ilaida

Synthetic panels · Real statistics

The AI writes the voices.
It never picks
the people.

ilaida runs focus groups with synthetic panelists — sampled from a statistical model of your audience by code, then brought to life by language models. Who they are is drawn; only how they talk is generated.

Fig. 01 — Panel composition, 6 panelistsseed 184529

RELIGIOSITY2 observant · 4 secular-leaning · target 40/60 · quota

INCOME BAND2 under ₺40k · 3 ₺40–80k · 1 ₺80k+ · target 30/45/25

AGE BAND1× 18–24 · 3× 25–34 · 2× 35–44 · target 20/45/35

Counts allocated by largest remainder. Reproducible from seed.

The difference

Ask a model for a “40% religious panel” and you get six stereotypes.

The usual way

Improvised panels

When a language model invents the panel itself, it samples from its imagination: each “religious” panelist becomes a caricature of the category, demographics drift run to run, and nobody can say what population the panel represents — or reproduce it tomorrow.

The ilaida way

Sampled panels

A population model states your audience's attribute distributions and how they interact — metro residents skew less observant, younger panelists skew lower income. Code draws each panelist from that model with exact quotas and a stored seed. The AI receives the result as fixed fact and writes one specific person on top. Composition is arithmetic, not improvisation.

Method · Brief to report in minutes

01

Write the brief

Product, audience, the questions keeping you up. Plain language.

02

Model the population

The system estimates your audience's attribute distributions and the dependencies between them — every number visible, every number editable.

03

Draw the panel

Code samples panelists from the model with a seeded random draw and exact quotas. Same seed, same panel — forever reproducible.

04

Run the session

A moderated focus group unfolds live: warmup, deep dive, concept test, wrap-up. Each panelist answers in character, shaped by what was sampled.

05

Read the report

Findings with confidence scores and verbatim quotes, prioritized recommendations, the full transcript in the appendix.

Fig. 02 — Population model (excerpt)
ReligiosityObservant 40 · Secular-leaning 60 · medium
UrbanicityMetro 55 · Town 30 · Rural 15 · high
Price posturePremium 25 · Value 45 · Skeptic 30 · medium

Rule — Metro residents skew less observant.

if urbanicity = Metro → religiosity × {Observant 0.6, Secular-leaning 1.6}

Every number disclosed before sampling — estimates labeled, never laundered.

Fig. 03 — Session transcript (excerpt)
Moderator

Same question to you, Emine — would the certification on the label change what you pick up?

Emine

For me it's not a label, it's the default. What I'd question is the price — devotion doesn't make me careless with money.

Can

Honestly I'd walk past it. But if my mother visits and it's in the fridge, that's one argument avoided. That's worth something.

Sampled attributes reach the conversation — a devout panelist argues like one, not like a label.

Exhibit — What a finding looks like

01

Certification reads as default, not differentiator

86% confidence · 5 of 6 panelists corroborate

Observant panelists treat certification as table stakes and evaluate on price; secular panelists are indifferent but value household harmony. Marketing it as premium misreads both.

“For me it's not a label, it's the default. What I'd question is the price.”

— Emine · sampled: observant, value-oriented

What you get

01

Auditable composition

Every panel ships with its receipts: target shares vs. actual counts, the seed that drew it, and honest notes when an option can't be represented at panel size.

02

Live moderated sessions

An AI moderator follows a discussion guide, probes contradictions, and pulls quiet panelists in — while insights surface in the margin as the session runs.

03

Findings with receipts

No vibes. Each finding carries a confidence score grounded in cross-panelist corroboration and the verbatim quotes behind it.

04

A method you can defend

Nothing is a black box: the population model, its confidence labels, the seed, the quotas — all of it is visible and goes in your readout.

05

Same person, new story

Regenerate a panelist's narrative without touching their sampled attributes. Composition holds; the writing changes.

06

Minutes, not weeks

No recruitment, no scheduling, no incentives, no no-shows. Write the brief in the morning, argue about the findings after lunch — then rerun with a new price point before dinner.

Stop interviewing
a model's imagination.

Run synthetic market research you can defend in the readout — sampled, seeded, and quoted.

Open the workbench