From Stanford to Marketing Intelligence: How We Build AI Focus Groupsمن Stanford إلى Marketing Intelligence: إزاي بنبني AI Focus Groups في Monzology

At Monzology we do not use AI just to write copy faster or hand you twenty ad ideas before lunch. We build AI-powered marketing systems that can sit inside bigger decisions:

• Will this positioning work in a market we have not entered yet?
• Will this message build trust, or plant doubt?
• Will this price read as attractive, or quietly lower the perceived value?
• Does the same campaign really travel from Egypt to Saudi Arabia unchanged?

That is how we ended up building AI Focus Groups. But instead of writing a prompt that tells a language model to "act like five customers", we started from deeper research into generative agents, and engineered an approach on top of it that fits the way marketing work actually runs.

It started with Generative Agents

In 2023, a team from Stanford and Google Research published Generative Agents: Interactive Simulacra of Human Behavior — software agents that simulate aspects of human behaviour convincingly enough to be studied.

The part that caught us was not that the AI "plays a person". It was the architecture. The agent keeps a record of its experiences in natural language, reflects on them to form higher-level conclusions, and retrieves what is relevant when it has to decide or act.

In other words: memory, reflection, retrieval, planning, behaviour — which is a very different thing from prompt, answer.

Memory, reflection, retrieval, planning, behaviour. The part of the 2023 Stanford / Google Research work that mattered to us was not that software can act human. It was the architecture: an agent that keeps a record of experience in natural language, draws higher-level conclusions from it, and retrieves what is relevant when it has to decide. Source: Generative Agents: Interactive Simulacra of Human Behavior.

For us that was a starting point. Not so we could take finished technology and put a logo on it, but so we could ask a better question: what happens if you take the logic of generative agents and turn it into an intelligence layer inside the marketing process?

From agent to AI Focus Group

Picture a brand preparing a major launch in Riyadh. Before the heavy spend starts — creative production, media, influencers, activations, landing experiences — there is one question worth answering: is the proposition itself understood and convincing to the different segments we are targeting?

That is where an AI Focus Group can enter, as an early layer in the decision process.

We define a set of audience profiles. Each persona carries a context, motivations, pain points, buying behaviour, objections and language. Then we put the proposition in front of them.

Not one persona speaking on behalf of the group. Each persona gets its own independent interaction — and that is an engineering distinction, not a cosmetic one.

Ask a single model to "give me the reaction of five customers" and you usually get one opinion rewritten five times. Each persona runs as its own interaction, with its own context, so the disagreement is built into the system rather than requested from the model.

We are not looking for the AI that says "nice"

This is one of the biggest problems with using LLMs in marketing. The model wants to be helpful. Show it your campaign and ask what it thinks, and you get: the idea is strong, the message is clear, maybe tighten the CTA.

Thank you. That is not a focus group. That is AI being polite.

So we design the interaction around the objection. The persona is not only asked to judge the idea. It is asked to explain what is making it hesitate — and then what would have to change for it to agree.

That changes the nature of the output. Instead of "the campaign is not convincing", you move toward "the campaign is not convincing because the value is not clear to me". And the next question becomes: what piece of information, if it changed, would change the decision?

For a marketer, that is actionable.

A verdict is not an insight. "I am not convinced" is where most AI panels stop. The sequence pushes past it to the objection, and then to the condition — the thing that would have to change. That last answer is the one a marketing team can actually act on.

What the newer research adds

Here it gets more interesting. In a more recent study, Stanford-led researchers working with collaborators at other institutions built agents to simulate 1,052 real people, using different sources of information about them.

They compared how the agents were grounded. Agents built from interviews with real people reached 83%. Agents built from structured surveys reached 82%. Combining both reached 86%. Agents built from demographic descriptions alone reached 74%.

Those numbers have to be read correctly. They are not absolute accuracy. They are measured against how consistent a human being is when asked the same questions again two weeks later. The human participant is the benchmark.

And that matters to us more than the number itself, because of what it says plainly: the more real information you have about a person, the better a simulation can represent them.

Read this carefully — these are not absolute accuracy scores. In the study, agents were built for 1,052 real people and scored against the General Social Survey. The yardstick is the human: how consistently the same person answers the same questions again two weeks later. An agent at 86% reproduces its person almost as reliably as that person reproduces themselves. Figures from the arXiv paper, arXiv:2411.10109.

And this is the part that changed how we talk about it

Our AI Focus Group is not currently built on real interviews with real customers. We build personas from what the marketer defines: age, city, role, behaviour, pain points, objections, language.

So, honestly: we are closer to the 74% condition. Not 83%. Not 86%.

And we are not going to use the research numbers as if they were Monzology's accuracy, because that is not what the research said. That discipline is exactly what turning research into commercial technology requires.

So what is it worth, knowing the limits?

The value is not that the AI decides for you. The value is that it widens the range of questions you can ask before a decision reaches the market.

A brand entering a new market

A company moves into a Gulf market with brand messaging that was originally built for a different one. The focus group can surface that the problem is not the creative at all — it is language, cultural context, perceived relevance, trust. That sends the team back to positioning before the problem shows up as campaign performance.

A premium proposition

A brand launches something premium. The team believes the price and the language communicate exclusivity. A second persona may read the same signal as overpriced. A third asks: what am I actually getting for this price? The focus group does not pick the winner. It exposes the difference in interpretation.

Banking

A bank is preparing a proposition for affluent customers. The question is not "is the ad beautiful". It is whether the message builds trust, whether it speaks the customer's language, whether the benefit is clear, and whether there is an objection hiding in privacy, complexity or the cost of switching. The focus group can build that layer of questions and objections before the campaign meets the market.

Telecom

An operator launches a proposition aimed at a particular segment. The same offer can be attractive to a price-sensitive customer and irrelevant to one who is buying convenience or premium service. Instead of treating "the audience" as one block, you break it into positions.

Real estate

A developer prepares a large launch. The decision is not whether the headline is nice. Is the value proposition understood? Is the location read as an advantage? Is the pricing signal right? Is the lifestyle promise believable? What is the first objection? And who in the audience rejects the premise the campaign is built on in the first place?

These are questions an AI Focus Group can start to take apart before market validation.

We build the system to disagree

We do not like AI that returns a nice-looking answer and stops. So the system carries constraints.

Each persona interacts independently. Objections are a mandatory part of the output. The scale is positive / reserved / negative / undefined rather than good / bad. A persona that fails to answer returns a failure, not an invented answer. And when the sample is small we report counts, not percentages.

If two out of three personas rejected it, we say 2 of 3. Not 67% of the market. We are building a decision-support system, not a machine for manufacturing scientific-looking numbers.

POSITIVE RESERVED NEGATIVE UNDEFINED FAILURE

What we report2 of 3 personas rejected it

What we refuse to report67% of the market rejects it

When a persona cannot answerFAILURE — logged, not filled in

What we refuse to doInvent a plausible answer

Four verdicts, not good-or-bad. Counts, not percentages. A logged failure, not an invented answer. A small sample does not become a market statistic because you wrote a percent sign after it. The constraints exist so the output stays a decision-support tool instead of a machine that manufactures scientific-looking numbers.

This is not an isolated SaaS product

The AI Focus Group is part of something bigger. At Monzology we build AI-powered marketing systems, which means AI enters at different stages of the work: research, intelligence, strategy, audience, creative, content, testing, optimisation.

The focus group can be a layer inside that system. Not to replace the marketer, and not to replace research — but to add another layer of intelligence before the decision reaches the market.

The focus group is not the system. It is one intelligence layer inside the way we build AI-powered marketing services — sitting between strategy, audience and creative, before anything reaches the market. It does not replace the marketer and it does not replace research.

And there is still something we have to prove

The virtual focus group is built and running. But we are not putting it in front of clients as a standalone product yet.

Before we do, we need to calibrate it against three to five real campaigns whose actual results we already know. Simulation against reality. Where was it right? Where was it wrong? Where was the persona description too thin? And where did the system need real evidence instead of a profile?

After that we can decide what deserves to become a wider commercial capability.

BuiltThe virtual focus group is engineered, running, and used inside our own work.

Not yet sold as a standalone productWe are not putting it in front of clients as a product until it has been calibrated.

Calibration — in progressThree to five real campaigns whose actual results we already know. Simulation against reality. Where was it right, where was it wrong, where was the persona description too thin, and where did the system need real evidence rather than a profile?

This is the part we are not going to hide. Only after that calibration can we decide what deserves to become a wider commercial capability.

Because to us, AI is not impressive for looking clever

The technology we work on is not a new way to write content. It is a new way to build decision layers around marketing.

We started from research into generative agents. We took an architecture and a set of ideas from it. We re-engineered them inside a marketing environment. And we use them as part of how Monzology builds AI-powered marketing services.

01

Published research

Stanford and Google Research published the generative-agent architecture in 2023, and a Stanford-led team published the 1,052-person simulation study in 2024. That is the source of the idea and the evidence about its limits — it is not our technology, and we do not present it as ours.

02

Monzology engineering

We took the architecture and re-engineered it for a marketing environment: independent persona runs, a mandatory objection, a fixed verdict scale, counts instead of percentages, and a follow-up sequence that ends in a change list.

03

Commercial application

Used as an early decision layer on real marketing questions — market entry, premium positioning, trust in financial services, segmentation in telecom, the promise behind a property launch.

The AI Focus Group is one example. But the bigger question for us is this: what happens when we stop using AI only to produce marketing, and start using it to test our own thinking before we produce anything at all?

That, to us, is where real Marketing Intelligence starts.

Stanford University · Google Research · UIST 2023

Generative Agents: Interactive Simulacra of Human Behavior

Park, O'Brien, Cai, Ringel Morris, Liang, Bernstein — research.google/pubs

Stanford-led, with collaborating institutions · 2024

LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals

1,052 participants. Originally published as "Generative Agent Simulations of 1,000 People" — arXiv:2411.10109

We build on published research. We do not claim it as our own, and we do not borrow its numbers to describe our own accuracy.