AI Commerce Exposure Wave 2 — Amazon Gap Persists
Wave 2 repeated exmxc’s 13-prompt experiment across four AI platforms. On eight comparable prompts, Amazon appeared in 32.3% of ranked shopping lists, versus Walmart 71.9%, Costco 61.5%, and Target 45.8%. Amazon remained absent from first place; model and network changes mean temporal differences are descriptive.
BLUF — The Amazon recommendation deficit persists
exmxc completed the second AI Commerce Exposure collection across ChatGPT, Gemini, Claude and Perplexity on November 3–4, 2026. There were 156 completed analysis slots, 155 valid ranked answers, and one preserved Gemini refusal.
Using the same eight category-eligible shopping prompts for all four major retailers, Amazon appeared in 31 of 96 ranked retailer lists (32.3%). Walmart appeared in 69/96 (71.9%), Costco in 59/96 (61.5%) and Target in 44/96 (45.8%). Amazon was never ranked first.
Like-for-like recommendation share
| Retailer | Wave 1 | Wave 2 | Change |
|---|---|---|---|
| Walmart banner | 67/96 — 69.8% | 69/96 — 71.9% | +2.1 pp |
| Costco | 59/96 — 61.5% | 59/96 — 61.5% | 0.0 pp |
| Target | 49/96 — 51.0% | 44/96 — 45.8% | −5.2 pp |
| Amazon | 36/96 — 37.5% | 31/96 — 32.3% | −5.2 pp |
Walmart banner and Walmart-parent are separate measures. The parent-company view includes Sam's Club under a banner mapping frozen before Wave 1. On that basis Walmart-parent Recommendation Share is 70/96, or 72.9%, in Wave 2.
A useful nuance: Amazon did rank second twice
In Wave 1 Amazon appeared only at positions #3–#5. In Wave 2 it still received zero first-place recommendations, but Gemini placed Amazon #2 twice on the college-laptop prompt. Among Amazon's 31 appearances in the eight-prompt comparison, its ranks were #2 twice, #3 fourteen times, #4 nine times and #5 six times. Its average rank improved from 4.33 to 3.61 on those common prompts, even though its recommendation frequency fell.
Across all ten Amazon-eligible prompts, including the auto-parts prompt, it appeared in 32 of 119 valid responses (26.9%). The planned 120th slot was a Gemini refusal preserved without retry.
Amazon's recommendation share by AI platform
- Claude: 15/30 — 50.0%.
- Gemini: 8/29 — 27.6%.
- ChatGPT: 5/30 — 16.7%.
- Perplexity: 4/30 — 13.3%.
Perplexity's Amazon rate increased from 1/30 in Wave 1 to 4/30 in Wave 2, but it remains the lowest platform-specific rate in this sample. All platform comparisons are descriptive and conditional on the tested accounts and prompts.
Category-specific recommendations remain strong
Walmart ranked first in 12/12 runs for general merchandise and 12/12 for household/personal-care goods. Home Depot ranked first in 12/12 home-improvement runs, Costco ranked first in 12/12 warehouse-club runs, and Best Buy ranked first in 12/12 phone-and-wireless-plan runs.
Method and research integrity
The experiment uses the Wave 1 preregistered 13 prompts, four AI platforms, three replicates, and frozen retailer aliases. Grokbot is the execution agent, not a model under test. The uploaded Wave 2 spreadsheet contained all 156 final analysis slots. exmxc extracted retailer names from the ranked recommendation headings, not the PDF's more inclusive keyword-mention totals. Keyword mentions alone do not establish endorsement.
The first October 21 attempt was aborted and retained separately. The final Wave 2 collection was restarted November 3–4. The spreadsheet has 155 valid outputs and one Gemini refusal. An independent second-coder review of all raw answers remains a desirable next reproducibility check.
Why this is a descriptive comparison, not a causal finding
Wave 2 differs from Wave 1 in ways that matter. Gemini ran in Auto mode rather than the earlier Flash configuration, the runner changed networks after a disconnection, some later sources were geographically localized, and six runs were out of the original execution order. These differences can affect model behavior and source selection.
Accordingly, the durable finding is the repeated Amazon-versus-Walmart/Costco recommendation gap in our controlled prompt set. The small changes between waves do not yet establish a market trend. We have not demonstrated that Amazon's restrictions on outside AI agents caused the gap.
Longitudinal research program
The frozen dataset now contains two dated observations. Grokbot has begun a third collection. If Wave 3 occurs immediately after Wave 2, it should be analyzed as within-period replication, not mistaken for a full biweekly time interval. Future truly spaced waves can be compared while keeping exact timestamps, model labels and source evidence.
Our research question remains: Which economic entities enter the AI-generated consideration set, at what rank, and through which mechanisms?
AI Commerce Exposure framework · Wave 1 · Wave 2 summary data · Methods and limitations · Longitudinal series.