Home

How it works

One brief in. One defensible report out. Here's everything in between.

ilaida runs focus groups with synthetic panelists. The whole trick is in who gets to sit at the table — so let's follow one study from start to finish. Pick an industry; every example on this page will follow it.

Act 1 · The brief

You describe the study in plain words.

No screener grids, no fieldwork briefs. What are you testing, who buys it, and what do you need to know. A paragraph is enough — everything downstream is built from it.

The brief · Consumer goods

Testing

A premium pistachio–tahini snack bar

Audience

Shoppers in Turkey who buy packaged snacks every week

Key question

Would they pay ₺95 when the standard bar costs ₺60?

Act 2 · The population

Before inventing anyone, we write down who your audience actually is — as percentages.

Ask an AI for “six customers” and you get six clichés. So ilaida first builds a small table: which traits matter for this study, and how common each one is in your audience. It also notes how traits pull on each other — real populations aren't flat.

Every number is visible before anything else happens. Each one is labeled as an estimate — or as verified when it comes from real statistics you've entered. And you can simply type over any of them.

Population table (excerpt)
Income bandguaranteed mixlower-middle 40 · middle 40 · upper 20
Where they shopmetro 45 · town 35 · rural 20
Snack postureingredient-readers 30 · taste-first 55 · price-only 15

Interaction — Metro shoppers are more likely to be ingredient-readers — so the city panelists lean label-conscious, the town panelists don't.

Act 3 · The draw

The panel is drawn like a lottery — weighted by those numbers, never improvised.

A random draw — plain arithmetic, not AI — assigns each of the six seats its traits from the table. Marked attributes get a guaranteed mix: a 40/60 split over six panelists is always 2 and 4, never “roughly”. The draw has a ticket number, so the exact same panel can be reproduced later, for anyone who asks.

Only then does the AI write each person: a name, a job, a life that fits the drawn facts. The facts are locked; the storytelling is layered on top.

Panelistdrawn, then written

Elif, 34

Pharmacist, Istanbul

middle income · drawnmetro · drawningredient-reader · drawn

I flip the package before I look at the price. If the list is short, we can talk.

Panelistdrawn, then written

Hasan, 51

Shop owner, Denizli

lower-middle income · drawntown · drawntaste-first · drawn

₺95 for a snack? It had better feed me, not just impress me.

Act 4 · The room

A moderated discussion unfolds — warmup, deep dive, your concept, wrap-up.

An AI moderator runs the room like a researcher would: easy questions first, follow-ups on what people actually said, quiet panelists pulled in by name. Each panelist answers in their own voice, shaped by their drawn facts — the label-reader reads labels, the burned investor asks about the exit.

Watch it live, or close the laptop — the session runs on our side and is waiting, finished, when you come back.

Session transcript (excerpt)
MOD

The bar costs ₺95 — about 60% more than the one you usually buy. Honest first reactions?

Elif

Depends entirely on what's in it. If it's pistachio, tahini and nothing I can't pronounce, I'd buy it for the office drawer. But I'd check the price per gram first — premium bars cheat on size.

Hasan

No. My customers wouldn't either. We know what a snack should cost.

MOD

Hasan, is there anything that would change your mind at that price?

Hasan

If it replaced a meal, maybe. As an indulgence, never. That's what lokum is for, and lokum is cheaper.

Act 5 · The report

The transcript becomes findings — each one with receipts.

The session is distilled into a handful of findings. Each carries a confidence score — based on how many panelists independently corroborate it, not on enthusiasm — and the verbatim quotes behind it. Then recommendations: actions, not themes.

The whole study files itself in your archive and exports as a branded PDF — population table, panel profiles, findings, full transcript. Everything a skeptical stakeholder would ask for is already in there.

Finding 01 (sample)

The premium price survives only with an ingredient story

72% confidence

Label-conscious city shoppers will pay ₺95 — but they justify it with the ingredient list, not the word “premium”. Taste-first shoppers reject the price outright regardless of framing.

I flip the package before I look at the price.

Elif · ingredient-reader

Recommendation — Lead shelf materials with the short ingredient list, not “premium”. Skip rural and town distribution at launch — the price is a non-starter there.

Where this is going

The roadmap is one idea: ground the population in ever more real data.

Live today

Verified market numbers

Admins can already enter real statistics — census splits, survey results — as the fixed base for a market. Those attributes show a “verified” badge with the source, and the AI may only express audience skew on top, never rewrite the numbers.

Up next

Official statistics, built in

Population bases imported directly from public statistical offices and large recurring surveys, kept current — so a study in a supported market starts from cited, dated numbers by default instead of estimates.

Up next

Bring your own polls

Upload survey results or crosstabs as a file and have them become distributions and interaction rules automatically — your tracker data becomes the ground your panels stand on.

Later

Survey-scale panels

The same sampling engine, but hundreds of respondents answering structured questions — quantitative reads with the same auditable composition, next to the focus-group depth.

Later

Calibration studies

Side-by-side runs against real fieldwork on the same brief, published as calibration reports — so you know where synthetic panels track reality closely and where to stay careful.

Now run one yourself.

Your first study takes about as long as reading this page did.

Open the workbench