AI Commerce Exposure Wave 3 — Same-Day Replication
exmxc's 156-run same-period Wave 3 replication finds Amazon in 28.1% of comparable AI-ranked retailer recommendations, versus Walmart 70.8%, Costco 63.5%, and Target 50.0%. Amazon remained absent from first place. Wave 3 was collected on November 4, immediately after Wave 2, so it tests repeatability rather than a new biweekly time trend.
BLUF: Amazon's recommendation gap survives a third test
On November 4, 2026, the exmxc AI Commerce Test Runner completed a 156-for-156 valid third wave across ChatGPT, Gemini, Claude, and Perplexity. Grokbot preserved three failed or blocked first attempts separately. All 156 final answer-text hashes checked against the submitted workbook, and the prompt texts matched those used in Wave 2.
In the eight frozen shopping prompts where Walmart, Amazon, Costco, and Target were all eligible, Amazon appeared in only 27 of 96 ranked lists (28.1%). Walmart banner appeared in 68/96 (70.8%), Costco in 61/96 (63.5%), and Target in 48/96 (50.0%).
Three-wave comparison
| Retailer | Wave 1 | Wave 2 | Wave 3 |
|---|---|---|---|
| Walmart banner | 69.8% | 71.9% | 70.8% |
| Costco | 61.5% | 61.5% | 63.5% |
| Target | 51.0% | 45.8% | 50.0% |
| Amazon | 37.5% | 32.3% | 28.1% |
Every percentage in this table uses an identical 96-observation denominator per retailer, per wave. Walmart banner is distinct from Walmart parent, which also includes Sam's Club under frozen entity mappings.
Amazon is not absent; it is ranked lower
Amazon received zero first-place rankings in Wave 3. It was ranked second twice (both Gemini's college-laptop prompt), third 12 times, fourth eight times, and fifth five times. Its mean position when recommended was 3.59, compared with 3.61 in Wave 2 and 4.33 in Wave 1.
Across Amazon's complete 120 category-eligible Wave 3 runs, it appeared in 27, or 22.5%. The Wave 2 equivalent was 32/119 (26.9%; one Gemini refusal), and Wave 1 was 36/120 (30.0%). These movements remain descriptive.
Amazon recommendation share by system
- Claude: 13/30 = 43.3%.
- Gemini: 7/30 = 23.3%.
- ChatGPT: 5/30 = 16.7%.
- Perplexity: 2/30 = 6.7%.
The most retrieval-oriented platform in our tested set continues to surface Amazon at a relatively low rate, but that observation does not prove a crawler-access mechanism.
Category leaders remain highly consistent
In Wave 3, Walmart ranked first for general merchandise in 12/12 runs and for household/personal-care goods in 12/12. Costco ranked first for warehouse clubs in 12/12 and OLED televisions in 12/12. Home Depot ranked first for home improvement in 12/12; Best Buy led laptops in 10/12; and AutoZone led auto parts in 11/12.
Why Wave 3 is a replication, not another two-week time point
Wave 1 was collected October 7. Wave 2 was delayed from its original October 21 attempt and completed November 3–4. Wave 3 was collected November 4—the same day Wave 2 finished. Wave 3 therefore tests short-horizon repeatability, not change over a full biweekly interval. We preserve the actual dates and classify this observation separately in the longitudinal index.
Comparability also requires care. All Gemini runs used Auto (underlying submodel may vary), and the runner used its own network, which exposed occasional Colorado/Utah-localized sources. Usage limits, approved retries, and four out-of-sequence make-up runs were logged. ChatGPT showed unlabeled Shop product cards, and one Perplexity source carried an ad-related query parameter without a visible Sponsored label. These were not counted as ranked recommendations unless the model's own ranked response named the retailer.
Methodology, verification, and limitations
Grokbot's summary PDF provides an alias-mention frequency table, not recommendation-share coding. Our exmxc analysis separately extracted the ranked retailer names from all 156 final answers, ignored incidental narrative/citation mentions, preserved original banner names, and applied the same frozen category eligibility and ranking method as Wave 2.
Individual raw-answer hashes and submitted prompts verified from the supplied spreadsheet. The separate raw ZIP archive and manifest described by Grokbot were not attached for independent archive-level validation. An additional independent second-coder adjudication remains desirable for formal inferential publication.
No causal conclusion is established. The study cannot yet separate Amazon's external machine-access policies from retailer reputation, return policies, category differences, model routing, search results, or account/location effects.
Source records
AI Commerce Exposure framework · Wave 1 baseline · Wave 2 · Wave 3 structured summary · Full research report · Versioned longitudinal index