Market research · Zürich, Switzerland

Market research with two engines — real people and simulated respondents.

Boomerang Ideas is the methods house behind two outlets: boomerang.ai — AI market research for SMEs, affordable and fast — and boomResearch — fully fledged research over new channels, for institutes and insights teams. We build both engines ourselves and measure them against each other continuously.

Best of Swiss Web 2026 – Marketing, Silver

Best of Swiss Web 2026

Two outlets

Market research for SMEs. Market research for institutes.

boomerang.ai

For startups and SMEs: AI market research, affordable and fast. Simulated studies in minutes — anchored with real respondents when it matters.

boomerang.ai →

boomResearch

For institutes, agencies and insights teams: fully fledged market research over new channels — recruiting, fieldwork, trackers, full methodological control.

boomResearch →

The institute

One methods house, three jobs.

Research

Method research with Universität Basel and Johns Hopkins University, validation studies with Universität Zürich — every claim tested before it ships.

Method management

The shared engine room: sampling methods, quality standards, benchmarks and Confidence Guard™ — maintained once, used by both outlets.

Governance

Boomerang Ideas steers both outlets — one scientific standard, one data-ethics line, whichever door a client comes through.

Our philosophy

Enabling informed, fast and confident decisions — in any situation.

Science firstValidated before claimedNo black boxes in the AI — we say when not to trust a simulationData minimalismNo BS

Easy, fast and affordable high-quality research — for marketing teams testing campaigns, product teams shaping features, executives validating a strategy, and analysts forecasting votes. The method changes with the question; the standard doesn't.

How we redefine market research

Two engines, one questionnaire.

Human Sampling

Real respondents, recruited where they already are

81% passed the attention checks, against 37 % on a commercial panel — University of Zurich validation study II

Recruited ad hoc on Instagram, Facebook, TikTok, Snapchat, LinkedIn and WhatsApp — micro-targeted for the specific question, reaching groups no panel holds.

Cross-quotas over age, gender and region, steered live by the engine.
Incentives benchmarked to fair per-minute pay, released after quality checks.
Attention checks, speeder and duplicate detection in every field.

How it was validated ↓

Silicon Sampling

Simulated respondents, answering the same questionnaire

6.9/10 average plausibility in August 2026 — scored by an independent judge model

In plain words: a language model that has learned from decades of representative, proprietary survey data from established partner institutes plays the respondents. Technically, persona-conditioned large language models (LLMs), fine-tuned on that data — decades of it from gfs.bern — one model call per synthetic respondent.

Measured on held-out Swiss surveys: overlap is 1 − mean .
Strict integrity rules: contaminated surveys dropped, seen questions excluded.
Every run scored by Confidence Guard™ (0–10), declared as simulation and labelled in every export, never mixed into a human sample.

How it was validated ↓

Science-first

Validation first, claim later.

Human Sampling Silicon Sampling

2021University of Zurich pre-study

Thirteen Boomerang polls over seven months (N between 416 and 976), checked against two national referendum results and against a control sample from an established Swiss panel. In 2 out of 3 political polls, Boomerang predicted the vote result more closely than the established polls.

2023/24University of Zurich validation study II — incl. Qualtrics head-to-head

Same questionnaire, our field vs. Qualtrics' commercial panel, USA, 1'000 respondents each: 81% vs. 37% passed the attention checks, 42% vs. 20% completed both waves.

12/2024gfs.bern federal vote surveys — for SRG, the Swiss public broadcaster

Since December 2024 our channel is the social-media field layer in gfs.bern's mixed-sample vote surveys — their forecast, our field. Federal ballots fall on up to four fixed dates a year, so each one gives an objectively measurable result to compare against and calibrate with.

2023Large Insurer A/B — control study

Head-to-head against a client's own fieldwork, on their instrument, to see how good our samples were. Results aligned with the client's internal benchmarks and confirmed the plausibility of both social-media samples and hybrid samples of social media and panel — all human respondents.

2024Swing State, US presidential election 2024

0.25% off the presidential result, against an industry average of roughly 3.6% — the accuracy came from coverage. One race: a case, not an error rate — internal analysis of our own survey, not a published validation.

2026Benchmark v1 — Aug 2026

Simulated answer distributions land on average 82.6% of the way to the human ones, across five Swiss surveys held out of training.

runningJoint research with Universität Basel

«Confidence-Driven Augmentation of Synthetic Survey Data», Innosuisse-funded, with Universität Basel, in collaboration with Johns Hopkins University, with SBB and gfs.bern as project partners. Joint research since spring 2026; first publication expected Q1 2027.

Method documentation: Institutsprofil & Methodik (PDF). Reports: Swiss Insights News #11 · University of Zurich Pre-Study 2021 · University of Zurich Validation Study 2023/24

Silicon Samples · simulated respondents

How close is a simulated answer to a human one?

Our first benchmark measures the distance between the distribution of answers real people gave and the distribution our production model produces.

82.6%

Boomerang Model — mean overlap with the human answer distributions

<52%

Score of untuned frontier models on the same test

Fig. 1 — Benchmark v1, Aug 2026. Overlap = 1 − mean total variation distance. Topline only, five real Swiss surveys held out of training, production model. Measures faithfulness to the human survey, not correctness about the world, which is tracked separately. Cohort-level scoring follows. Full method and all models in the technical write-up.

Confidence Guard™

A simulation that always answers is worth little. The useful part is knowing when it shouldn't be trusted — so every result carries a plausibility score, and the platform says when to go and ask real people instead.

Plausibility

0low10high

Assigned by an independent judge model.

Rates each simulated result for consistency with the question, target group and answer logic. It measures trust in the simulation, not correctness about the world.
Validation of the criterion against human re-fieldings runs with the University of Basel and Johns Hopkins.
Automatic escalation per cohort is in development with our research partners.

Every human sample feeds back into the model, which is why the curve climbs — dips included. 5.6 → 6.9.

Fig. 2 — Average plausibility across all simulations per month, Dec 25 – Aug 26. Everyday use, not lab conditions: arbitrary topics, started by users themselves. Scored against every human sample we field, anonymised.

Startup mode

A research company that ships like a startup.

Boomerang Ideas AG, Zürich — founded 2021. Institute member of Swiss Insights, in strategic partnership with gfs.bern and Panter, and social-media sampling partner for gfs.bern's SRG Trend surveys. Innosuisse-backed, now in a third study in joint research with SBB, gfs.bern, Universität Basel and Johns Hopkins University.

A small team that runs its own field and its own benchmark, on models we run ourselves — and publishes the misses along with the hits.

Partners & memberships