Accuracy Disclaimer

Splitroom is a synthetic audience platform. Given two ad creatives and a target audience description, it runs a multi-engine AI pipeline that:

• Extracts a symmetric inventory of what each creative contains, and cross-verifies both images to catch anything a single-pass reading would miss.
• Scores both creatives on a curated set of canonical dimensions per product category.
• Generates a synthetic panel of consumer personas with weighted priorities.
• Asks each persona to react through their segment lens, using dimension-constrained reasoning.
• Aggregates results into a report with a verdict, per-theme editorial narrative, dimension-level distributions, a debate transcript, and a recommendation.

The report is a structured description of how a simulated panel might respond. It is meant to be read the way a strategist reads a research summary: as an input to thinking, not as a forecast.

Not a predictor of real-world performance

Nothing in the report tells you what a real audience will do with your creative in market. Splitroom does not predict click-through rate, view-through rate, conversion rate, cost per acquisition, return on ad spend, sales, brand-lift, ad recall, or any other real-world metric. If you are making a media-spend decision, Splitroom is one input among several — never the only one.

Not a substitute for real research

Synthetic panels do not replace surveys, interviews, focus groups, usability testing, live pilot campaigns, or any other primary consumer research method. A synthetic panel is faster and cheaper than a real one, but it is also a simulation — the personas are not real people and their reactions are not real reactions.

Not statistically representative

The panel is not a random sample of any real population. Even when you configure demographic filters, the personas are generated by a language model, not sampled from a demographic register. Confidence scores in the report describe internal pipeline diagnostics — how much variance the panel produced, how similar the two creatives are, whether the verdict looks overclaimed — not statistical confidence intervals on a population estimate.

Not a compliance or brand-safety review

Splitroom is not a legal, regulatory, medical, financial, or brand-safety review. It does not tell you whether your creative is compliant with advertising standards, industry codes, or platform policies in any jurisdiction. You are responsible for those reviews.

Verdict labels

The verdict is expressed as a verbal band — “Directional lean,” “Moderate lean,” “Strong lean,” or “Very strong lean.” These bands are calibrated to be modest: when the two creatives are similar and the panel returns a very lopsided percentage, the pipeline caps the band at “Directional lean” and surfaces a Spread sanity check warning. Trust the band, not the raw percentage.

Confidence tiers

Reports show three internal confidence scores — Observation, Reasoning, and Verdict — each on a 0–100 scale. Low scores are not a bug; they are the pipeline honestly reporting that a stage did not behave well. A low Reasoning score means the panel converged too quickly on one dimension; a low Verdict score means the panel’s preference is not well supported by the underlying differences.

Spread warnings

When the two creatives are highly similar but the panel returns a very one-sided verdict, the report surfaces a warning explaining that a real audience would not likely produce such a split. Take those warnings at face value.

Segment-level analysis

Alternative-strategy recommendations (“run the other creative for this segment”) are heuristic. They surface where the losing creative performed relatively better inside the panel. They do not guarantee that segment will behave the same way in market.

The outputs are generated by large language models operated by third-party providers (currently including Anthropic and OpenAI via OpenRouter). Language models can be:

• Inconsistent across runs — the same inputs can produce slightly different outputs.
• Wrong about facts — including facts about your creative material or your product category.
• Anchored on training data — reflecting patterns learned from prior text and image data that may not match your specific audience or market.
• Susceptible to prompt sensitivity — small changes to phrasing can change the verdict.

Our pipeline includes deliberate guardrails (symmetric measurement, dimension-constrained reasoning, spread-sanity checks, monolithic verdict warnings) to reduce these failure modes. It does not eliminate them.

You are solely responsible for how you use the report. In particular:

• Do not present outputs as human research or primary consumer research.
• Do not skip live testing or pilot campaigns because Splitroom said one creative won.
• Do not use outputs to make consequential decisions about individuals or protected classes.
• Do disclose the use of AI-generated audience simulation when your client, platform, or regulator requires it.

To the maximum extent permitted by applicable law, Splitroom and Satak Technologies disclaim all warranties as to accuracy, reliability, and fitness for a particular purpose. We are not liable for any decision you or anyone else makes based on the outputs, including without limitation media-spend decisions, creative decisions, and campaign strategy decisions. See our Terms of Service for the full disclaimer and limitation of liability.

For questions about how to read a specific report, contact us at hello@satak.tech. For legal questions about this disclaimer, contact legal@satak.tech.