# CarPixAI 2026 VDP Photo Benchmark v1.1 — Multi-Platform Replication Methodology

Release: `v1.1-replication-2026-07-28`
Frozen rubric: `v1.0-2026-07-28`
Collection date: `2026-07-28`
Fixed source frame: `122 public used/CPO VDP URLs`
Included replication sample: `111 VDPs`
Excluded without replacement: `11 VDPs`

## Status and relationship to v1

This is a separate observational replication of the frozen 108-VDP v1 benchmark. It does not revise, replace, pool, or selectively rescore the v1 records.

The replication was launched after a previously dispatched source-research batch returned a substantially more diverse platform pool. Keeping it separate prevents post-hoc replacement of the original sample after its results were known.

Both releases are purposive convenience samples. Neither is a probability sample or a statistically representative study of US dealerships, dealer website platforms, or used-vehicle inventory.

## Research question

> When the frozen 12-point VDP photo rubric is applied to a separate, more platform-diverse fixed source frame, what observable photo-merchandising patterns appear, and which descriptive findings remain similar or change?

The replication does not test sales outcomes, dealer quality, platform quality, vehicle quality, editing provenance, or causality.

## Source-frame construction

Three independent sourcing lanes were completed in parallel before scoring:

1. 40 live public used-vehicle VDPs from five dealer-owned independent-dealer domains.
2. 40 live public used/CPO VDPs from eight franchise-dealer domains.
3. 42 live dealer-owned used-vehicle VDPs from six DealerOn domains discovered through public sitemap conventions.

The combined fixed frame contained:

- 122 unique VDP URLs;
- 19 dealer domains;
- 11 states;
- 40 independent-dealer rows and 82 franchise rows;
- seven normalized website-platform families:
  - AutoManager;
  - CarsForSale.com;
  - DealerCenter;
  - Dealer.com;
  - Dealer Inspire/Cars Commerce;
  - DealerOn;
  - Overfuel.

Every candidate was verified as a live individual vehicle page by the sourcing lane before the frame was frozen. The private source-pool CSV retains source URL, domain, state, platform class, segment, verification timestamp, and sourcing-batch provenance.

The frame is purposive rather than random. Domain counts and platform counts reflect what the fixed research lanes found and verified, not platform market share.

## Unit of analysis

One public dealer-owned used or certified-pre-owned vehicle-detail page as observed on the collection date.

The benchmark evaluates the photo gallery attached to that VDP. It does not evaluate the dealership or vehicle.

## Access and robots handling

Before processing a domain, the collector requested its public `robots.txt` using an identified research user agent.

- A parsed rule disallowing the actual VDP path caused exclusion.
- Declared crawl delays were honored, capped at 10 seconds per sequential request lane.
- When a direct robots request failed, the collector attempted a renderer-backed retrieval.
- In accordance with RFC 9309 section 2.3.1.2, an HTTP 4xx robots response was recorded as unavailable rather than as a prohibition.
- Network, TLS, renderer, or 5xx failures that left robots policy unreachable failed closed.
- Disallowed APIs, feeds, AJAX inventory endpoints, and lead/customer forms were not used.

Ten fixed-frame VDPs from two domains were excluded because robots policy remained unreachable both directly and through the renderer. They were not replaced.

## Inclusion criteria

A source-frame row was included when:

1. it remained retrievable as a public dealer-owned used/CPO VDP;
2. the VDP path was not disallowed by applicable robots rules;
3. the collector extracted at least three unique vehicle-image candidates;
4. at least three representative images could be downloaded and decoded;
5. the frozen scoring model returned valid scores for all five visual dimensions.

## Exclusion criteria

A fixed-frame row was excluded, without replacement, for:

- robots disallowance or robots policy remaining unreachable;
- removed or inaccessible listing;
- fewer than three extractable vehicle images;
- fewer than three decodable representative images;
- page extraction failure;
- invalid model response after retry.

Final exclusions:

- 10 `robots_disallowed_or_unavailable`;
- 1 `insufficient_extractable_images`.

The final replication therefore contains 111 of the 122 frozen candidates across 17 domains, all 11 source-frame states, and all seven platform families.

## Multi-platform image extraction

The collector decoded image URLs from page HTML and used platform-aware evidence to isolate the vehicle gallery:

- VIN-linked image URLs where at least three were present;
- recognized dealer-inventory image hosts and paths;
- `image_src` or Open Graph hero-image metadata;
- same-gallery URL grouping;
- thumbnail, icon, logo, tracking, stock-placeholder, and resize-variant filtering;
- unique normalized source-image identities.

The published hero was moved to position one when a matching metadata image was found. Representative images were then selected from the beginning, middle, and end of the detected gallery.

Image extraction differs across website platforms. A detected gallery count can omit lazy-loaded assets or retain platform-specific duplicates. Gallery-count comparisons across platforms must therefore be treated as pipeline measurements, not perfect counts of every image visible to every browser session.

## Frozen 12-point rubric

The v1 rubric was reused without changing thresholds after the v1 results were known:

| Dimension | Method | Points |
|---|---|---:|
| First hero image | Model-assisted visual classification | 0–2 |
| Gallery completeness | Deterministic detected-image count | 0–2 |
| Condition proof | Model-assisted representative-image classification | 0–2 |
| Background consistency | Model-assisted visual classification | 0–2 |
| Mobile crop safety | Model-assisted hero classification | 0–2 |
| Visible edit safety | Model-assisted warning-sign classification | 0–2 |

Gallery completeness remained:

- 0 points: fewer than 12 detected unique vehicle images;
- 1 point: 12–19;
- 2 points: at least 20.

The complete definitions remain in `used-car-photo-benchmark-2026-rubric.csv`.

## Model-assisted review

Primary visual reviewer: `google/gemini-2.5-flash` through OpenRouter.
Temperature: `0`.
Representative images: up to eight per VDP, sampled from beginning, middle, and end.
Image normalization: auto-rotated JPEG, maximum 1024 × 1024, quality 72.

Model-assisted scores are classifications, not human findings or ground truth. The visible-edit dimension identifies visible warning signs only; it cannot determine whether an image was edited or whether every depicted detail matches an unavailable source photograph.

## Platform-stratified QA

A second model, `openai/gpt-4.1-mini`, independently rescored two rows from each of the seven included platform families:

- QA rows: 14;
- visual dimensions per row: 5;
- total comparisons: 70;
- exact agreement: 38.6%;
- within one point on the 0–2 scale: 90.0%;
- mean absolute difference: 0.71 points.

The low exact agreement is material. Visual percentages should be read as model-sensitive directional observations. The QA is a sensitivity analysis, not human inter-rater validation, and neither model is treated as ground truth.

## Missing-data handling

No score was imputed. A candidate that failed inclusion remained an excluded source-frame row. Unlike v1, no replacement URL was selected because the replication frame was frozen before scoring.

## Anonymization and publication

Public row-level files retain:

- anonymous sample and dealer IDs;
- source-pool ID;
- SHA-256 source URL hash;
- dealer segment and broad region;
- normalized website-platform family;
- retrieval date;
- body-style classification;
- detected image count and reviewed-image count;
- six dimension scores and total;
- standardized issue/evidence codes;
- concise non-identifying priority fix;
- rubric, release, and review-method versions.

Public files omit:

- dealer name and domain;
- VDP URL;
- VIN and stock number;
- exact state and location;
- source title;
- image URLs and source images;
- license plates, signage, customer information, and raw screenshots.

No dealer photographs are republished. Public examples are derived text scorecards only.

## Descriptive comparison rules

The v1 and v1.1 samples remain separate. The comparison download reports unweighted differences in observed means and full-pass percentages.

Do not:

- pool the row sets without publishing a weighting plan;
- call percentage differences platform effects;
- infer causality;
- generalize either sample to the dealership industry;
- use the results to rank named dealers or platforms.

Dealer composition, vehicle composition, platform composition, source-frame construction, and extraction behavior all differ between releases.

## Known limitations

1. Purposive convenience sampling; no known selection probability.
2. Platform and domain quotas are not market-share weighted.
3. Franchise rows outnumber independent rows in the fixed frame.
4. One platform family contributes 41 of the 111 included rows.
5. Gallery extraction varies by platform and rendering behavior.
6. Only representative images received multimodal review.
7. Model scores are reviewer-sensitive; exact second-model agreement was 38.6%.
8. No human inter-rater validation was completed.
9. No original-versus-edited image pairs were available.
10. Listings are volatile and may change or disappear.
11. No customer, traffic, lead, sales, pricing, or time-to-sale outcome was measured.
12. Cross-release differences are descriptive and confounded by composition.

## File lineage

Private/internal:

- `v1.1-replication-source-pool.csv` — frozen 122-row source frame;
- `v1.1-replication/source-manifest.csv` — 111 included source records and image provenance;
- `v1.1-replication/collection-failures.json` — 11 fixed-frame exclusions;
- `v1.1-replication/model-review-records.json` — primary review records;
- `v1.1-replication/qa-review-records.json` — second-model QA with private image provenance;
- `tmp/vdp-photo-benchmark-v1-1-work.json` — ignored resumable local work state.

Public:

- `used-car-photo-benchmark-2026-v1-1.csv`;
- `used-car-photo-benchmark-2026-v1-1.json`;
- `used-car-photo-benchmark-2026-v1-1-summary.json`;
- `used-car-photo-benchmark-2026-v1-1-qa.json`;
- `used-car-photo-benchmark-2026-v1-1-methodology.md`;
- `used-car-photo-benchmark-2026-v1-1-data-dictionary.csv`;
- `used-car-photo-benchmark-2026-v1-vs-v1-1.json`.

Executable scripts:

- `scripts/import-vdp-replication-source-pool.mjs`;
- `scripts/build-vdp-photo-benchmark-v1-1.mjs`;
- `scripts/qa-vdp-photo-benchmark-v1-1.mjs`;
- `scripts/build-vdp-benchmark-comparison.mjs`.

## Citation guidance

Appropriate:

> CarPixAI’s separate 2026 multi-platform replication included 111 public used/CPO VDPs from a fixed 122-URL purposive frame across seven website-platform families. It used the frozen v1 rubric and reported model sensitivity separately.

Inappropriate:

> CarPixAI proved that a stated percentage of all US dealerships follows or fails a practice.

The second statement is unsupported because this is not a representative industry sample.