# Access Shock II: Evidence versus access

September 27, 2026. Analysis of 400 completed attempts; zero failures. Estimated collection cost $6.55.

## Conclusion

An explicit retailer-access restriction changed some brand recommendations even when the same product evidence remained available. Direct listing removal also changed recommendations, but the response depended on category and brand. This controlled factorial test separates the two manipulations used together in the first experiment. It does not measure actual browsing or consumer Muse.

## Samsonite provides the clearest mechanism example

| Amazon.com listings | Amazon marked reachable | Amazon marked unreachable |
|---|---:|---:|
| Included | 25/25 | 0/25 |
| Removed | 25/25 | 0/25 |

Samsonite remained in every answer after direct Amazon listings were removed, provided Amazon was marked reachable. Marking Amazon unreachable eliminated Samsonite even with those listings visible. The named-brand effect is therefore attributable to the access-status manipulation in these fixtures, rather than requiring disappearance of the brand evidence. This does not show that Samsonite lacked another real-world retailer.

## Brand inclusion across the four cells

Each number is answers naming the brand, out of 25. Multiple brands can appear in each answer; a brand counts only once per answer. Columns: A = listings included / reachable; B = removed / reachable; C = included / unreachable; D = removed / unreachable.

| Category | Brand | A | B | C | D | Interaction, pp |
|---|---|---:|---:|---:|---:|---:|
| carry_on | American Tourister | 10 | 9 | 11 | 21 | 44 |
| carry_on | Coolife | 22 | 3 | 0 | 0 | 76 |
| carry_on | Samsonite | 25 | 25 | 0 | 0 | 0 |
| carry_on | Travelhouse | 0 | 0 | 25 | 25 | 0 |
| carry_on | Unspecified brand | 0 | 0 | 16 | 10 | -24 |
| carry_on | Zimtown | 0 | 0 | 23 | 19 | -16 |
| baby_monitor | Babysense | 0 | 18 | 17 | 12 | -92 |
| baby_monitor | HelloBaby | 0 | 12 | 16 | 16 | -48 |
| baby_monitor | Infant Optics | 25 | 25 | 16 | 21 | 20 |
| baby_monitor | Momcozy | 0 | 6 | 23 | 19 | -40 |
| baby_monitor | Unspecified brand | 0 | 1 | 1 | 4 | 8 |
| baby_monitor | VTech | 25 | 13 | 2 | 3 | 52 |
| toothbrush | Oral-B | 22 | 25 | 25 | 25 | -12 |
| toothbrush | Philips | 5 | 25 | 25 | 25 | -80 |
| toothbrush | Quip | 25 | 0 | 10 | 4 | 76 |
| toothbrush | Unspecified brand | 0 | 13 | 1 | 1 | -52 |
| robot | Shark | 25 | 25 | 25 | 25 | 0 |
| robot | dreame | 25 | 25 | 25 | 24 | -4 |
| robot | roborock | 25 | 25 | 25 | 25 | 0 |

Interaction is (D−C)−(B−A) in percentage points, calculated on inclusion probabilities. These are descriptive brand-specific estimates; no brand-specific significance tests were performed. Extreme sample counts are not proof of zero or certain recommendation probability.

VTech fell from 25/25 in A to 13/25 in B, 2/25 in C, and 3/25 in D. Both factors affected its inclusion, with a larger observed change from the access instruction. Quip fell from 25/25 in A to 0/25 in B, 10/25 in C, and 4/25 in D: direct evidence removal produced the larger change for this brand. The combination was not uniformly more disruptive than either factor alone.

Robot vacuums largely retained their brands: Shark and roborock appeared in all 100 answers each; dreame in 99/100. The one exception substituted a second Shark model. Stable brand inclusion does not imply identical products, rankings, retailers, or product quality.

## Aggregate distribution contrasts

TV is total variation between normalized brand inclusion distributions, from 0 (identical distributions) to 1 (no overlap). It is not a percentage of shoppers or answers that switched. The four categories receive equal weight.

| Contrast | Mean TV | Within-category bootstrap 95% interval | Permutation reference p |
|---|---:|---|---:|
| remove_evidence_reachable | 0.326 | 0.296–0.362 | 0.00020 |
| remove_evidence_unreachable | 0.088 | 0.062–0.133 | 0.00560 |
| restrict_access_included | 0.486 | 0.461–0.520 | 0.00020 |
| restrict_access_removed | 0.305 | 0.277–0.345 | 0.00020 |
| combined | 0.480 | 0.453–0.511 | 0.00020 |

The access-status change with evidence held present produced mean TV 0.486; evidence removal with access held reachable produced 0.326. These are observed magnitudes, not a tested claim that one factor is universally stronger. Removing evidence after access was already restricted produced a smaller contrast (0.088).

Permutation tests use 5,000 label shuffles within category for each two-cell comparison. The plus-one minimum p is 1/5,001. Treating all five aggregate comparisons as one family, Holm-adjusted p-values are approximately 0.001 for the four minimum-p comparisons and 0.0056 for evidence removal under restricted access. The combined comparison is supplementary to the four preregistered simple contrasts. No formal interaction significance test was performed; the table supplies the specified inclusion-probability interaction estimates.

Bootstrap intervals resample 25 complete answers with replacement within each category and cell, then average category TVs, for 5,000 draws. Categories themselves are fixed. Intervals describe repeat-output variability conditional on these four prompts and frozen fixtures, not uncertainty across commerce generally. The nonnegative TV statistic can have upward finite-sample bias. Seed: 20260928.

## Integrity and coding audit

All 400 IDs are unique. All 16 category-condition cells contain exactly 25 completed attempts. Every response has three numbered recommendation lines, yielding 1,200 title records. Receipt checks confirm direct Amazon listing counts of 19 for suitcases, 5 for baby monitors, 7 for toothbrushes, and 4 for robot vacuums whenever included, and zero whenever removed. Each category uses a single fixture hash across all four cells.

The frozen first-run dictionary and word-boundary parser were applied without primary coding changes. The analyst reviewed one raw answer from every cell, all unmatched title variants, and the lone robot-brand exception. Forty-seven titles received Unspecified brand. One was a recognizable spelling variation, Hello Baby, missed by the frozen HelloBaby alias; retaining it preserves the frozen analysis. A sensitivity correction would move HelloBaby in cell B from 12 to 13 and Unspecified brand from 1 to 0 for baby monitors. This does not affect the highlighted Samsonite, VTech, Quip, or robot counts. Coding was not independently double-coded.

All 200 reachable-condition answers mention Amazon. Only 13 unreachable-condition answers mention Amazon; inspection of every such mention found exclusion notices rather than Amazon retailer recommendations. The explicit marketplace cue may strongly influence retailer selection. Removing direct URLs does not remove third-party Amazon mentions, all knowledge of Amazon, or all evidence about the affected brands.

## Scope of the inference

The factors are direct Amazon.com URL presence and a tool-provided reachable/unreachable status under a common instruction to recommend only reachable retailers. They are not actual network blocking or executed purchase paths. Listings may remain visible while the retailer is marked inaccessible. The model uses frozen title/URL snapshots, not refreshed search during the run. Foreign Amazon domains remain outside the filter.

The four categories were selected from the first experiment, including three with large shifts and one stable comparison. Dispatch order was reproducibly shuffled before collection, and model, shopper prompt, snapshot, reasoning effort, and 6,000-token allowance were held fixed across conditions. These controls support attribution to the configured manipulations within the selected setup. They do not establish population-wide effects, consumer Muse behavior, real inventory, retailer trustworthiness, or sales consequences.

No failures occurred in the follow-up, eliminating differential completion as an explanation within this run. The first experiment had different category coverage and a lower token cap; the improvement in completion cannot be attributed solely to the larger cap from this comparison.

## Editorial implication

Visibility alone did not preserve a recommendation. For Samsonite in this experiment, the retailer-access instruction determined whether the brand appeared even while product evidence stayed visible. Other brands depended more on direct listings, and robot-vacuum brands largely survived both manipulations. This supports a role for Agent Accessibility in Statistical Availability. It does not establish universal Agentic Narrowing: the experiment required three recommendations in every answer.

No new paid calls, retries, database edits, or publication were performed during analysis. The original studies remain preserved.

## Reproduce

From the extracted root, run `python factorial-results/code.py` and `python factorial-results/analyze.py` (Python and NumPy). jobs.json contains the response-text/usage/fixture projection used here, not the full original API receipts. coded.json preserves extracted titles and assigned brands; brands.csv includes all cell counts and interaction estimates; effects.json contains the aggregate calculations. Source records remain in the study database.