← All studies

The candidate message scored no better than the current one.

A self-initiated Message Lift test for a global hotel group: the current executive-lounge message against a candidate rewrite with cocktail-messaging, run through a simulated buyer panel. The candidate was stopped.

Simulated31 Jul 2026n=30 per messageDeepSeek · ungroundedNot a client

The brand, a global hotel group, was not a client, did not commission this test, did not participate, and has not endorsed the findings. The brand is anonymized throughout; the value of publishing it is the no.

02 / The test

One Current message, one New message, only the message changed.

Both arms ran the same shared contexts: a proposed executive-lounge message for the group’s current framing against a candidate rewrite that leads with cocktail messaging. Both messages were quoted verbatim in the prompts, and the exact prompt set is disclosed to clients on every report. Buyer response is a comparative simulation: a synthetic panel reacts in free text, and the reactions are scored on a 1 to 5 expressed-intent construct with anchor sets from independent peer-reviewed research (arXiv:2510.08338).

01 / The result

No lift in the displayed means.

A descriptive comparison on the tested construct. This is not a statistical significance or equivalence claim.

Current message
3.45Buyer response, 1 to 5 expressed-intent construct
New message
3.41Same contexts, same construct. Only the message changed
Shift
−0.04No lift. The candidate does not ship

Thirty simulated buyer reactions per message (six prompt cells, five repetitions each), all on one model route: DeepSeek via its disclosed API, ungrounded. Every score traces to a stored, timestamped answer.

03 / Limits

What this test does not show.

  • Simulated, not human. The panel is synthetic. Reactions are benchmarked against the independent method’s anchor sets, not observed human behavior.
  • Comparative only. The result compares two messages inside one test. It is never a sales forecast, never a revenue prediction, and never a statement about the brand’s real performance.
  • One route, one construct. The scores come from one model route and one expressed-intent construct; they do not generalize to other routes or to human guests.
  • A snapshot. Valid for the tested prompts, the one model route, and 31 Jul 2026. Models change, and so will this.
Request a brand audit Read the Insta360 study