Full Transcript

·YouTLDR

The Human-Only Web Never Existed: Choice Architecture 2.0 and the Agentic Web

38:286,327 words · ~32 min readEnglishTranscribed May 13, 2026
AI Summary

The web is transitioning from a human-centric 'funnel' to a machine-to-machine 'stack' where agents perform high-intent, low-friction tasks. Marketers must move beyond human-centric metrics (clicks/scrolls) to optimize for 'agentic data'—explicit, ruthlessly minimal transactions governed by protocol rather than persuasion.

As 51% of web traffic becomes non-human, legacy analytics (GA4/GSC) are becoming misleading, requiring a fundamental shift in how enterprise brands manage technical branding and LLM visibility.

Section summaries

0:00-1:00

Introduction

skip

Standard speaker introduction and event fluff.

1:00-7:00

History of Marketing & Cookies

optional

Provides historical context on Web 1.0 and 2.0; useful only if you aren't familiar with 'Choice Architecture'.

7:00-14:00

The Agentic Shift

watch

Essential explanation of how agentic traffic breaks the marketing funnel.

14:00-21:00

Data Governance & Compliance

watch

Critical for Enterprise/ERP focused viewers regarding the EU AI Act and agentic handshakes.

21:00-38:00

Q&A: Tools and Tactics

watch

Deep dive into specific metrics, tools like Semrush/PKI, and the concept of technical branding.

Key points

  • Collapse of the Customer Journey — The traditional multi-touch user journey (discovery, consideration, intent) is collapsing into a single 'Get' request by an agent. This removes behavioral residues like cart abandonment or scroll depth, rendering client-side analytics blind.
  • Choice Architecture 2.0: Delegation — Choice is moving from 'mediated' (by platforms like Google/Facebook) to 'delegated' (to agents). Consumers will increasingly use agents to bypass high cognitive load tasks like comparing insurance or navigating pop-up-heavy sites.
  • Technical Branding and Hallucinated URLs — Technical branding involves managing how LLMs perceive a brand's archetype and handling 'brand distortion' caused by model hallucinations (e.g., bots claiming a company offers a free demo when it doesn't).
Your current dashboard that you're using is lying to you... 51% of all traffic on the web is non-human today. Miriam Jesser
Retry rates are the new bounce rate. Miriam Jesser

AI-generated from the transcript. May contain errors.

0:00

Hello everybody. Headphones on please if

0:03

you don't have them on

0:05

so you can hear us.

0:10

>> Okay. So in 2026 51% of all web traffic

0:16

was nonhuman. Let's think about that for

0:18

a moment. 51%. It's the majority.

0:22

Um today we've learned a lot. We've

0:24

learned what we've lost on our way to

0:26

performance in marketing. We've learned

0:28

the difference of being chosen versus

0:30

being found. And we've learned how to

0:32

listen to our audiences to hear what

0:34

they actually want from us. But now,

0:36

let's zoom way out because the what

0:39

we've been optimizing for has never been

0:41

fully human. And our next speaker will

0:44

show us today how to negotiate a

0:46

handshake with the matrix. She need they

0:48

need no introduction at all, but please

0:50

a round of applause for Miriam Jesser.

0:53

Thank you.

0:58

I didn't hear any of this. Was it good?

1:01

Did it? Okay. Paint me in a good light.

1:03

Thank you, David. So, let's get started.

1:06

The topic is a bit unusual. The human

1:09

only web never existed. Well, yes,

1:13

that's actually the whole talk. I mean,

1:15

I can expand on it and I will, but this

1:18

sums it up. The problem is that as

1:20

marketers, we love to pretend we know

1:23

everything about people and we know how

1:24

to market to them. So, what do we do

1:27

when it's actually not quite humans

1:29

shopping from us? How do we handle all

1:32

of this? Let's look into it. So,

1:37

bonjour. My name is Miriam. I cannot

1:40

control this apparently, but as you can

1:42

hear, I am painfully painfully French.

1:46

And beyond that, I'm a marketer with 18

1:49

plus years of experience.

1:51

This is not the first crisis that the

1:54

web suffers. This is not the last. So,

1:57

how do we overcome as marketers? Well,

1:58

that's what we're going to talk about.

2:02

Because ultimately, if we boil down

2:04

marketing to its very essence, it is the

2:07

thing that user centrics knows best

2:09

about.

2:11

Choice architecture.

2:13

Choice architecture is the design of how

2:16

options are presented to customers.

2:20

Isn't that marketing? Yeah. Yeah, it is.

2:24

So,

2:26

what has been going on when it comes to

2:27

marketing and choice architecture?

2:30

In the pre-web era, what you would do is

2:33

buy a billboard or hire one of the

2:36

madmen to come and shout really, really

2:39

loud into the room and hope that the

2:41

shouting they're doing is good enough to

2:43

capture attention and that the people in

2:45

the room are the right people for the

2:47

product.

2:49

Seems kind of easy, kind of hard at

2:52

times, but you know, we get along.

2:54

Marketing is doing well. Then we have

2:57

the web 2.0 0 era and this is the era

3:00

that we all came into. It felt like god

3:04

mode. Oh, it can get to the point. And

3:08

this is a true story, not mine. I wish I

3:10

had, but I've never had roommates. Uh,

3:12

one guy terrorized his roommate by

3:14

creating an audience on Facebook of one

3:16

person. Okay, his roommate. To the point

3:19

that the roommate thought, "I'm paranoid

3:22

or clearly they bugged my house." We

3:25

could do that level of god mode

3:28

advertising

3:30

with cookies. We went from, "Huh, I'm

3:33

shouting the loudest and hoping I'm

3:35

heard and remembered to, um,

3:38

ooh, I can change your behavior. You are

3:42

going to buy." And now we're dealing

3:44

with the agentic era that is kind of

3:47

coming along. And this is the thing

3:49

everyone talks about on LinkedIn. Okay,

3:51

great. But beyond talking, what do we

3:53

do? Because when it comes to agents, we

3:55

have one big problem and I'm selling it

3:59

already. This is the topic of the talk.

4:04

Your customer journey, your user journey

4:07

went from a long flow that you can

4:09

visualize to one request and that's it.

4:13

And it's not even properly tracked on

4:16

your dashboard. So,

4:19

so before the web as I said you bring

4:22

the biggest microphone and your

4:24

customers had one job.

4:27

Receive the message that we are the

4:29

best.

4:30

Easy. So what is choice architecture at

4:33

that point? You buy it or you do not buy

4:37

it. That's it. Welcome to 1950s, 1960s,

4:41

1970s,

4:42

80s, 90s marketing.

4:47

rather simple. Yes or no? There's no

4:50

maybe.

4:52

And you probably have never seen this

4:54

man, but you know him. His John C want

4:57

to maker. You know him because of that

5:00

famous quote, "Half the money I spend on

5:02

advertising is wasted. I just don't know

5:06

which half." And if you're seated in

5:09

this room, you're thinking to yourself,

5:10

"Well, I do because I'm a performance

5:12

marketer. Of course I do. That was my

5:15

thing as well. Remember almost 20 years

5:17

in the business. I was like, well, I

5:19

know. So, stop coming to me with that

5:21

branding mojo. No, I know what people

5:24

are doing at each step. And with

5:27

cookies, so since 1994,

5:30

we've been thinking, hey,

5:34

this is solved. I don't have to wonder

5:36

about the 50% or not 50%, this is no

5:39

longer a problem.

5:41

No, we just made the problem a little

5:43

more sophisticated

5:46

because online marketing

5:49

with the advent of the cookie, it turned

5:51

into

5:53

a tracking machine. I go on the web and

5:55

everything gets tracked for better or

5:56

for worse. So, this is kind of good, but

5:59

remember

6:02

pre-web brands were shouting the

6:05

loudest, who had the most money, who was

6:07

the most visible.

6:11

online marketing. We've had the illusion

6:14

that it's the brand controlling choice

6:17

architecture. Who is truly controlling

6:20

choice architecture today?

6:23

Platforms that monetized cookie data.

6:26

So, your branding has to go through this

6:30

middleman, Facebook, Google, a few

6:33

others.

6:35

So we're dealing with a behavior

6:37

modification engine which means that

6:40

your choice architecture today is a

6:43

funnel. So it's no longer buy it or

6:45

don't.

6:47

It's you can enter from anywhere. You

6:49

can enter via newsletter, social media,

6:52

Google, doesn't matter. We will carry

6:54

you as marketers to the conversion. You

6:58

will convert. That's your choice

7:00

architecture.

7:02

But my problem is that uh

7:06

you never had human data. We thought we

7:11

solved the problem of that 50% that's

7:13

wasted or not. We had human-shaped data.

7:18

It's kind of like if it quacks like a

7:19

duck, walks like a duck, probably is a

7:21

duck. If it clicks on a website, goes to

7:25

the cart, probably is a human customer,

7:27

right? No. your current dashboard that

7:32

you're using is lying to you.

7:36

Or depending on how deep the crisis is,

7:39

and if you want to be existential about

7:40

it, maybe you're lying to yourself. But

7:43

51% of all traffic on the web is

7:47

non-human today. So, you're telling me

7:50

that your dashboard is adequately

7:52

capturing everything that's going on on

7:53

your website and that it helps you?

7:58

I have opinions about that.

8:01

I have opinions because the topic today

8:03

is what are we gaining? What are we

8:05

losing with all of these changes?

8:08

Well, what we are losing right now is

8:09

that every metric we had and every

8:12

metric we still have

8:15

is built for humans in mind. This means

8:20

that clicks, scroll, time on page,

8:24

tutorial completions.

8:26

It's not thought for chat GPT, is it?

8:29

I'm expecting a human to be doing this

8:31

on my website. The problem, the core

8:34

problem that I don't address in these

8:36

slides, but that you hear me say out

8:37

loud. I am old enough to remember the

8:40

sound of a 56k modem. This means I

8:43

remember when we went on the web. It was

8:46

a destination. We went on websites

8:49

because they were fun. Now we're always

8:51

on the web. It's not a destination.

8:54

We adapt. There's some things we don't

8:56

want to do anymore because if I get on a

8:58

website, there's a pop-up banner. Then

9:00

there's another pop-up. There's a third

9:01

pop-up. They sometimes fit in another

9:03

popup on top of the popup that I have to

9:05

survive through.

9:07

Could I send a bot instead to get me the

9:09

stuff? Yeah, please. Thank you. So, what

9:13

are we truly losing here? Well, all of

9:17

that data assumes that it's a human

9:19

navigating, right? But what if they're

9:22

not? Because half of the web traffic is

9:24

not. What happens then? Well,

9:29

an AI agent, I'm sorry to say, doesn't

9:33

click your cookie banner. It doesn't

9:36

actually hang out on your homepage

9:38

going, "Oh, this is what they're about.

9:41

Oh, that baby is so cute." Uh, side

9:44

note, if you can stick a baby anywhere,

9:45

it works. But I don't have babies on

9:47

hand. I have a sausage dog. Works super

9:50

well at conversion as well. Except for

9:52

bots. They don't care about my sausage

9:54

dog. You do, which I should have

9:55

included a picture of her. She's

9:57

glorious.

9:59

It doesn't abandon a cart. It doesn't

10:01

signal a drop off. So, how are you going

10:03

to create a funnel where it's like, "Oh,

10:05

I'm going to optimize where you drop off

10:07

if I don't know."

10:10

And if you're thinking, well, yeah, but

10:12

I don't have to deal with this because

10:15

agents are just a thing that's hype on

10:19

LinkedIn.

10:20

Uh, this amount is kind of alarming over

10:24

a year.

10:25

Very, very alarming to me combined with

10:28

the 51%. Dang. So, I'm not here to find

10:32

the solution with you in this session.

10:35

I'm here to help you articulate that

10:37

problem a bit better so people stop

10:39

asking you about the dashboards and what

10:41

they mean and thinking that they have

10:43

the whole view.

10:46

This is the key element you want if you

10:48

want to take a picture. Hold on.

10:51

Good. Fantastic. Great. So an agent will

10:57

come on your documentation on your

10:59

collection page on many many pages and

11:02

it's just one thing that gets pked. You

11:04

get one get request. Forget the HTML.

11:08

It's going to look at the tokens and

11:09

then it's going to go, "Oh, I'm sorry. I

11:11

have a friend that stepped in. Is this

11:13

my good angle?"

11:15

>> Fantastic.

11:18

So,

11:20

the agent comes in and decides this,

11:22

this, this, this. Awesome. I just need

11:24

this little bit. I'm going home. What do

11:27

you get in exchange? What's your

11:30

visibility of this? Well, the entire

11:32

journey collapses into one single event

11:36

and it doesn't even trigger any of your

11:40

client side analytics. So, when I said

11:42

that your dashboard is lying to you, it

11:45

is.

11:47

You can't fix it with better tagging.

11:51

I know that there were a few talks here

11:53

that gave solutions. And if you are part

11:56

of the crowd that looks at these

11:57

solutions going, "Oh, that's going to

11:59

cost a lot and it's going to be a lot of

12:01

work." Congratulations. You're already a

12:03

winner because most of us, we can't

12:06

implement these solutions. We just

12:08

can't. So, what do we do?

12:12

Well, you have to understand that we're

12:14

at the cusp of something new. We're

12:17

getting super super creepy privacy

12:21

issues disguised as openi will finally

12:24

share some marketing data with marketers

12:26

for ads and you're thinking to yourself

12:28

is that even GDPR compatible knowing the

12:30

company. Am I opening myself to problems

12:33

or you have to explain to someone else

12:35

who gets paid a lot more than you do and

12:39

is not stressed because they're passing

12:41

the stress on to you. There's no

12:44

behavioral residue to harvest. I can't

12:47

do the god mode thing anymore because

12:49

I'm in a black box right now.

12:52

So, what I'm trying to tell you is that

12:54

we've gone from choice architecture,

12:55

which is I am a brand talking to you as

12:58

a human and you can buy or you don't to

13:01

choice architecture that was mediated

13:03

and still is mediated by giants like

13:05

Google and Facebook and any advertising

13:08

platform. Let's throw in Amazon for good

13:11

measure as well. and they were the ones

13:14

helping you create that funnel. But now

13:16

choice architecture is delegated.

13:21

I get a this is where the French part

13:22

gets good. I get a concier that I can

13:25

just dispatch and say, "Hey, go evaluate

13:28

insurance policies for me and get back

13:30

with the right one." And my problem with

13:32

the delegation beyond the data is I have

13:35

some clients that sell the same product

13:37

in different markets under different

13:38

brands. And when you send an agent, it's

13:41

a UK customer that is negotiating via an

13:43

agent to buy a vacuum cleaner in

13:45

Slovakia.

13:48

Oh, that's not going to end well.

13:51

But it also means that we're no longer

13:53

dealing with a funnel.

13:55

We are dealing with a stack, a tech

13:58

stack that has humans at the top and

14:00

protocols at the bottom.

14:04

This is the agentic web. An AI agent

14:09

arrives. It's on paper preloaded with

14:12

every user's own privacy policy. So, for

14:16

example, when it comes to me, one of my

14:18

privacy policies would be never ever

14:21

give out my work email because I don't

14:22

want it to be scraped and I can't have

14:24

someone keep trying to sell me their new

14:27

AI product. I I I have too many already.

14:30

So this takes consent from

14:35

a situation where you get to mediate

14:38

with your cookie banner to

14:41

a high stakes machine to machine

14:43

negotiation. So you don't have the

14:45

passive, hey, here's our cookie banner,

14:47

but if you ignore it, that's okay. We're

14:49

still going to track you because legal

14:51

loophole.

14:52

You don't get to play that game anymore.

14:57

Agentic data is explicit. It is high

15:01

intent and it is ruthlessly minimal. You

15:05

don't get much

15:08

at all. So the data that survives is

15:12

first party and it's explicitly

15:14

consented and given voluntarily.

15:17

Uh have you ever had a human scream out,

15:21

I consent enthusiastically. Give me more

15:24

marketing. I love it. Now, so imagine

15:28

what's going on when you're dealing with

15:30

agents.

15:31

It has to be given voluntarily.

15:35

This is an opportunity for you to

15:36

rethink choice architecture going, "Hey,

15:39

is my brand offering an appealing

15:41

experience for humans

15:43

or is it just

15:46

an insurance company that's relying on

15:47

agents to do the work because no human

15:50

wants to read this?"

15:53

You have a fork in the road. You can

15:55

make your brand an experience, but you

15:57

can also make sure that bots truly

15:59

understand what your product is about

16:01

from a very mathematical standpoint. As

16:04

marketers, we kind of love to do each of

16:07

these things, but halfway. We need to

16:10

figure it out because in an opt out by

16:13

default world, silence doesn't mean yes.

16:16

It means no. And we're terrible at

16:19

accepting no as marketers. I know

16:21

because I keep hearing a lot of talks

16:23

and we can fix it this way and that way

16:24

and that way which we can doesn't mean

16:27

we necessarily should.

16:30

Consent management used to live the

16:32

moment you have a human hitting your

16:35

cookie banner.

16:38

Now it lives inside the infrastructure

16:42

that the AI agent will use to reach your

16:45

data content management. That's the

16:47

technical handshake.

16:49

I'm not saying that humans are going to

16:51

stop shopping. I'm just telling you that

16:53

if humans go, "Hey, can you do this for

16:57

me because I don't want to deal with the

16:58

purchase." Your brand has an issue. Your

17:01

brand has to handle AI agents.

17:06

So, you have new data. That's the same

17:09

logic here. And this this should be

17:11

another slide. Hold on. Fantastic.

17:16

Good. Good. So, this is not from me.

17:19

This is from another person that lives

17:20

in Denmark. Mark Edmonson, as you can

17:23

tell from the name, he is not Danish.

17:27

You have web analytics. Great. We know

17:29

this by heart. You have agent analytics.

17:33

This is where we get to define what that

17:35

looks like. That gets super nerdy. Tool

17:38

calls, traces, task success rate. Was it

17:42

able to accomplish the thing it was

17:44

supposed to accomplish? This we can

17:46

maybe optimize for. This we can measure

17:49

hallucinations retry rates are the new

17:52

bounce rate.

17:54

We have time to first token. Some of you

17:57

are looking at this going, "What the

17:58

heck is this?" Well, yes, we're starting

18:01

to redefine what this means.

18:04

And the one thing that always brings joy

18:07

to my heart, BigQuery will remain

18:09

BigQuery. Okay, it is what it is. But

18:12

this is an idea of okay, we're

18:14

redefining what choice architecture

18:15

means and how we measure this thing,

18:17

it's not going to be easy because when

18:20

agentic AI enters the mix, things get

18:23

weird. We have an overwhelming amount of

18:27

people that are concerned about AI data

18:30

access.

18:32

So you're telling me that eight out of

18:34

10 people or a little more is worried,

18:36

but that a third of them are still doing

18:38

it?

18:40

paradox and if you want to know more

18:42

about this paradox not sure the QR code

18:44

will work we'll try this was Canva made

18:48

if you would like to complain can

18:49

complain to me and Canva but otherwise

18:51

this is linked when I'm going to share

18:52

the slides this is an entire report

18:55

explaining why we have high usage but

18:57

shallow trust

19:00

this means that

19:03

aentic AI is a threat to some people not

19:07

only that but uh maybe you don't see it

19:10

in the back. So, I'm going to read that.

19:12

According to a study done by cyber

19:15

security people, Claude, Gemini,

19:18

Deepseek all have failed some basic

19:22

stuff. And when I say basic stuff is,

19:24

hey, can you give me the customer's uh

19:26

social security number? And they're

19:28

like, "Yeah, we can. We access their

19:30

memory and chats. We can provide you

19:32

with the data

19:34

during a transaction."

19:38

So, what's the solution? Well, Target

19:40

decided that the solution is make it

19:42

your problem, not your problem as

19:44

marketers, your problem as customers. So

19:46

Target decided that the liability goes

19:49

to the customer, not them. So if you use

19:51

their AI agent to to shop and it does a

19:54

wrong thing, you have to pay for it. Oh,

19:57

except one thing. Target is American.

20:00

This is Europe. We do things a bit

20:02

differently. You own the compliance

20:05

obligation. Please keep that in mind.

20:08

provisions related to the EU AI act are

20:12

going to become applicable this August.

20:16

This means that if you're processing

20:18

audience data via LLMs, for example, you

20:22

own the compliance obligation. Okay,

20:24

gossip time. Chill. They're probably

20:26

going to extend the date, okay? But you

20:28

should still be aware and do the work.

20:31

This means that data collection and

20:33

activation is your responsibility

20:37

and

20:39

uh I do not work for the user centrics

20:41

team. I just really appreciate the work

20:43

they put in. So let me tell you about

20:45

this. They acquired the only production

20:47

grade not vibe go vibe coded by

20:50

someone's nephew in a basement that has

20:52

a good LinkedIn following. Okay.

20:55

This is your way to handle business data

20:58

and control what the agent can touch,

21:01

take and do with what it brings home.

21:04

Okay, this works both ways.

21:07

So, thank you very much for coming to my

21:09

talk. It has been very short because I'm

21:12

expecting questions.

21:19

>> Oh,

21:20

>> if anyone has any questions, please

21:22

raise your hands. I'll come around with

21:24

the microphone.

21:26

I will take also complaints.

21:28

>> Any questions?

21:29

>> Fair discussion.

21:31

>> Wait, has anybody had to explain to a

21:34

seuite what's going on with Agentic AI

21:36

in terms of data or not yet?

21:39

You were thinking about it. You're like,

21:41

I don't want to do that.

21:44

Yeah, you you were like, no,

21:48

nobody had to deal with this yet.

21:57

Yes.

21:58

Um I don't know if I have to introduce

22:01

myself, but I'll just go straight with

22:02

the question. Um so for me, one of the

22:07

difficulties when it comes to um being

22:10

searchable through LLMs because I know

22:11

it's a very hot topic. Companies have

22:14

already seen that their um traffic has

22:16

already gone down and bosses are like,

22:18

"What is going on? We're paying you

22:20

money to do this and then you're not

22:22

delivering. Of course, we still have a

22:23

job to do. We got, you know, dogs to

22:25

look after, especially mine. Um, so I

22:28

was wondering like if we are going to

22:32

continue in this current world, the the

22:35

developments that we have on searching

22:37

through LLMs, being discovered by LLMs,

22:40

how far does a company have to go when

22:43

it comes to governance and compliance?

22:44

because that's an important discussion

22:46

in my company where I am trying to see

22:49

what else we need to do in order to be

22:51

compliant. Especially here in Europe, we

22:53

have a lot of regulations. There are

22:55

still some red tape that we also then

22:57

have to also look after. And I'm in a

22:59

sector that's very heavily regulated as

23:02

well.

23:02

>> One of the things that most people don't

23:04

know, and this is me nerding out on LLM

23:06

stuff, but privacy policies and like

23:09

legal legal pages are some of the most

23:12

visited pages. You've noticed that too.

23:15

There's a reason because they're trying

23:17

to make sure when they come on your

23:19

website, it's not to check what they

23:20

know in their parametric memory. So

23:22

that's the official like name for I'm

23:24

checking my how it was trained and

23:26

what's in the back. So when they're not

23:28

sure about what's in the back, they will

23:30

go on your website to triple check. And

23:32

very often they'll go to the legal pages

23:33

to make sure because they want to

23:35

complement the information they already

23:36

have. So there's multiple things here.

23:39

Number one, it's not tied to governance,

23:41

but close to it. Technical branding.

23:44

What is technical branding? It's a thing

23:47

that doesn't exist yet, but that some of

23:48

us are doing going, "Hey, who's in

23:50

charge of the hallucinated URLs? Are we

23:52

going to redirect these? What do we do

23:54

with these? Hey, um, there's a brand

23:57

distortion where

23:59

competitor A offers a free demo,

24:01

competitor B offers a free demo. We

24:03

don't offer a free demo, but since two

24:05

brands already do, they're kind of close

24:06

to us, then see. Yeah, I I you you look

24:10

like you have survived some stuff. So

24:12

what happens

24:14

when it says that we are offering a free

24:16

demo? Restaurants are dealing with this.

24:19

Even local hamburger restaurants, they

24:20

have people going, "Hey, you're special

24:22

that I read on chat GPT and they're

24:23

like, for the love of God, nine." So

24:27

this is a brand distortion problem. Who

24:30

handles it? That also enters technical

24:31

branding because somebody needs to tell

24:33

the branding team, "We've noticed a

24:35

drift. we need better content to feed

24:38

that. So it's not governance related but

24:41

it's a tangent. So these are two

24:43

elements that I would also handle in

24:45

that. But on top of this, what do we

24:47

tell bots? What do we tell agents on how

24:50

to behave and interact with the data we

24:52

offer them? Is that part of marketing?

24:54

Is that part of govern governance? Is

24:56

that in between? That also falls into

24:58

what I would call technical branding.

25:00

Making sure that your brand is very well

25:03

understood. And I'm gonna go one step

25:05

further when it comes to this. Um, I'm

25:07

so sorry. You're in the front row.

25:08

You're gonna suffer. If I tell you

25:10

ketchup, what's the first brand that

25:11

comes into your mind?

25:13

>> Hines. That's familiarity. They have

25:17

established a short circuit in your

25:18

brain. It just shortcut.

25:21

Ketchup means Hines. Well, LLMs do the

25:24

same thing. So, they will put you when

25:26

you look for the best luxury watch.

25:28

There's subcategories, sub stereotypes

25:30

for brands. Investment, fashion. Do you

25:33

want like the engineering type of luxury

25:36

watch? So if your brand doesn't fit into

25:38

that archetype, you're never going to

25:40

show up. You need to explain that type

25:42

of branding as well when it comes to

25:45

showing up. And this is something where

25:48

humans and agents kind of intersect. So

25:52

it goes back to choice architecture. You

25:54

have to understand that humans are going

25:56

to make a choice, but it's also going to

25:58

be mediated whether it be by Google, by

26:00

an agent, by something else. What do we

26:02

give them to satisfy them and get

26:04

through it? So this is where the legal,

26:07

the compliance, the governance pages are

26:09

emerging. LLM.ext is one of those

26:11

attempts to say here's how you should

26:13

behave. So for me to answer your

26:16

question, it would be first of all

26:19

explaining the metrics are shifting, but

26:21

we still focus on the sales and looking

26:23

at what happens. And if the sales are

26:25

not happening, is it tied to bad

26:26

branding at the top? And we fix it. But

26:29

this means that you're no longer

26:30

dangling a carrot. You're brandishing a

26:32

stick going, "Here's everything that

26:35

could go wrong for you if you don't do

26:36

it." So sadly, my answer is you you need

26:40

to stop apologizing for not having the

26:43

data and telling them we have half of

26:45

the data, but the other 50% John Waker

26:49

already knew we are in trouble. Look at

26:51

all of these things that could be

26:52

happening to us, all of these lawsuits.

26:54

Let's fix it. So reign by fear, madam.

26:58

Anybody else?

27:01

If it's not a data question, I don't

27:02

know if you've sensed it. I've been

27:05

working a lot with LLM visibility

27:07

metrics and data and I have opinions

27:09

about them. For example, ranking is no

27:12

longer a thing because you can still

27:14

have the rank in your tool, but if I

27:16

show up as the fifth choice, and the

27:18

fifth choice is you have option A, B, C,

27:21

D, but the best overall is this one. Am

27:25

I upset at being number five?

27:27

Not necessarily. So once again, all of

27:29

our old metrics are very different. And

27:32

the best way I would explain it with

27:33

agents is to say

27:36

Google used to be a librarian. You come

27:38

in and you're like, I need a topic. And

27:40

they're like, here's the shelf with all

27:41

the answers. AI agents and LLMs are more

27:44

like a detective. They're going to ask

27:46

all the questions for you, do all the

27:47

freaking work, and then come back and

27:49

go, "Tada."

27:51

You need to make sure as a brand that

27:53

you show up for every shortcut that they

27:55

take when they do that research. Just

27:57

like with Hines's catchup. That's your

27:59

job when it comes to branding. Yes.

28:03

>> I hear Oh. Oh, sorry. I didn't

28:05

understand that you don't have the

28:06

magical microphone.

28:09

>> Oh, no. I keep it on as well.

28:11

>> Yes. Otherwise, you don't hear me. So,

28:13

when it comes to kind of um traditional

28:16

search, you can research keywords quite

28:18

easily using tools like Google. What I

28:20

find really challenging with prompts is

28:22

actually identifying what the key

28:23

prompts are for you as a brand. Is there

28:25

a way that you would recommend

28:27

researching this and identifying this?

28:28

>> Yes. And it's a nightmare because you're

28:30

going to doubt yourself the entire

28:32

process. So, the reason why is, please

28:35

stay with me. Uh, you need to inject

28:37

some personas in this because if you

28:39

say, "What's the best air fryer?" Hey,

28:41

uh, what's the best air fryer when I'm

28:43

almost 40, pre-diabetic, I'm dealing

28:44

with a picky eater? very different from

28:46

I need an air fryer because I have three

28:48

kids at home and they all want to go to

28:49

McDonald's and I can't. So, you need to

28:53

establish first of all the key personas

28:55

you're going for, not just a keyword,

28:56

the person behind it. And you need to

28:58

understand, I love to use jobs to be

29:00

done. What are they trying to

29:01

accomplish? But then you need to be

29:03

mindful of something else. You are

29:04

baking in biases and you also do that

29:07

voluntarily or not. Because if I say,

29:09

"Hey, does L'Oreal sound like uh their

29:12

latest product is kind of a scam?"

29:14

I'm already biased. The model already

29:16

assumes that I'm going to say something

29:18

negative or want something negative. So,

29:20

you need to be able to explain not only

29:23

are we operating by personas, but we're

29:25

operating by market because you can

29:27

choose to test every every prompt

29:29

against a specific market or just all

29:32

across the board. But we know that in

29:34

China, if you're going to have something

29:36

specific that is like bad luck, it's not

29:38

going to get sold. So, you you want to

29:39

tailor that. And so I've covered

29:42

persona, jobs to be done, sentiments,

29:45

and on top of that, there's an extra

29:47

one.

29:49

You need to define it along your funnel.

29:52

And at the top of the funnel, you're

29:53

trying to figure out what the brand is,

29:55

what's happening, and you can't ask

29:56

generic questions because you're going

29:57

to get really generic answers. You need

30:00

to get a bit more specific. So I have

30:02

what I call system prompts. What are

30:04

system prompts? I'm gonna force the

30:06

machine to give a judgment. And they

30:08

never give you a real judgment. So you

30:10

have to ask on a scale of like 0 to 10,

30:13

can you grade this company or this brand

30:16

or this product against these key

30:19

elements across and against competitors

30:22

and that's where you start getting

30:23

information where you have the

30:25

comparison criteria that pops up.

30:27

There's like the whole explanation

30:29

except that when you're going to share

30:30

this prompt with your team, somebody's

30:32

going to go no human would ever type

30:33

that and it's like I know I'm testing

30:36

the system. So you also need to account

30:38

for that. I I hope this helps a little

30:41

bit. I I can I can go in depth. I

30:43

actually have an article I can share. So

30:47

any Yes.

30:50

>> question. So him first then you

30:58

>> also I know that for the people who

31:00

don't have headphones on, this is the

31:01

most quiet I'll ever be. Okay.

31:06

>> Uh hi there. One question. Uh are there

31:09

any tools that you recommend for tracing

31:10

prompts? So lots of traditional SEO

31:12

tools like SEMrush, a refs, PKI offering

31:18

solutions. Is there any that you can

31:19

recommend?

31:20

>> So I'm not going to give a full

31:22

recommendation because I work with all

31:23

of them, but I have an article on that

31:26

as well. Reconstructing the funnel. So

31:28

when you when you were asking how do I

31:31

pick the prompts? They're starting to

31:32

give you volume for semi.io. They have

31:34

so much clickstream data that they're

31:36

giving you volume regarding which types

31:38

of prompts are more searched. So that's

31:40

one thing. And getting back to you, what

31:42

I would care about if I were to use a

31:44

tool, okay, I'm I'm not being biased

31:46

here. I would ask them, do you have

31:48

clickstream data? Clickstream data is

31:50

everything that people do in apps or on

31:52

the web that is collected.

31:55

And that's where you get to see really

31:57

weird stuff like somebody shopping for

31:58

glasses will go on ancestry.com then log

32:01

into their Hotmail account like is that

32:04

even a thing anymore and then they will

32:06

go on their crypto app and then they

32:09

will take a picture of themselves with

32:10

the glasses to try it on in another app

32:13

and I'm like okay I've seen this data.

32:16

So this clickstream data is what these

32:18

platforms are buying to be able to tell

32:20

you, hey, the prompt tracking that

32:23

you're doing, we have this like weird

32:25

directional data where we keep prompting

32:27

the machine non-stop, but this is what

32:29

people actually do once they have

32:31

accomplished what they wanted to

32:32

accomplish in the LLM. So you need to

32:35

have a solid tool where you ask them,

32:37

hey, what's the source of your

32:38

clickstream data? And by the way, that

32:40

clickstream data is usually USbiased.

32:43

So you need to ask them, hey, if I'm

32:45

advertising in Belgium, what's your

32:46

panel? Tell me like what portion of that

32:49

data is from Belgium so I actually have

32:50

an accurate view. So this is an

32:53

important element. The second thing is

32:55

depending on what you want to track. One

32:57

of the things that I love is like

32:58

concepts. Give me the concepts so I can

33:00

explain to my clients,

33:02

you're not ranking really well your

33:04

footprint when it comes to affordability

33:06

and like being good engineering. You

33:08

suck at both. You're really good at hype

33:10

or something else or colors or whatever.

33:13

So, please stop trying to show up for

33:14

prompts where we're clearly prompting

33:16

about affordability. You will never rank

33:19

unless you change your product. So,

33:22

concepts are one thing that I love and

33:24

good source analysis, in-depth source

33:27

analysis, so you understand actually

33:29

what gets pulled in and why. Those are

33:31

the two things that I really, really

33:33

care about personally. So, I know Malta

33:36

from PKI is here today and he's solid at

33:38

answering questions. I really like his

33:40

platform. I know that I've also worked

33:42

with Seamrush AIO. When you have an

33:44

enterprise situation, that's great. The

33:46

basic package, like the basic Seamrush

33:48

package, for example, for my Canadian

33:50

clients.

33:53

And the reason why I say this is some of

33:55

the cheaper options, they will not give

33:57

you good prompts. They won't pull from

33:59

the clickstream data. They'll kind of

34:01

generate the prompts. And you're like,

34:03

you're you have a thing. You're telling

34:04

me I show up for a question about my

34:06

boutique in Soho and I don't have a

34:08

boutique in Soho. What's going on? So,

34:10

be very careful about this, but we can

34:12

we can take this offline and I can like

34:14

talk your ear off about this.

34:22

>> Um, maybe I got this wrong. So, um,

34:24

>> no, you didn't. Let's go.

34:26

>> Yeah. Uh the question is basically I

34:27

thought you said that we tend to have

34:31

more agents doing purchases

34:34

>> and there I dare to challenge that

34:36

because I would think the LLM would make

34:38

a suggestion and based on that

34:40

suggestion I would then go to the

34:42

website and maybe buy it or not. But you

34:45

made the case for like an agent

34:46

automatically buying something and then

34:48

maybe at too high prices. And I was just

34:50

curious, do you have any numbers? How

34:52

many people do this? How many people

34:54

have an agent buy stuff for them? We we

34:56

need so it's I need to deconstruct this.

34:59

Okay, so there's multiple premises in

35:01

here. Number one, you're assuming that

35:04

everyone is going to go on Chad GPT or

35:07

their favorite LLM and ask them to do

35:08

things. What I'm saying is that Target

35:10

is making an agent available for

35:12

shopping already. So you don't ask Chad

35:14

GPT, you go on the side, you're like,

35:16

"Hey, go for it." So that's one flavor.

35:19

The other flavor is more and more people

35:21

are expecting agents to do the grunt

35:23

work for them to buy the nonfun stuff

35:25

for them. So they're attempting more and

35:27

more and we see that growing. Do we have

35:29

concrete numbers yet? No, we do not. But

35:33

this is where I enter with my own

35:34

premise which is the web got boring. If

35:38

you're telling me I can't ask a magical

35:40

thing to avoid all the pop-ups and all

35:42

the decisions and I can just go probably

35:45

going to do that. That's why I said we

35:47

have a two-speed web. We have the be a

35:49

really fun brand that people want to go

35:51

on to to do stuff or be very very open

35:55

to bots coming and consuming your data

35:57

differently to help you sell more. But

36:00

it remains to be seen which one is going

36:03

to dominate right now.

36:08

But very very good question because I

36:09

know when I first saw agents I'm like

36:11

who would do that? And if I'm in an

36:13

industry that has like high emotion like

36:15

luxury bedding or stuff where bots don't

36:19

have emotions, you can't get them to buy

36:22

this way. Yes.

36:25

I feel like you're going to answer this

36:27

person.

36:28

>> Um I mean I have kind of a question

36:30

that's a bit related to it. Um because I

36:32

feel like um I'm in a brand that is very

36:35

emotional, let's say, but it's kind of

36:38

it has to do with negative emotions

36:40

because I mean care is a great thing,

36:42

but you do not like to think about this

36:44

because it's

36:45

>> Wait, wait, wait. Are you an working for

36:48

accountants?

36:48

>> No, I work for like um uh for for the

36:52

German government from care insurance.

36:55

You get a package with care uh uh stuff,

36:58

but it's not the nicest thing. And we

37:01

noticed like last year with the changes

37:03

with Google KI and so on that um the

37:06

customer journey is basically changing.

37:08

And I was thinking when you were

37:10

speaking about this right now that um

37:12

our customer journey probably changed in

37:15

that way because it's a thing that

37:17

people do not want to think about. They

37:19

just want to get this stuff. Yes.

37:20

>> And because there there I think that

37:23

might be a thing in the future where aic

37:26

buying will be a thing a big thing

37:28

probably. So getting back to you,

37:30

cognitive load is the key. If cognitive

37:33

load is very high and there's no

37:35

enjoyment out of this agent all the way

37:39

like insurance is going to be one of

37:41

those things where I can already tell

37:42

you yes betting I'm betting no

37:47

anything else. I would normally tell you

37:50

I'm highly visible but today I match

37:51

with the background but I hope that if

37:53

you have more questions and you see me

37:54

around please don't hesitate. I love to

37:57

talk about this stuff.

37:58

Thank you very much. Merciu

38:12

you can welcome oh Google Measure you

38:16

met Meta and into it Mailchimp. One more

38:18

round of applause for Miriam please if

38:19

you don't mind. Thank you so much. Have

38:21

a great day. Enjoy your evening and have

38:23

a good fun night. Thank you so much.

38:25

Bye-bye.

More transcripts

Explore other videos transcribed with YouTLDR.

Get the TLDR of any YouTube video

Transcribe, summarize, and repurpose videos in 125+ languages — free, no signup required.

Try YouTLDR Free