Leonard Kinzinger

How do changes in a video's visual features over time affect how consumers respond to ads? We use machine learning to build a multi-dimensional measure of visual variety and apply it to more than 2,300 real video ads across TV and social media and several product categories. Across three studies, visual variety has an inverted U-shaped relationship with intended and actual ad clicks and purchases. A controlled experiment shows why. Stimulation rises with visual variety, while comprehension peaks at moderate levels and drops when variety becomes excessive, and both mediate the effect in parallel. Moderate variety therefore works best. We derive guidelines for advertisers and provide a free, open-source web tool that scores videos on visual variety.

LLM-based digital twins promise faster and cheaper market research, yet most existing twins are either generic personas built from a few demographics or detailed twins built from purpose-collected surveys and interviews. We build individual-level twins from data firms already hold, using the German Socio-Economic Panel (SOEP). Across three open-weight LLMs, five entropy-ranked information depths, two embedding methods and two reasoning modes, we score more than 2.1 million twin responses from 500 participants on 183 held-out questions. Twin quality rises with information depth but flattens beyond the 75th entropy percentile. Feeding the raw dialog history of past answers instead of a narrative persona summary improves held-out accuracy in every configuration, and explicit thinking improves rank-order correlation without changing accuracy. The best configuration reaches 78.8% accuracy and a Fisher-z correlation of r = 0.590.

Waveform-domain audio watermarks withstand many conventional distortions but can degrade substantially under neural codec resynthesis. We test whether continuous neural codec latents offer a better embedding space. Building on a frozen pretrained EnCodec model, a feedforward embedder maps a multi-bit payload to an additive latent perturbation that is decoded through the unchanged codec decoder. Compared with AudioSeal and WavMark, the latent watermark degrades more gradually under repeated and low-bitrate EnCodec resynthesis, transfers to the unseen DAC codec, and keeps high detection under most waveform distortions. Much of this robustness appears even without codec-resynthesis supervision, and codec-aware training strengthens it further. Learned perturbations survive codec cycling better than random perturbations of equal norm, and preservation depends more on channel-specific allocation than on temporal structure. Perceptual quality stays close to the frozen EnCodec reconstruction. Continuous codec latents are therefore a promising embedding space for watermarks that stay robust to neural codec resynthesis.