Semantic Similarity Rating (SSR)

What It Is

SSR is the elicitation method that makes synthetic-consumer-panels reliable. Instead of asking the LLM for a number, you:
1. Elicit a textual reaction from the persona ("How relevant is this concept for you?" → free text).
2. Embed the reaction and compute embedding-similarity (cosine) against a set of anchor-statements written for each Likert level.
3. Convert the similarities into a probability mass function over the Likert scale.

Why It Beats the Alternatives

The paper compared three strategies:
- Direct Likert (ask for a number) — ρ≈80% correlation attainment but badly unrealistic distributions (KS 0.26–0.39; "safe 3" regression).
- Follow-up Likert (a second LLM scores the free text) — intermediate.
- SSR — wins on both distribution similarity (KS > 0.85) and concept ranking.

Operational settings (paper-verified): GPT-4o / Gemini-2.0-flash (chosen for consistency, not frontier-best); temperature 0.5, top_p 0.9; n=2 samples per respondent is sufficient (not 10 — ~5× cost saving); persona/demographic conditioning is mandatory.

How It Applies to Marketing Factory

SSR is the engine inside the v1 experiment-loop gate. The key operational rule: never let an agent emit a direct numeric self-rating — always elicit text, then map via embeddings. This avoids the "safe 3" failure mode and yields the qualitative reasons that make the panel's output actionable.

Referenced from: ssr-paper-arxiv-2510-08338