that? Because

5:09

it could be used to find more

5:11

users that look like that concept of

5:13

a look like audience. And so you

5:16

use this information and you find users

5:18

that look like that user. Now that

5:20

look alikiness is defined by any number

5:22

of factors we don't know that's not

5:24

public knowledge any number of features you

5:26

think about spectral clustering and that kind

5:29

of thing like there's any number of

5:31

features that could go into that into

5:33

figuring out who was like this person

5:35

based on their feature set but Facebook

5:37

did that and they found other users

5:40

that could advertise to that looks like

5:42

that user that was proved to be

5:44

a good user now that benefited the

5:46

that advertiser but it also benefited every

5:48

other advertiser. because you've got that new

5:50

piece of data that this user buys

5:53

pants or shoes or whatever. And now

5:55

I know that about them. And so

5:57

when another advertiser is looking to sell

5:59

that type of item or an item

6:01

that people like that tend to buy,

6:04

then I can use that information for

6:06

that advertiser's benefit. So there was this

6:08

incentive to. contribute to essentially this data

6:10

co-op and Facebook and other public platforms

6:12

such as Facebook. Facebook's the canonical go-to

6:14

example, but lots of platforms operate in

6:17

this way. Search is a little bit

6:19

different. You have to put search aside,

6:21

but YouTube for instance, snap, tick-talk, they

6:23

operate in this way. Search is a

6:25

little bit different. You have to put

6:27

search aside, but YouTube for instance, snap,

6:30

tick-talk, and they all operate in. They

6:32

use this. to having that day to

6:34

be sent back on an app-by-app basis.

6:36

Now you might say, okay, this is

6:38

only iOS, this is only apps. This

6:41

doesn't sound like it's that meaningful, well,

6:43

no. First of all, iOS is the

6:45

majority of revenue. Almost everything's on mobile

6:47

now. If you looked at Metas or

6:49

Facebook's Avenue composition before ATT, it's 95%

6:51

of the revenue is mobile. First of

6:54

all, that answers to the mobile question

6:56

on mobile is not that big. Well,

6:58

yes, yes, it is. The only question,

7:00

no, no, no, it's not, it's not,

7:02

it's not, absolutely. It's only for the

7:05

purposes of the limitation, but it applies

7:07

to wherever that user's going post click

7:09

and wherever the data's coming in. And

7:11

so the thing is if meta or

7:13

Facebook app, Instagram app, don't get the

7:15

permission, then they can't accept that data,

7:18

that data, that Facebook app, Instagram app,

7:20

don't get the permission, then they can't

7:22

accept that. If the ad click leads

7:24

to a website, it's okay, because it's

7:26

not an app. And so that app

7:28

doesn't have permission. It comes down to

7:31

the publisher's permission. Now that click is

7:33

going to an app and they also

7:35

don't get permission. Well, then they can't

7:37

do anything with that data at all.

7:39

But even if they do get the

7:42

permission, unless both Facebook and that app

7:44

got the permission from that user, then

7:46

the data can't really be used meaningfully.

7:48

But even if it leads to a

7:50

desktop browser, first of all, there's no

7:52

IDFA, when you send the data back

7:55

by policy. Facebook is not supposed to

7:57

ingest it and use it for ad

7:59

targeting. The whole point is you may

8:01

not use my data for ad targeting

8:03

after I say no. There's no contingent.

8:06

or conditions on that. And so that's

8:08

the paradigm that we existed. And there's

8:10

been other data restrictions too. GDPR, I

8:12

think, is thrown into that bucket. It's

8:14

not really a data restriction per se.

8:16

It's actually just a bunch of obligations

8:19

that data processors have with respect to

8:21

user data, like the right to be

8:23

forgotten, the right to know what you're

8:25

collecting and all that stuff, consent. But

8:27

this was an actual data restriction. And

8:30

there have been other since then. So

8:32

you've probably heard about Google's proposed cookie

8:34

deprecation scheme that we don't know when

8:36

that's actually going to go into effect.

8:38

They're going to use a consent mechanism

8:40

that sounds like it's going to look

8:43

like ATT. So that'll be another data

8:45

restriction. There's the DMA and the DSA

8:47

and Europe both have elements of data

8:49

restriction. And then actually some interpretations of

8:51

the GDPR have led to some interpretations

8:53

of the GDPR have led to the

8:56

GDP, some interpretations of the GDPR have

8:58

led to interpretations of the GDP, or

9:00

the media. Going back being in college

9:02

in the early 2000s when Facebook was

9:04

being created, I think that I came

9:07

from a generation and talking to friends

9:09

that we just had this dystopian give

9:11

up. I have no control over my

9:13

data, Facebook and Google, because I like

9:15

to use it. I just accept that

9:17

it's everywhere and I have no control

9:20

in that it leads maybe to better

9:22

ads. Was the impetus for ATT driven

9:24

by government or privacy activists? Or was

9:26

it actually a strategic economic benefit to

9:28

places like Apple to say we're going

9:31

to... lessen the power of Facebook and

9:33

Google. I don't know that there's any

9:35

cogent argument that it wasn't just a

9:37

competitive ploy. There essentially is no domestic

9:39

regulatory pressure to institute privacy restrictions. In

9:41

Europe there is, but in the US

9:44

there essentially is none. I mean there

9:46

is no federal privacy law. Now I'm

9:48

talking about those faint red lines. Apple

9:50

has a history prior to entity of

9:52

implementing privacy restrictions. They have a framework

9:54

called ITP intelligence tracking protection in safari

9:57

that blocked third party. cookies wholesale. They

9:59

put limitations on first party cookies. So

10:01

in that way, I mean, Apple had

10:03

a history of doing this. The ATT

10:05

didn't come out of nowhere, but this

10:08

is much more meaningful. ITP and

10:10

Safari, there's no mistaking what the effect

10:12

was. It had a tremendously negative impact

10:14

on CPM, the price of impressions in

10:17

Safari. But there was always an alternative

10:19

in a viable alternative, and actually a

10:21

bigger alternative, which is Google Chrome. But

10:23

when you implement this in IOS, there

10:26

isn't really a bigger alternative on the basis

10:28

of revenue because IOS was the biggest one.

10:30

There is really nowhere for the money to go.

10:32

And so that was one of the big differences.

10:34

But why I think it was mostly motivated

10:37

by... competitive impulses was Apple

10:39

expanded its ad network very quickly

10:41

after that after they implemented this.

10:43

I think they saw Facebook or meta

10:45

as a competitive threat across a number

10:48

of different factors. This is obviously the

10:50

most important part. It is the business

10:52

essentially is advertising. I do think there

10:54

was some sort of sensitivity towards privacy

10:57

within the company, but you don't

10:59

have a company of that size making

11:01

decisions that are this far reaching on the

11:03

basis of vibes or just goodwill. It just

11:05

can't happen. a company that avaricious

11:07

has to only do things that benefit

11:10

it. Now they did use privacy as

11:12

an advertising campaign, you know, Apple,

11:14

this is privacy, that doesn't seem to

11:16

be the motivation for this,

11:18

just to sell or iPhones, I mean

11:21

there is no real iPhone device growth,

11:23

and the device replaces cycles are getting

11:25

longer, I don't think this shrank

11:27

them in any consequential way. But I

11:30

do think the reasoning was a little

11:32

bit more nuanced and just we want

11:34

to a grow, that's like... meaningful to

11:37

Apple at Apple scale. I think that

11:39

the bigger catalyst here, the bigger motivation

11:41

was seeing that ads had become

11:44

the primary distribution mechanism

11:46

for content on Iowa's.

11:48

And if you allow that to happen,

11:50

at some point, the device gets

11:52

commoditized. Most content

11:54

interaction models are moving

11:57

across platform, multi-hardware forms.

11:59

factor configuration. You think about how

12:01

do you interact with Netflix? How do

12:04

you interact with the apps that you

12:06

use every single day, not just on

12:08

one form factor? Well, how do you

12:10

interact with iOS apps on iOS? And

12:12

so if ads become the primary distribution

12:14

mechanism, I think consumption is downstream of

12:16

that. And so if ads or how

12:18

I get content into people's phones, I

12:20

can get content onto all of their

12:22

devices. And the biggest fear that Apple

12:25

had still has, I think is that.

12:27

You lose that very strict one-to-one relationship

12:29

between this content only exists on this

12:31

device and is only monetized on this

12:33

device and allows us to take a

12:35

cut of that monetization. If you lose

12:37

that exclusivity and this content can be

12:39

made available everywhere, then it can be

12:41

monetized everywhere and the developer can sidestep

12:43

the fee. They add revenue is probably

12:46

just icing on the cake, but the

12:48

core services revenue from the App Store

12:50

is very material to Apple. And so

12:52

if you allow ads to be the

12:54

primary distribution mechanism instead of the App

12:56

Store, instead of the App Store being

12:58

the locus of discovery and distribution, if

13:00

you allow ads to do that, then

13:02

ads can't put content to every single

13:04

surface area, every single hardware form factor,

13:07

and then also be engaged with on

13:09

the iPhone. I think that's also why

13:11

they've obstructed game streaming on the iPhone,

13:13

because they don't want to be one

13:15

of many. They want to be the

13:17

only way you interact with this content.

13:19

And so they've given out a lot

13:21

of exceptions to the no app stores

13:23

in the app store rule and the

13:25

ability to monetize outside of the app

13:28

store with the staple products like Netflix

13:30

with the reader app rule and the

13:32

multi-platform exception. But you're fighting very much,

13:34

but you're fighting very much, but you're

13:36

fighting very much, but you're fighting very

13:38

much, with no-platform, but you're fighting very

13:40

much the idea that you can link

13:42

out with no friction or that you

13:44

can have alternative right to monetize. the

13:46

iPhone. And so I think that probably

13:49

plays a bigger role. That makes sense.

13:51

Help me understand from the position of

13:53

the chief marketing officer that controls the

13:55

digital ad spend dollars. Pre-ATT-T, you walk

13:57

through this very symbiotic relationship, I pay

13:59

Facebook, I give Facebook, it gets better

14:01

and better, and I'm happy because what

14:03

I'm trying to do is get in

14:05

front of the right customers at the

14:07

right time to sell my products. So

14:10

how has the money from the spending

14:12

standpoint changed before and after this tectonic

14:14

shift, as you call it? Anybody listening

14:16

up to this point is probably thinking

14:18

like, what is this guy talking about?

14:20

If this was so disastrous to meta,

14:22

why is meta sitting at all time

14:24

high? Especially today as we record two

14:26

days after the election and everything's ripping.

14:28

Seems like meta turned out okay. Yeah,

14:31

well, you also remember meta at 90.

14:33

Things got really bad before they got

14:35

really good and what you'd attribute that

14:37

recovery to is meta executing the wholesale

14:39

re-engineering of the company down to the

14:41

name. Now I think the naming stuff

14:43

was more related to the Facebook files

14:45

and things like that, but they changed

14:47

the core content interaction model of Facebook

14:49

and Instagram. They changed it from the

14:52

Friends of Family Network, Friends Graph, to

14:54

the Open Graph. That is a fundamental

14:56

transformation of the company. And they rebuilt

14:58

the ads architecture too, but let me

15:00

talk about these things and then let

15:02

me answer your question. Because I think

15:04

all the arc of Facebook's navigation of

15:06

this I think is very instructive. to

15:08

understanding the broader ecosystem. What happened post-dinosaurs.

15:10

Facebook had all this data and they

15:13

could find the most relevant ad for

15:15

anyone. So first of all, that didn't

15:17

go anywhere. Like the first age, AT&T

15:19

was in effect. They still had a

15:21

lot of data. The thing is they

15:23

stopped getting new data. Their bank of

15:25

data became stale on a percentage basis

15:27

just over time. That gave them time

15:29

to figure this out. and they also

15:31

have plenty of resources. But how Facebook

15:34

adapted this was, I think, by recognizing,

15:36

okay, our targeting will never be as

15:38

good as it was the day before

15:40

ACT, or it'll never be that precise

15:42

on a user level basis. So what

15:44

do we have to do? Well, we

15:46

have to find ways to increase our

15:48

ad revenue. Well, there's a couple ways

15:50

you can do that. One is increasing.

15:52

So when a user comes into the

15:55

app, they have a specific session length,

15:57

call it like 10 minutes, instead of

15:59

seeing five ads in 10 minutes, you

16:01

see 10 ads in 10 minutes. Well,

16:03

I just showed more ad impressions, I'm

16:05

going to increase my ad revenue. What's

16:07

the problem with that? You can churn

16:09

users. You can churn users. That's the

16:11

problem with that. You can churn users.

16:13

That's not the idea. You can churn

16:16

users. That's not the ideal way to

16:18

churn users. That's not the idea. You

16:20

can churn users. That's not the idea.

16:22

That's not the idea. You can churn

16:24

users. That's not the idea. That's not

16:26

the idea. That's not the idea. You

16:28

can churn users. That's not the idea.

16:30

That you can churn users. That's not

16:32

the idea. That you can churn users.

16:34

That's not the idea. That's not the

16:37

idea. That you can churn users. That's

16:39

not the idea. That's not the idea.

16:41

That's not the idea. That you can

16:43

churn users. The other is you increase

16:45

the time that people spend. So ad

16:47

load is the same per minute of

16:49

time, but they spend more time in

16:51

the app. Well, that's great. That's actually

16:53

a really good thing to do. You

16:55

make the app more engaging, everybody wins.

16:58

And the fourth way is you make

17:00

the ads more valuable, everybody wins. And

17:02

the fourth way is you make the

17:04

ads more valuable. You do that through

17:06

targeting, and that's off the table. So

17:08

what they did was a combination of

17:10

your friends and family. That's a constraint.

17:12

If it's not limited by anything, if

17:14

it's open to the entirety of the

17:16

Facebook user base, well, there's much more

17:19

of a set of opportunities of things

17:21

to show you. And so if you

17:23

get really good at picking across that

17:25

much larger pool of what to show

17:27

you, they could get you to spend

17:29

more time in the app. So they

17:31

did that in two ways. But when

17:33

it went open graph, then it shows

17:35

reals. It's video. It's more engaging. People.

17:37

It's more engaging. People. It's more engaging.

17:40

People. People. People. People. People. People. People.

17:42

People. People. People. People. People. And they

17:44

have grown the user base, not on

17:46

a percentage base, it's pretty small, quarter

17:48

of a quarter, but they still grown

17:50

it. The targeting, there's just nothing much

17:52

they can do because they've lost access

17:54

to that data. But there was another

17:56

opportunity for Facebook to grow its ad

17:58

revenue is to decrease waste. So, it's

18:01

to decrease ad revenue is to decrease

18:03

waste. So, okay, if I can't make

18:05

targeting better. I can at least reduce

18:07

the ways that advertisers are wasting their

18:09

money. Let me take the reins. from

18:11

managing their campaigns. And that's AI-driven ads

18:13

optimization. That's what they call Advantage Plus.

18:15

Advantage Plus is a suite of tools

18:17

that advertisers can use to make all

18:19

the decision-making related to their campaigns, like

18:22

budget optimization, bid optimization, which adds to

18:24

show, audience targeting experimentation. It's basically a

18:26

one big experimentation engine across all levers

18:28

that the advertiser has. for making their

18:30

campaign perform better. And where with human

18:32

media buyer, you've got vacation time, you've

18:34

got weekends, you've got sleep, you've got

18:36

the lack of ability to spin up,

18:38

10,000 different ad targeting slash ad creative

18:40

combinations, their machine learning infrastructure doesn't. And

18:43

so they call that AI, that's fine.

18:45

I call it just plain old machine

18:47

learning. But what they did was they

18:49

made those automation tools available to everyone,

18:51

tools available to everyone. or ads optimally

18:53

efficient at whatever level of potential existed

18:55

for that company. So they did that

18:57

on the advertising facing side, and then

18:59

on the user facing side, they just

19:01

increased engaging. So people are spending more

19:04

time in the app, they're seeing more

19:06

ads. Right now, the ads are not

19:08

getting more valuable in and of themselves.

19:10

People are just seeing more of them,

19:12

and therefore the revenue goes up. And

19:14

they built this model. They eliminated these

19:16

compartmentalized smaller models in favor like the

19:18

much bigger models that are computationally more

19:20

intensive. but perform better, they made these

19:22

big models for doing all of that,

19:25

for doing ad selection, for doing content

19:27

selection, which really are quite the same.

19:29

But they invested in this AI infrastructure

19:31

that allowed them to pick content better

19:33

to make the experience stickier to increase

19:35

engagement time, which allowed them to show

19:37

more ads. But in doing all of

19:39

that, ad targeting remained structurally worse than

19:41

it was. So what does that do?

19:43

That opens up the opportunity for people

19:46

that can target. because they have native

19:48

contextual relevance to a specific commercial goal

19:50

or outcome. Call Facebook, they were the

19:52

everything store that's pre-medier. Well, and they

19:54

still are, I mean, every advertiser goes

19:56

there, but they're not as capable of

19:58

linking up, caught like some niche products

20:00

to a user. there's just not enough

20:02

information. I talk about the measurement piece,

20:04

if we want, so that goes even

20:07

further into the weeds. But essentially, they're

20:09

not getting the data back to know

20:11

that this specific user, like this specific

20:13

thing, so they're having to move one

20:15

abstraction layer higher and say, well, when

20:17

we advertise this kind of stuff to

20:19

the groups of people that we can

20:21

define with the data that we do

20:23

have, it tends to work out. So

20:25

that's what they've done. They're not targeting

20:28

individual users anymore. They're targeting individual users

20:30

anymore. Well, what did that do that

20:32

open the door to the companies that

20:34

can say, wait a second, we have

20:36

all of this first party data about

20:38

our users. And it is very granular.

20:40

It is very specific. We didn't collect

20:42

it from some third party. We collected

20:44

it through native use of our products.

20:46

And what we can do is we

20:49

can monetize that because Facebook can't anymore.

20:51

They can't target those individuals based on

20:53

their individual preferences. But we can because

20:55

we collected that data with first party

20:57

privileges. And so it's fair game to

20:59

use and targeting because HTT doesn't limit

21:01

that use because we collected the data.

21:03

We didn't collect it from some third

21:05

party. So that created the everything is

21:07

an ad network. dynamic, which allowed every

21:10

company that has any sort of first

21:12

party data to spin up an advertising

21:14

business, because they already had the asset,

21:16

which is the data. All they had

21:18

to do is stick some ad impressions

21:20

in the product. And so this ushered

21:22

in this era of everything being an

21:24

ad network, because now there's a premium

21:26

on having first party data that has

21:28

commercial relevance based on context. Whereas Facebook

21:31

was the everything store, I'm the beauty

21:33

store for ads. And so on that

21:35

point, I did find fast saying that

21:37

it led to this for companies like

21:39

on my phone, I was playing a

21:41

Bass Pro or Instagram, I'm buying products

21:43

on that website. So I don't have

21:45

to opt in, they know my purchases,

21:47

they know what I'm interested in. Did

21:49

they start to then increase their ad

21:52

revenue away from a place like all

21:54

that money going to Facebook? So the

21:56

way performance advertising works is not zero

21:58

sum like that. If you offer. performance

22:00

for my ads, I will spend as

22:02

much as I can on your platform

22:04

while also spending as much as I

22:06

can on every other platform. To some

22:08

degree, there is a competition for the

22:11

budget, but in the long run, in

22:13

the grand scheme of things, there isn't.

22:15

Because of the way performance advertisers

22:17

think about return. Let's say you

22:20

go into some scaled app or

22:22

some ecom company or like some DDC

22:24

brand and you go talk to their

22:26

head of marketing. You ask them, what's

22:28

your row as? For what channel? For

22:30

Facebook, our 90 day row as is 1.1. So

22:32

we spend a dollar on Facebook, we get

22:34

a dollar 10 back within 90

22:36

days. But no one talks about

22:38

row as in the abstract. They

22:40

talk about it, usually channel specific,

22:42

sometimes geo channel specific, but always

22:44

time index. But thinking about that,

22:46

if you're making 110% in 90

22:48

days, where else can you get

22:50

that return? Well, it sounds great.

22:52

Keep going. Yeah. 110 which which

22:54

compounds. Which compounds, which compounds.

22:57

and turns into another 1.1 and any

22:59

other days. And so the thing is,

23:01

because of the generally short turnaround time

23:03

in this money, you could borrow money. If

23:06

you could prove this out, you could

23:08

borrow all the money you need to

23:10

invest on a 90-day basis. Essentially infant.

23:12

Essentially uncapped. And so if a new

23:14

channel emerges and says, hey, look, we

23:16

can get you 1.1 at 90 to spend here,

23:18

I can put on my credit card

23:20

and make money on that channel. There's

23:23

not a fixed limit to my budget

23:25

and I have to decide. Again, to

23:27

some extent there is, and I can

23:30

talk about the nuance of that, but

23:32

it's probably not interesting to this audience.

23:34

That's more like a performance marketing audience,

23:37

strategizing around the waterfall, and I've got

23:39

limited head counts, I can't operate on

23:41

every single channel, but you think about

23:44

this, just from the abstract sense, essentially,

23:46

infinite budget, because you could always borrow

23:48

against 90-day return on 1.1 if it's

23:51

demonstrable. channels will offer you a credit

23:53

in essence. They might bill on a 30-day term,

23:55

so I just need to fix my row as

23:57

30 days and I'm playing with their money. But

23:59

there's... to do some relationship here. I'm

24:01

investing across all channels that can deliver

24:03

the ROA standard that my CFO says

24:06

I need to achieve. Let's say that

24:08

I'm a local restaurant or whatever. It's

24:10

no-brainer to go on Uber Eats. I'm

24:12

a food brand, no-brainer to go on

24:14

Insticart, but the list of these, these

24:17

are called retail media networks generally. These

24:19

companies that emerge, they've got retail data,

24:21

and they've just spin up an advertising

24:23

business on top of it, generally called

24:25

retail media networks. So they've proliferated. there's

24:27

many of them now I think there's

24:30

probably too many more than the market

24:32

can bear because again you do run

24:34

into operational constraints as I think we

24:36

will start to see some either consolidation

24:38

there or bundling more just some of

24:41

them failing but you saw this mass

24:43

proliferation of these retail media networks that

24:45

did have contextually relevant first-party data for

24:47

some vertical not universally not like Facebook

24:49

ad not for everybody but if you're

24:51

operating in that space you want to

24:54

reach customers that buy the stuff that

24:56

you sell then you go there Now

24:58

the King Kong of all retail media

25:00

networks is Amazon because they are the

25:02

everything store. They were the first everything

25:05

store and they have an advertising platform.

25:07

And so they were I think one

25:09

of the principal beneficiaries of ATT and

25:11

they were the one essentially that disrupted

25:13

the duopoly. I wrote a piece back

25:16

in 2019 saying can Amazon upset the

25:18

duopoly? So I think it's going to

25:20

become a triumvirate. And that's essentially what

25:22

happened. They basically maintained this double digit

25:24

growth rate as the big walled gardens

25:26

of ATT. because a lot of that

25:29

money, the DDC money and the retail

25:31

money, was shifting to Amazon because they

25:33

have the data in a first party

25:35

capacity, first party privileges, and they have

25:37

an ad network. But all the retail

25:40

media networks that emerged following that, some

25:42

of them have grown to become very

25:44

big. Can we talk about what happened

25:46

with X and their ads? I know

25:48

there's all a little politics involved, but

25:51

it seemed to me... Back when the

25:53

first model, pre-totonic shift, it was the

25:55

hub and spoke model, if you had

25:57

all the eyeballs, you could advertise and

25:59

you could make money. And then this

26:01

change upset the duopoly and now other

26:04

people have opportunity. Why isn't an app

26:06

with that popularity have the ability to

26:08

generate ads? That is a very interesting

26:10

question. I think. client in X's ad

26:12

revenue stems from a self-inflicted wound. But

26:15

following the election I question whether that

26:17

wasn't the intention all along. But let

26:19

me start with X or Twitter prior

26:21

to Elon's acquisition. So Twitter actually had

26:23

pretty decent ARPU. If you look at

26:25

the last quarter for which Twitter was

26:28

a public company, it's ARPU was $5.

26:30

It's a little bit different because Twitter

26:32

had reported M Dow. So it's not

26:34

exact Apple's to apples to oranges. with

26:36

Twitter. But call it adjusted, our poo

26:39

was $4.96, and this is globally, meadows

26:41

was 982. So it was about half

26:43

of meadows, but it was more of

26:45

the snaps, it was more of the

26:47

pictures. So Twitter is somewhere in the

26:50

middle of the path. What is it

26:52

in our poo, average revenue per unit?

26:54

What is the unit for social media

26:56

companies? User, user, revenue per user, but

26:58

quarterly. And their revenue essentially had billion

27:00

a quarter. Snap, last quarter was 1.1.1.1.1.

27:03

with digital ads. They never really built

27:05

out the mechanism for receiving that data

27:07

back. I talked about the Hubb and

27:09

Spoke model. They weren't really a Hubb

27:11

and Spoke model. You went there and

27:14

you targeted it against people's interests or

27:16

their social graph. But you didn't really

27:18

target against knowledge of their purchases because

27:20

you didn't really target against knowledge of

27:22

their purchases because Twitter never really built

27:25

the infrastructure for that. It's called that.

27:27

as essentially a cost of revenue. And

27:29

so you want that that scales. It

27:31

scales to meta-size business. That's what you'd

27:33

prefer to build. Twitter is like always

27:35

has been like more of a brand

27:38

oriented platform and that the SMEs were

27:40

spending, but I couldn't really scale that

27:42

budget. It wasn't a direct response. Direct

27:44

response means I'm spending money because I

27:46

want you to positively associate. sentiment with

27:49

my brand and the next time you're

27:51

in market for something you buy it.

27:53

So there's a difference. So if a

27:55

brand spend comes from like the very

27:57

biggest companies, the companies where next time

27:59

you walk into the supermarket, you're going

28:02

to buy their products. positive associations with

28:04

their brand, not some DDC brand where

28:06

they want you to click and buy

28:08

it right in that instant. And so

28:10

this is the difference. SMB support comes

28:13

from building the direct response walls. That's

28:15

the measurement, the attribution, the targeting. Brands

28:17

spend just means you get viable. That's

28:19

all you need. Following the Elon acquisition,

28:21

the targeting, Brands spend just means you

28:24

get by. That's all you need. Following

28:26

the, he's making certain voices heard on

28:28

the platform, and they just stopped spending.

28:30

That has nothing to do with data

28:32

privacy. That's purely his customer base just

28:34

didn't want to buy from him after

28:37

he bought the company. It's fascinating the

28:39

difference between how powerful the engagement model

28:41

was for meta to help advertisers. You

28:43

talked about this semi-relationship. I send you

28:45

information, you send me information, which led

28:48

to this engagement, which led to this

28:50

flywheel, which led to this super company.

28:52

And then the fact that Twitter had

28:54

all that, presumably, and never built that,

28:56

it's kind of a fascinating thing to

28:59

me, I didn't realize. And also that

29:01

they preference towards this non-scalable, slow growth,

29:03

very high-touch customer relationships of brand advertising

29:05

because brands are very fickle. They're a

29:07

little bit hard to manage. They have

29:09

a lot of demands. And there's not

29:12

that many of demands. And there's not

29:14

that many of them that spend large

29:16

amounts of money. Whereas there's lots of

29:18

SMBs. And the thing is what you'd

29:20

prefer to do is you'd prefer to

29:23

build a system that's essentially fortified. Do

29:25

you remember the boycott? All these brands,

29:27

Boycott had met a, it wasn't even

29:29

a blip in their earnings. And Zuckerberg

29:31

even said that in that earnings, he

29:33

said, like, look, a lot of people

29:36

think that our business is built on

29:38

a handful of big averages. No, it's

29:40

not. It's built on 10 million SMBs.

29:42

One or many big brands turn their

29:44

spendoffs. We won't even really notice it.

29:47

So genius. Moving to Google, the other

29:49

side of it, and the stuff with

29:51

cookies. I don't think I fully followed.

29:53

I fully followed. this whole thing of

29:55

websites are tracking me. What is Google

29:58

doing with this notion of getting rid

30:00

of cookies. I can talk about what cookie

30:02

is. There's two types. The distinction

30:04

is important. So a cookie is a

30:06

file that's saved onto your hard drive

30:09

in your browser that just stores some

30:11

information about your interaction with that website.

30:13

The way it's used in advertising is you

30:16

have what's known as like a third party

30:18

cookie. A third party cookie means, well, the

30:20

site itself doesn't store some information

30:22

in your browser, but actually third

30:24

party does. We invoke some JavaScript

30:27

from a third party source. they save this

30:29

information into your browser and then

30:31

that information is useful to know

30:34

who you are usually kind of like 90.

30:36

So that sounds like people might not

30:38

object to that. I was like, okay,

30:40

well, some third party that I don't

30:42

know of is saving something my browser.

30:44

Maybe I'd prefer they wouldn't do that,

30:47

but I don't see that to be totally

30:49

objectionable. It's probably benign. But the

30:51

problem with that is that third

30:53

party being some sort of advertising

30:55

company. They can't really do much

30:57

if they know who you are.

30:59

They need everybody to know who you

31:01

are. What that company either wants

31:03

to do is sell ads for you to

31:06

some advertising company, or they want to

31:08

bid on the inventory that is being

31:10

sold to be shown to you. So

31:12

they're either on the SSP side or the

31:14

DSP side. I'm selling the inventory that will

31:16

be shown to you. If I know who

31:19

you are, but the people that are buying

31:21

that inventory don't know who you are,

31:23

they don't know how to bid. If I'm

31:25

an advertiser. You're on that website, I

31:27

could identify you because I've seen you

31:29

elsewhere, I have an ID for you, but

31:31

that website, the company that's selling the

31:34

inventory for you, doesn't really know you

31:36

are, but that doesn't help anybody either.

31:38

So what we do is, the company that has their

31:40

third party cookie placed in the website, they

31:43

do a thing called cookie sinking, where

31:45

they essentially invoke all of these other

31:47

services, and there could be thousands of

31:49

them that are essentially to vote, because

31:51

you could chain invoke. They say allow them

31:53

to, through this header system, they'll essentially allow

31:55

all of them to know who you are. So

31:58

we've gone from one company setting a third.

32:00

party cookie to essentially sharing that cookie through

32:02

this process called cookie ID through this process

32:04

called cookie sinking with another service. And then

32:06

that service who didn't interact with you in

32:09

any way shares that idea with a bunch

32:11

of other services. And those services web is

32:13

out from there. So in loading a website,

32:15

a thousand services may have some sort of

32:18

new ID for you that is synced with

32:20

everyone else's. So it's basically like a table.

32:22

It's like a ledger entry. It's like my

32:25

idea is this. But this company's ID for

32:27

you is this, and this company's ID for

32:29

you is this. And all of this is

32:31

just some version of an identifier for you.

32:34

That is the privacy and basis of aspect

32:36

of cookies. Now the reason all that's necessary

32:38

is because there's not a singular publisher here.

32:40

It's many publishers. It's the web. By definition,

32:43

it's a web of publishers. Some of them

32:45

are big enough to sell their own inventory,

32:47

but most aren't. They have essentially a broker,

32:49

sell inventory for them. And then advertisers want

32:52

to go buy that inventory. Well, everyone in

32:54

this ecosystem that participates has to have some

32:56

way to understand who is who, when they're

32:58

on website A or website B or website

33:01

Z. And so Apple pushed back on this

33:03

first with ITP, this launch of 2017 and

33:05

safari. Now the problem is when they did

33:08

that, Facebook has a cookie. It's their pixel.

33:10

They want their cookie to work. Apple blocked

33:12

third party cookies wholesale. So what Facebook figured

33:14

out was, okay, well, let's find a way

33:17

to have our cookie, not be third party,

33:19

but first party. But a first party, well,

33:21

it's my website, I can do whatever I

33:23

want. And also, if you're logging in, I

33:26

need to drop a cookie. I need to

33:28

be able to know that when you come

33:30

back that you're logged in. You can't block

33:32

that. And so what Apple did was they

33:35

said, okay, we see what's happening here. which

33:37

still loads this third party information, but it's

33:39

the JavaScript is native in the page in

33:41

the header. And so that's considered to be

33:44

a first party cookie. And so when Apple

33:46

did was, okay, like when there's first party

33:48

cookies that are. invoked by known ad platforms,

33:50

we limit the lifetime of those to 24

33:53

hours. So if you have your login cookie,

33:55

that's fine. But if it's this cookie that

33:57

we believe is actually a third party cookie,

34:00

but it's dressed up as a first party

34:02

cookie, then we're going to limit the viability

34:04

of that to just 24 hours. And so

34:06

then all the big platforms switched to what's

34:09

known as a cappy, convergence API. So that's

34:11

a server side end point. And so what

34:13

they said was look, we could play this

34:15

game of cat and mouse forever. Instead of

34:18

using a cookie to send this data back

34:20

to us, what the cookie did was it

34:22

just had the Facebook ID. And so whenever

34:24

that user was there, they had the Facebook

34:27

ID and when they did stuff, they landed

34:29

on the page, the pixel got invoked, which

34:31

is a file on Facebook's web server that

34:33

had the Facebook web server that had the

34:36

Facebook ID that had the URL, that had

34:38

the Facebook web server that had the Facebook

34:40

ID. So they had the Facebook web server,

34:43

that had the Facebook's first party. Let's see,

34:45

look, I mean, why don't we just do

34:47

this? Let's create a server endpoint. And when

34:49

someone buys something, just send a ping us

34:52

from your server to our server, just do

34:54

totally server side. So we're going to do

34:56

all this outside the browser. So they have

34:58

the CAPI. So first party cookie integration, it

35:01

required you to go and copy paste in

35:03

JavaScript. That's some work. It's not that much.

35:05

But you actually had to do it in

35:07

the advertiser process with Facebook. CAP. And the

35:10

big ones. Now, the issue with the open

35:12

web and cookie deprecation is, what Google said

35:14

is, we're going to deprecate third party cookies.

35:16

Okay, well, why doesn't everyone just move to

35:19

a first party to cookie? Because how are

35:21

you going to go to a first party

35:23

cookie? If I'm a website operator, how do

35:26

I know which of these services that are

35:28

all being linked through cookie sinking? It needs

35:30

to be invoked. I have no idea. I

35:32

have no relationship with them. I have. into

35:35

my website. So that just becomes non-viable. Just

35:37

to circle back to a question you asked

35:39

before. But the idea of phase stuff was

35:41

Apple under any regulatory pressure to do this?

35:44

No. But was Google under any regulatory pressure

35:46

to depregate cookies? No. Maybe competitive, because Safari

35:48

did it, maybe they should do it. But

35:50

what happened was, once they said they wanted

35:53

to do it, the competition and markets authority

35:55

in the UK stepped in. So no, wait

35:57

a second. If you do that, it's going

35:59

to unfairly benefit your advertising properties. Google has

36:02

the biggest DSP, which is the programmatic. architecture

36:04

for serving ads on the web. It's called

36:06

DV 360. But that business is a sustained

36:08

decline for Google. It's called the network business.

36:11

You look at their earnings every quarter, it's

36:13

down. It's been down for nine consecutive quarters,

36:15

the revenue. What happens if selling ads programmatically

36:18

on the web becomes non-viable? Well, the advertisers

36:20

buying that inventory, they have to go somewhere

36:22

else. Where are they going to go? YouTube,

36:24

search, which don't rely on third-party cookies at

36:27

all. YouTube's in the browser, but you're logged

36:29

in. searches in the browser, but you're logged

36:31

in generally in any way, a lot of

36:33

that's keyword-based. So who wins if cookies get

36:36

deprecated? Well, Google takes a big bite maybe

36:38

out of its network business, but that network

36:40

business has tiny margins. It's basically just a

36:42

broker, which businesses have massive margins, is well,

36:45

that owned and operated inventories, but Facebook margins

36:47

are enormous. So you don't even need to

36:49

retain that much of that top line revenue

36:51

to retain the net revenue. Wow, do you

36:54

think the odds of this happening are likely?

36:56

It's inevitable. So what happened was the CMA

36:58

stepped in and you might say, well, okay,

37:01

a regulator stepping in, they probably want to

37:03

ensure that this is all done in a

37:05

privacy safe way. Well, no, it had nothing

37:07

to do with privacy. Did it? And wait

37:10

a second. Google, you're just going to self-preference

37:12

and kill the open web. You've got to

37:14

manage this process with us. You have to

37:16

do regular reporting with us. You have to

37:19

do regular reporting with us. You have to

37:21

share data reporting with us. You have to

37:23

share data with us. And I think part

37:25

of that was informed by the experience from

37:28

AGT because the CMA did investigate Apple with

37:30

AGT. Ultimately, they They couldn't do anything about

37:32

it. But they did. If you read the

37:34

reports, I'm quoted in them, I talked to

37:37

them, I missed, but I think they saw

37:39

what happened is like, hey, wait a second.

37:41

But a big platform, a trillion dollar company

37:44

says they want to do something just for

37:46

the benefit of consumer privacy. They're so gracious

37:48

with the way they invest all these resources

37:50

into these initiatives that don't benefit them at

37:53

all. It's really just to protect the protected

37:55

consumer. I think we should be suspicious. This

37:57

is really going to self-preference their own services

37:59

and give them a competitive advantage in some

38:02

way. So the CMA stepped in, it's okay,

38:04

you got to manage this process with us,

38:06

you will slow this down, we're going to

38:08

let you know if it's okay. And then

38:11

what happened was, they went back and forth,

38:13

and then what happened was, they went back

38:15

and forth to CMA kept saying, look, you've

38:17

made progress, but we can't let you know

38:20

if it's OK, and let you know if

38:22

it, if it's okay. And then, they went

38:24

back, they went back, they went back, they

38:26

went back, they went back, they went back,

38:29

they went back, they went back, they went

38:31

back, they went back, they went back, they

38:33

went back, they went back, they went back,

38:36

they went back, they went back, they went

38:38

back, they went back, they went back, they

38:40

went back, they went back, they went back,

38:42

they went back, they went back, they went

38:45

back, they We're going to let them say

38:47

whether they want cookies or not. Well, it

38:49

sounds like ATT to me. And why would

38:51

they be incentivized to not implement it like

38:54

ATT? Now, this may get blocked, this change,

38:56

but I think it's a little bit hard

38:58

to do that when the regulators have been

39:00

saying, look, consent is the goal, is the

39:03

goal, is standard, but I think it's a

39:05

little bit hard to do that when the

39:07

regulators have been saying, look, consent is the

39:09

choice, you let we're giving, we're giving you

39:12

the choice, what, what's the choice, what's the

39:14

choice, what's the choice, what's, what's, what's the,

39:16

what's, what's the, what's the, what's, what's, what's,

39:19

what's, what's, what's, what's, what's, what's, what's, what's,

39:21

what's, what's, what's, what's, what's, what's, what's, what's,

39:23

what's, what's, what's But we're just going to

39:25

ask for consent. Now, we're going to control

39:28

what the wording is when we ask for

39:30

consent, and we might put our thumb in

39:32

the scale in that way, but ultimately, we're

39:34

just giving users more choice. Isn't that a

39:37

good thing? And so that's what Google said

39:39

recently. That's the model that they're going to

39:41

use, and that'll be what regulates the possibility

39:43

of cookies in the Chrome browser, which is

39:46

by far the largest browser by usage. So

39:48

it's similar to what happened with ATAs. When

39:50

people, I feel like, debate the web and

39:52

business models with one one million through your

39:55

knowledge, I feel like there's this mental game

39:57

people play of trying to build businesses without

39:59

ads. And I know you've written everything as ad

40:02

network, so I don't expect you to say

40:04

it's okay. But I guess on one side, is

40:06

it ever possible to build an economic

40:08

model on the internet that isn't funded

40:10

by ads, whether that's paywalls or other

40:12

things, or that'll just never get to

40:14

a scale that's relevant? And then why

40:17

is it the best form of digital

40:19

commerce? That's a great question. It

40:21

gets pretty philosophical. If you want to

40:23

take this sort of quantitative approach to

40:25

answering it, I mean there. How much

40:27

friction is acceptable if we want to

40:29

get the outcome that we want? The

40:31

problem is going from zero to something

40:34

presents a lot of friction. And

40:36

then from there, there's just gradation. So

40:38

the second you go from free to

40:40

paid to one cent, free to one cent for

40:42

life, if you have to get someone to

40:44

pull the credit card out, the conversion

40:46

rate, which is just the usage of

40:48

the product, drops dramatically. The equilibrium

40:51

there is just a function of that. That's

40:53

why the premium model is so compelling.

40:55

And when you're talking about internet

40:58

scale, premium oftentimes makes a lot more

41:00

sense than a paid model. It's ironic

41:02

that I say this because I have

41:04

a subscription on my website. But I'm

41:07

going after a very nice audience and a

41:09

lot of them are willing to pay

41:11

a lot of money because they find

41:13

a great amount of value in this.

41:15

A lot of the subscribers are professional

41:17

investors. And this is valuable information to

41:19

them, so they're willing to pay. If

41:21

I have a recipe website, you know.

41:23

Well, I've got to monetize via ads. So

41:26

I think when you talk about the

41:28

internet, the whole point of the internet

41:30

is to reach humanity scale in terms

41:32

of distribution. And if you want to

41:34

do that, the distribution can be universally

41:36

correlated with friction, modernization friction. And ads

41:38

have the lowest modernization friction. There's just

41:40

a function of what's the natural TAM.

41:42

And if it's very large, well then

41:44

the idea of monetizing that with the

41:46

subscription. probably is less compelling given the

41:48

conversion dynamics than just ads. And now those economics

41:50

get altered by the degree to which you can

41:52

make those ads relevant and have people want to

41:54

pay a lot of money for them. Let's go

41:57

back to that recipe website for example. Let's say

41:59

have no targeting. is that inventory worth?

42:01

Almost nothing. If it's just junk inventory, it

42:03

is recipe stuff, there's no commercial intent there,

42:05

baking something, I would maybe target my baking

42:07

products if I ran a baking product company

42:10

to them. But I have no knowledge on

42:12

the zero context website that the inventory is

42:14

worthless. But what if, you know, everything about

42:16

those users, you knew everything about those users,

42:19

you've somehow had their entire consumption history for

42:21

life. It just so happened they were on

42:23

that website in that moment. Well, you can

42:26

be potentially very valuable to you. And now

42:28

the question is, how valuable is inventory? Well,

42:30

it depends on how many people know that

42:32

about them. Well, everyone knows that about them.

42:35

Well, then they're going to compete. If only

42:37

you know that about them, then you could

42:39

probably get them for very cheap and take

42:41

a big delta on what you paid versus

42:44

what they're worth. But if everyone has that

42:46

information, well, then you've got to compete. And

42:48

you compete in digital advertising through the auction.

42:50

Inventory sold in an in an auction process

42:53

in an auction process. And so you. And

42:55

so you compete in an auction process. And

42:57

so you compete in an auction. And so

42:59

you've. or this information is available to at

43:02

least one person. It's that this information is

43:04

available to many of the participants in the

43:06

auction that makes the inventory valuable. But if

43:09

you had total transparency until some person's commercial

43:11

history, well then it might not be worth

43:13

anything because you know that they're not worth

43:15

anything because you know that they're not worth

43:18

anything because you know that they're not in

43:20

a position to buy anything, or it could

43:22

be very very valuable. This is exactly the

43:24

perfect user free to target. And then, like,

43:27

if you think about value proposition from the

43:29

users, without having to show a paywall or

43:31

anything. The bigger I get, the more money

43:33

I make, sort of conversion is 100% because

43:36

everyone who reads my website monetizes. But if

43:38

the inventory is junk or like the ads

43:40

that people are buying are junk or irritating

43:42

or they're totally irritating or they're totally irrelevant,

43:45

it's actually going to cost friction on engagement

43:47

because people are saying, I don't want to

43:49

spend time on that website because the ads

43:52

are annoying or they're irrelevant. But what if

43:54

the ads were very relevant, they're very relevant

43:56

because all the advertisers had all the information

43:58

about the information about the person? Well, hey,

44:01

that actually might increase engagement. There have been

44:03

studies that have shown that relevant ads actually

44:05

increased engagement. in the source product. Think about

44:07

this from the user's perspective. Well, hey, would

44:10

I rather, I want to go get some

44:12

recipes. There's three options. I pay to access

44:14

that content, well, that's not ideal. I go

44:16

to the content and all, that's not ideal.

44:19

I go to the content and all the

44:21

content and all the ads are horribly distracting

44:23

and relevant to me. Or all the ads

44:25

are very relevant for products that I might

44:28

be interested in buying, whether or not. I've

44:30

called Targeted Effortize Public Goods. Everyone benefits from

44:32

it. If you think about the economics of

44:34

this, obviously everyone's better off if the ads

44:37

can be targeted. Yeah, I totally agree. I

44:39

feel like when it goes well, it always

44:41

felt like when it goes well, it always

44:44

felt like when it goes well, it always

44:46

felt magical, even though when it first happened,

44:48

people thought, oh my God, they know everything

44:50

about me, is Amazon listening to me, how

44:53

does it know that I wanted a pair

44:55

of shoes when I was, when I was

44:57

going on that. First of all, I mean,

44:59

I think people should be informed about how

45:02

their data is collected and used to not

45:04

do that. And for people to know that

45:06

it's happening, it's corrosive for trust. I think

45:08

actually in the big picture scheme of things,

45:11

it erodes the credibility of institutions. And broadly,

45:13

not just the idea of surfing the web

45:15

while I'm being surveilled, I think broadly it

45:17

makes people anxious and nervous. It makes them

45:20

less likely to express themselves freely. And in

45:22

that way, it's erodes the credibility and the

45:24

power of institutions. Because all these things are

45:27

tied together now. I mean, I get my

45:29

information about everything from the web. It's an

45:31

indispensable part of life. So I do think

45:33

it's important for people to understand that their

45:36

data is being collected, to have some sort

45:38

of say, to have some sort of agency

45:40

in the collection of that data. And people

45:42

use these products with wild misconceptions about what

45:45

data is collected and how it's used. Nonetheless.

45:47

and they still use these products with wild

45:49

misconceptions about how this data is collected and

45:51

what data is collected. Perfect example is there's

45:54

this persistent mythology that your phone is always

45:56

actively listening to everything. you do and that

45:58

data is being collected in someone and used

46:00

for ads targeting. Zuck was asked about this

46:03

in congressional testimony. Because that's point blank, does

46:05

Facebook's microphone listen to what I'm doing and

46:07

target ads on the basis of what it

46:10

hears me doing? He said no. And yet

46:12

people still believe that to be the case,

46:14

A, and B, still continue to use Facebook.

46:16

If you saw Facebook was surreptitiously listening to

46:19

everything you did and using that for ads,

46:21

I want you to delete the app immediately.

46:23

I want you to delete the app to

46:25

delete the app immediately. which is a misconception.

46:28

It's patently false. So I don't think Zuck

46:30

would lie under oath to Congress. He just

46:32

wouldn't do that. I hope not. So if

46:34

you believe that to be true, wouldn't you

46:37

just delete the app from your phone immediately?

46:39

So people hold this misconception and yet they

46:41

still continue to use the apps and the

46:43

products. Yeah, I think the average person I

46:46

talked to has no idea all the stuff

46:48

that you've shared with us today, which is

46:50

fascinating. And I think this is the dystopian

46:53

thing of when technology works so well. There's

46:55

a cross between magic and conspiracy that happens

46:57

where I'm talking to you today and I

46:59

mentioned that I like shoes. And I walk

47:02

home today and just because of all the

47:04

things that happened, I see my Amazon Alexa

47:06

in kitchen pops up and it offers me

47:08

shoes. And people have this thing where they're

47:11

like, how the hell didn't know I was

47:13

looking for shoes? because they're not connecting a

47:15

week ago or whatever, the cookie sink, whatever

47:17

magic the internet did deliver that point of

47:20

information to them, the magic that could be

47:22

good if it wasn't being used on bad

47:24

information invading my privacy, I think freaks people

47:26

out at times and there's no way they

47:29

have the ability to have the subject matter

47:31

expertise to figure out how the hell that

47:33

happened at that moment. So there's a couple

47:36

of things. One explanation is just recency bias.

47:38

You get shoes ads all the time. Yeah.

47:40

The only reason it caught your eye was

47:42

because we just talked about shoes. So that's

47:45

one explanation. Another explanation is you like shoes

47:47

and before seeing that ad, you were searching

47:49

for shoes, you were on the same IP

47:51

address and you were targeted in that way.

47:54

That could happen. But I think there's something

47:56

more astonishing happening and astonishing in a good

47:58

way. And it's actually that. you'd leave of

48:00

exhaust trail of interest everywhere you go

48:02

on the internet. And that could be

48:04

on Facebook and that could be wherever, Google

48:06

Search, whatever. And so the knowledge of you

48:09

being in the market for shoes at this

48:11

moment in time was not the impetus for

48:13

that ad to be shown. It was that

48:15

you fit the profile of someone who would

48:17

be in the market for shoes at this

48:19

moment in time. And that's why you

48:21

were targeted with the shoes ad. And

48:23

those are two different things. One is

48:25

more of a probability statement. One is

48:27

very much a discrete observation. And so

48:29

if the probability statement is actually what

48:31

surface that ad, that is an amazing

48:34

feat of technology. Exactly. And that's

48:36

why it's magical that it would freak

48:38

someone out because they don't have a

48:40

recency bias. They don't remember when they

48:42

did it. But because the probabilities worked

48:44

and all the agency and targeting hit

48:46

it, at that moment, it blows your mind.

48:48

Exactly. If the machine learning is so good

48:50

that it could take all of these disparate signals

48:53

and parse them together. into the understanding

48:55

that you fit the profile of someone

48:57

who would be in the market for

48:59

shoes at this moment time, that is

49:01

a technological feat that we should be

49:03

celebrating. But you can imagine that's what

49:06

freaks people out. Was this thing listening

49:08

to me? Of course. Because I haven't searched

49:10

on it. I think that's what sometimes causes

49:12

the when it works so well, which actually

49:14

I wanted to go play one more thought

49:17

experiment with you on this point. You had mentioned,

49:19

could you imagine if you knew all my data?

49:21

my entire Amazon search history, or my

49:23

entire Facebook. We'll put Facebook as we

49:26

can talk about those three. And you

49:28

are going to sell ads to me. Of

49:30

those three, what is the most valuable

49:32

data set to build anything on top

49:34

of? Amazon for sure. I mean, the

49:36

social media stuff would be not very

49:39

valuable at all. And search would be

49:41

mixed. You might be searching for stuff to

49:43

buy. You might be searching for a recipe

49:45

or whatever. It'd be the Amazon data

49:47

set for sure, I think. It seemed obvious,

49:49

but I was wondering if you were going

49:52

to tell me that because the richness of

49:54

all these other things of my networks

49:56

or of my search that these models

49:58

get further and farther afield, start saying,

50:00

oh, you would want this thing. I was

50:02

just wondering if that was anyway more valuable

50:05

than just, this is what they've been buying,

50:07

they're buying flashlights, they need batteries, this is

50:09

an obvious sell. I think that's a misconception.

50:11

That was one of the things that people

50:14

got wrong about ATT at the start. It

50:16

was something that I've spent a lot of

50:18

time trying to disabuse people of this idea

50:20

that your social media interactions are from an

50:23

advertising standpoint irrelevant. The fact you like some

50:25

picture or whatever, you like some picture or

50:27

whatever. Maybe with a lot more of them,

50:29

but you'd still need the feedback loop. You'd

50:32

still need, okay, this guy likes these images

50:34

and he also happens to buy shoes a

50:36

lot or whatever, buy pants or whatever, like

50:39

this brand. But then if I know that

50:41

you like the brand anyway, I guess I

50:43

don't need the other information, I don't need

50:45

to build those associations. Now I'm sure that

50:48

social media platforms in the absence of all

50:50

the day they had before, I have tried

50:52

to parse more signal from native interactions, but

50:54

native interactions, but my sense is they aren't

50:57

very valuable. They aren't very valuable. They aren't

50:59

very valuable. They aren't very valuable. And that's

51:01

why Amazon has done so well. I mean,

51:03

I called it an advertising behemoth. I mean,

51:06

they grew 19% year over year this quarter,

51:08

Q3. At that scale. And the idea here

51:10

is that everything is commercial in 10. You're

51:12

not on Amazon to discover recipes or look

51:15

up the local park. You're there to buy.

51:17

Yeah. It makes all sense. This has been

51:19

a lot of fun. The last thing I

51:21

wanted to talk about, because you mentioned a

51:24

little bit with the machine learning that Facebook's

51:26

doing, how has this push of AI changed,

51:28

if at all, any of these business models

51:31

for better or for worse? Oh, it's been

51:33

a massive accelerant. I think it's absolutely underrated.

51:35

I think across all these stocks, any of

51:37

the companies that can build this infrastructure, they're

51:40

undervalued. Because it's going to be transformative, it

51:42

has been transformative already, and they're just getting

51:44

started. And I think if you look at

51:46

the most obvious of these levers of these

51:49

levers, I talked about the on the generative

51:51

AI for creative, that will be absolutely transformative.

51:53

So right now, the way ad creative production

51:55

works, I make an ad and I test

51:58

and see if it works, and if it

52:00

works, I'll make some variations of it, change

52:02

the background color or if it's a person.

52:04

maybe I'll make them bigger in the ad

52:07

or smaller in the ad, do these cosmetic

52:09

changes and then I'll test those. What Facebook's

52:11

done already is automate that. You upload one

52:13

ad, we'll produce 10 variations of that and

52:16

that's already added impact. And then Google is

52:18

doing the same, credit, memorable AI. This is

52:20

an obvious thing. This is an obvious thing.

52:22

This is an obvious thing to do. That's

52:25

a couple of ads, 10. And so the

52:27

best ad for the best audience segment will

52:29

be shown. What happens if you can shrink

52:32

that group? the right ad to the right

52:34

person always with the highest probability of engagement

52:36

which be a click or ultimate like a

52:38

downstream purchase. And what if I could do

52:41

that by taking into account all of the

52:43

historical data that I have about this person

52:45

as well as all of the data that

52:47

is relevant to the session that they are

52:50

in at that moment. I can take into

52:52

account the three posts above that they just

52:54

scroll down and looked at. All of that

52:56

can be an input into the ad creation.

52:59

so that whatever ad is shown to you

53:01

at that game point in time has been

53:03

produced specifically in real time for that moment

53:05

and deemed to have the highest probability of

53:08

conversion. I think that is the end point

53:10

of all this and if you think about

53:12

that it's going to be just a dramatically

53:14

different environment for digital advertising. That's going to

53:17

be wild. Eric this has been a lot

53:19

of fun. I've learned a lot today. Thank

53:21

you for doing it. Eric. Thank you for

53:24

having me. There you will find every episode

53:26

of this podcast along with transcripts, our weekly

53:28

newsletter, and resources to continue your learning.