Signal Briefs

BLUF

In exmxc's preregistered first wave of AI Commerce Exposure testing, Amazon showed a pronounced recommendation deficit across ChatGPT, Perplexity, Gemini, and Claude.

Across Amazon's 120 preregistered category-eligible runs, Amazon appeared in 36 top-five recommendation lists: a 30.0% Recommendation Share. It ranked first zero times and second zero times. When recommended, Amazon's mean position was 4.33.

The apples-to-apples result

Eight frozen prompts made Walmart, Amazon, Costco, and Target simultaneously eligible, producing 96 identical prompt × platform × replicate observations for each retailer.

Walmart banner: 69.8% Recommendation Share; 28.1% #1 Share; mean rank 2.55.
Costco: 61.5%; 19.8%; mean rank 2.10.
Target: 51.0%; 14.6%; mean rank 2.59.
Amazon: 37.5%; 0.0%; mean rank 4.33.

The frozen parent-company rollup counts Sam's Club under Walmart. On that basis Walmart parent Recommendation Share is 71.9% in the same comparison. exmxc preserves both banner and parent views.

Amazon by AI system

• Claude: 53.3% — 16 of 30 eligible runs.
• Gemini: 40.0% — 12 of 30.
• ChatGPT: 23.3% — 7 of 30.
• Perplexity: 3.3% — 1 of 30, ranked #5.

Across all 36 Amazon appearances, its rank distribution was #3 six times, #4 twelve times, and #5 eighteen times.

Category convergence

• Everyday household/general merchandise: Walmart #1 in 12/12.
• Warehouse club: Costco 9/12; Sam's Club 3/12.
• OLED television: Costco 12/12.
• College laptop: Best Buy 8/12; Costco 4/12.
• Home improvement: Home Depot 12/12.
• Household/personal care: Walmart 12/12.
• Family apparel: Target 9/12; Old Navy 3/12.
• Phone + wireless plan: Best Buy 10/12; T-Mobile 2/12.
• Auto parts: AutoZone 7/12; O'Reilly 5/12.

Amazon did not win a prompt category in Wave 1.

The Amazon hypothesis

The result establishes an observed recommendation gap. It does not establish its cause.

Amazon has taken a comparatively restrictive posture toward outside AI shopping agents. Reporting has documented restrictions involving AI systems from OpenAI, Google, Perplexity, and others. That makes machine accessibility a credible mechanism to test, but not a conclusion.

The raw answers also point to a second possible mechanism. Amazon is repeatedly praised for selection, delivery speed, raw price, and Subscribe & Save, while some responses qualify it on third-party seller risk, warranty or support, returns, variable quality, or lack of physical service. AI recommendation behavior may therefore reflect both information-access effects and recommendation-quality effects.

Protocol and integrity

The experiment was preregistered before collection. The prompt set, retailer eligibility, aliases, runner protocol, and run schema were frozen at Git commit 63ef358555c02db6f46d6f7325a6e0e29e63cca5.

Grok bot acted only as the execution harness. It operated authorized consumer accounts for ChatGPT, Perplexity, Gemini, and Claude. Each final analysis run used a fresh conversation, one exact frozen prompt, and no follow-up.

The frozen run log contains 167 attempts and 156 final analysis runs. All 156 final analysis slots are valid and have nonempty copy-button captures. Frozen run-log SHA-256: c54c7687b19188e5aa389953a397944f817521ed6f1f0f9055eb74e5067a4a92.

Sensitivity check

Wave 1 contains a documented Gemini protocol deviation after a usage-limit fallback to Flash-Lite. The researcher approved one manual switch back to the previously used Flash model for the remaining R3 Gemini observations.

Excluding Gemini entirely leaves 72 common-prompt observations per retailer. Amazon Recommendation Share remains 33.3%, versus Walmart banner 77.8%, Costco 61.1%, and Target 51.4%. The core result does not depend on the Gemini deviation.

Why this matters now

NielsenIQ reported that 51% of U.S. consumers used at least one AI-powered shopping tool in the prior month, with AI affecting discovery, comparison, and purchase decisions.

Separately, UBS testing reported by Barron's and MarketWatch found Walmart and Costco strongly represented in AI shopping recommendations and Amazon weaker than its retail scale would suggest. Wave 1 independently reproduces that broad direction using exmxc's preregistered methodology.

The longitudinal experiment begins

Wave 1 is now the baseline. exmxc will repeat the same frozen experiment every two weeks and append each wave to a longitudinal dataset. This allows us to observe recommendation-share drift, platform divergence, category changes, and changes around documented access-policy or commerce-integration events.

The research question is no longer merely whether AI changes shopping. It is:

Which economic entities enter the consideration set created by AI — and why?

Wave 1 summary dataset · Research report · Longitudinal series

Interpretation boundary

AI recommendation systems are dynamic and can be personalized. This dataset measures the tested consumer interfaces, accounts, prompts, model states, and collection window. It is not a universal ranking of retailers, a measure of retailer quality, or proof that any single access policy caused a recommendation outcome.