PurchaseDecision

Problem

Brands are being told to spend on “AI visibility” without evidence that it changes what anyone buys. The research that could settle it has to answer two different questions. The first is descriptive: when in-market shoppers turn to an AI assistant, what do they ask, which brands come back, and what do they do with the answer? The second is causal: does appearing in an AI answer or a search result change the purchase, compared with an otherwise identical shopper who never saw it? Traditional research answers each slowly and separately. Surveys capture what people say rather than what they do, and the old forced-exposure studies ran on bolt-on rules that broke whenever a study was copied.

Approach

Built PurchaseDecision end-to-end with Claude Code as one platform with two products that share the same participant engine: panels, screeners, consent, quotas, quality rules, five languages, and redirects.

Observational studies record how people actually use AI to shop. Recruited, in-market participants hold a real conversation with ChatGPT, Claude, or Gemini inside a research replica of each app, backed live by the OpenAI, Anthropic, and Gemini APIs with web search turned on. Each participant is routed to the assistant they already use. Every turn logs the provider, model, latency, and tokens, so a “which brands do assistants recommend” finding can be reproduced. The engine also covers assistants with no public API, such as the shopping assistants in Amazon’s and Walmart’s apps, and people’s own ChatGPT, Claude, or Gemini accounts. For those, the participant shares a screen recording or browser window. Frames are extracted on their device and the video never leaves it. A vision model then rebuilds the transcript. Transcripts are coded against an operator-approved codebook that is checked for reliability, de-identified before any client sees them, and quoted in the report only verbatim, linked back to the exact words in the conversation.

Experimental studies measure cause and effect. Participants answer an identical battery of measures before and after a single controlled exposure to a realistic Google results page (with or without an AI Overview or ads), a ChatGPT experience, or Google AI Mode. They are randomized into test and control cells, and each picks the search most relevant to them, so the exposure stays realistic. Telemetry captures what they clicked and how long they stayed. Each session ends with an observed purchase decision: the participant uploads a photo or screenshot of what they bought, which an AI checks against their stated brand, product, and retailer. Analysis runs in a separate Python stats service rather than being hand-rolled in TypeScript. It computes lifts, significance tests, logistic regression, and CEM or IPW matching, and the report model places numbers but never writes them.

Around both sits the same automated business layer, adapted from Why Not Buy. An intake chat designs either study type and quotes it, then payment, fielding, analysis, reporting, and an “Ask the data” chat follow. Clients can also query their studies over MCP.

Outcome

Live at purchasedecision.ai. The experimental side productizes the method behind the 604-buyer controlled experiment on whether AI visibility moves purchase. The observational side runs live conversations across all three major assistants plus the retailer apps. Every signed-in client can browse a full example study: the design, what participants saw, the report, the conversations, and Ask the data. The two products are built to sell as one engagement. Phase one observes how shoppers use AI in a category. The questions they asked become the topics for phase two, and real assistant answers become draft stimuli for a controlled test of what actually moves the purchase.