The Prompt Set Is Part of the Measurement: AI answer visibility for destination brands: the persistence of a chosen frame, what a prompt set is made of, and why pinning the model does not pin the instrument
Year: 2026
Type: preprint
Abstract: Brand visibility in AI-generated answers is reported as a rate over a set of prompts chosen by the measuring party, and on the evidence below that choice moves the reported number by more than the measured surface moves on its own. The evidence is drawn from one vertical, travel destination marketing. Six prompt frames were built to one brief by documented rules, three of which ask for the identical object, and used to measure 46 destinations in two substitution categories, small coastal towns in California and destinations in the American southwest, on Gemini, OpenAI and Perplexity across four captures. A frame chosen because it most favored a destination in one measurement period still favored it in the next. That holds on Gemini over twelve, twenty-five and thirty-seven days, and on OpenAI over thirty-four days with the model version held fixed. The median gain over the six-frame portfolio is 12.43, 11.46 and 10.76 points of mention rate on Gemini and 8.54 on OpenAI, each against a drift of 0.83 to 1.36 points when nothing is chosen. The advantage a chosen frame confers is a reordering of the destinations against one another, not a rise in the level of the whole market, and only the first of those can be selected for a single brand. A factorial experiment on 600 further prompts says what the prompt set is made of: topic content carries more of the variance in a destination's figure than the breadth of the category asked about, which in turn carries more than the request template, in all 24 combinations of engine, cluster, population and scale. Pinning the engine is not a remedy. Over thirty-four days, with the same request to the same pinned model, the share of answers citing a source fell from 0.620 to 0.514 and destination rates drifted past their sampling floor. Over the same study one engine withdrew the model version under measurement while continuing to list it. Every figure is conditional on this vertical, these 46 destinations and the levels these designs contain, and nothing here is shown to hold in another category. No commercial product was measured.
Author: Birba, Trevor
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7222904407