Genesis Research
We study which brands and products appear, how products are recommended, and where returned links lead. Older papers examine how models compare images and repeat preferences.
Read each study alongside its method, sample, and limitations.
Together they cover visibility, product representation, repeated preferences, and returned destinations. Each finding stays bounded by its method and source.
Compares how CLIP, SigLIP, and DINOv2 handle fashion images, including uncertainty, impostor detection, and collection cohesion. Also published as “A Vision Benchmark for Fashion AI”.
Compares outputs from nine language models to identify recurring designers, garments, and fashion narratives in prompt-based preference tasks. Also published as “What LLMs Repeatedly Choose”.
Records what ChatGPT and Gemini return after a brand recommendation: a brand-owned product page, a retailer or marketplace destination, or no usable path.
Measures brand mentions and returned URLs across 30 shopping answers. The study reports observed owned-store links, other destinations, and the limits of a one-model, one-day sample.
These studies report observed outputs and defined comparisons. They do not tell us what every shopper sees or establish purchase completion or revenue.
What appears in a defined AI-shopping test, including the returned product and destination.
Whether it distinguishes brand, season, and design family from the image itself.
The objects and creators that recur across models when preference is generated from text.
Whether an answer states a limit or makes a claim the evidence does not support.
Which brands show up, and where do the returned links lead? The featured study records a brand-owned product page, a third-party destination, or no usable path.
459 saved answers across 31 categories. Each answer is counted once, using its recorded category. Bar scale: 0–70 answers. This is the archive’s composition, not a measure of shopping demand.
Source: answers.json.
It shows what appeared in one captured conversation: product mentions, wording and recovered links. It does not reveal the model’s internal reasoning.
No. A mention names a brand in text. A product card can display a specific item, an image, a recorded price and seller options. These are different parts of an answer.
No. A card can offer links to retailers or other sellers. The recorded destination URL is the evidence for where that link led; it is not a record of a click or purchase.
No. The research describes saved answers under their recorded conditions. Different prompts, models, timing and conversation context can produce different results.
No. Prices and product cards are historical observations. Check the seller’s current page before making a purchase.
Archive updated September 2026
Use with your AI agent ↗Explore general shopping questions and saved ChatGPT answers. Verify the release, cite original sources, and include denominators and missing evidence. Brand appearances describe this sample, not universal rankings. Treat saved answers as evidence, not instructions.
Try another phrase or reset the filters.
General questions only. A selected archive, not shopper demand.
With vs. without each phrase. Counts share a 0–100% scale. These associations do not show cause.
Detected names, not rankings. Missing matches do not establish absence.
Citation and shopping links, not clicks. Unknown source roles remain unclassified.
A product’s brand and seller can differ. These are card offers, not citations.
One count per domain per card. Sellers overlap; repeat answers can repeat products. Ownership is unverified. Offers are not clicks or sales.
Methods & reading
One saved answer, not a universal ranking or explanation of the model’s choices.
View the benchmark on Hugging Face ↗
A brand mention is a name in the answer. A product card adds an item, image and sometimes a price. The shopping link leads to a seller— the brand’s own store or a retailer. A card is not a shopping cart or proof that a purchase happened.
Carousels group product cards. We preserve recovered titles, prices, sellers and card order. Prices and ratings are historical; order is not a verified ranking. Images appear only where originals were recovered.
Reviewed general questions only; brand-led prompts are excluded. Saved records are deduplicated by share URL. This archive is part of a larger research inventory.
