Used-Car VDP Photo Benchmark: Data + Replication

Quick answer: On September 15, 2026, CarPixAI tested its current Clean Dealer Studio workflow on 74 used-vehicle source photos. The dedicated Integrity Review agent completed 74 of 74 evaluations (100% coverage) and recorded 72 full passes (97.3%). A full pass means the vehicle material was preserved, the scene stayed consistent, and the result was postable. This product qualification is not pooled with the dealer-VDP observational studies below.

CarPixAI workflow qualification (tested September 15, 2026)

CarPixAI tested its current image-processing workflow on 74 used-vehicle source photos using Clean Dealer Studio results. 71 are untouched first-pass holdout outputs, 1 is a same-vehicle Standard reprocess, and 2 are higher-resolution replacement vehicles used after the original dealer files proved too small and the listings were no longer available. Generation completed for all 74 vehicles with 0 paid-generation failures. Test date September 15, 2026. This qualification is separate from the public-VDP observational studies below and is not pooled with them.

How to read this result: The full-pass rate is 97.3% among the 74 valid evaluations produced by CarPixAI's dedicated Integrity Review agent on September 15, 2026. A full pass means the vehicle material was preserved, the scene stayed consistent, and the result was postable. The 2 flagged results are not all vehicle-integrity failures: 1 confirmed vehicle-material exception — source-hidden wheel or tire evidence made newly visible. None of the flags are scene-only. 1 returned insufficient candidate evidence for a verdict and is reported as inconclusive rather than failed. Scored by dimension, 72 of 74 passed vehicle-material integrity (97.3%, 90.7%–99.3%) and 74 of 74 passed scene consistency (100%). The 0 unavailable evaluations are reported separately. They are neither passes nor vehicle-integrity failures; counting unavailable evaluations as failures would misstate the pass rate.

No secondary operator verdict is merged into these results. This bounded test does not establish perfect vehicle preservation or automatic publication safety. Source and result images remain private; the public download contains aggregate methodology and results only.

Download the vehicle-integrity aggregate

Observational dealer-VDP study (collected July 28, 2026)

Separately, CarPixAI scored 108 live, public used-vehicle detail pages from 9 dealer websites across 8 states on July 28, 2026. The median listing scored 8out of 12. This is a frozen observational benchmark of dealer listing galleries on that collection date—not a test of CarPixAI's image-processing workflow, not a nationally representative industry study, and not evidence that a photo score causes sales.

A separate multi-platform replication then applied the same frozen rubric to a fixed 122-URL frame. It included 111 VDPs across 17 domains and seven platform families, with a median of 8 out of 12. The two observational releases remain separate from each other and from the product qualification above. Neither VDP release is industry-representative.

The strongest pattern in this sample was gallery quantity: 76.9% of listings exposed at least 20 unique vehicle images. Presentation quality was less consistent. Only 17.6% earned the full two points for the published hero image, while 20.4% earned full marks for repeatable, non-distracting backgrounds.

The public release includes the complete anonymized row-level dataset, aggregate calculations, rubric, data dictionary, methodology, limitations, and reproducible collection/scoring script. Dealer names, domains, listing URLs, VINs, stock numbers, titles, states, and source images are excluded from the public downloads.

Want a decision-focused review of your own set? Upload up to 12 photos to the free Car Listing Photo Quality Checker to review lead-image strength, missing coverage, consistency, reshoot needs, and eligible presentation fixes. Its answer is directional and does not turn these purposive samples into a market-representative score.

Download the anonymized CSV · Download row-level JSON · Download aggregate JSON · Download QA results · Download methodology · Download rubric · Download data dictionary

Observational v1.1 multi-platform replication

After observational v1 was frozen, a previously dispatched sourcing batch returned a broader fixed frame of 122 verified VDP URLs across 19 dealer domains, 11 states, and seven website-platform families. CarPixAI ran that pool as a separate observational replication rather than replacing favorable or unfavorable v1 rows after seeing the original results. It is not a re-run of the CarPixAI workflow qualification above.

The replication included 111 VDPs across 17 domains. 11 fixed-frame candidates were excluded without replacement: 10 because robots policy remained unreachable and 1 because fewer than three vehicle images could be extracted. All seven platform families remained represented.

The replication mean was 0.35 points higher, while both medians were eight. Full-pass rates were higher in the replication for hero image, gallery completeness, condition proof, and background consistency, but lower for mobile crop and visible edit safety. These are unweighted descriptive differences—not platform effects—because dealer, vehicle, platform, and extraction composition all changed.

Replication QA: Two rows from each platform family were independently rescored with a second model. Across 70 comparisons, exact agreement was 38.6%, 90% were within one point, and mean absolute difference was 0.71. This again shows that the visual scores are directional and reviewer-sensitive.

Download v1.1 CSV · Download v1.1 row JSON · Download v1.1 summary · Download v1.1 QA · Download v1.1 methodology · Download v1.1 data dictionary · Download descriptive comparison

What the observational v1 sample contained

Descriptive findings from the sampled VDPs

Across the 108 sampled listings collected on July 28, 2026, the mean score was 7.79 and the median was 8 out of 12. Scores ranged from 3 to 12. These figures describe only this frozen observational sample and collection date, not CarPixAI's image-processing workflow.

Gallery count was the most common full pass: 83 listings had at least 20 unique detected vehicle images. Condition proof was less complete in the representative image sample: 43 listings showed enough breadth for a full score. This does not prove a missing proof image was absent from every unsampled gallery position.

Visual-score sensitivity: A stratified 12-row second-model QA pass rechecked 60 visual-dimension scores. Exact agreement was 40%, 85% were within one point on the 0–2 scale, and mean absolute difference was 0.75. The gallery-count metric is deterministic, but the visual percentages above are model-sensitive directional observations—not ground truth. The disagreement is published rather than hidden.

Most frequent published-hero issues

The explicit 12-point rubric

Every VDP receives zero, one, or two points in six dimensions. A zero indicates missing or unusable observable evidence, one indicates a usable result with material limitations, and two indicates a full pass under the published definition. The rubric was frozen before the full sample was scored.

The visible-edit score is deliberately narrow. A two does not prove an image is unedited, and it does not prove every vehicle detail matches an unavailable source photo. It only means the reviewed images showed no obvious warning sign under this rubric.

Observational VDP methodology

1. Source selection: Nine US dealer-owned websites were purposively selected to cover independent and franchise inventory across several regions. Aggregators and marketplaces were excluded.
2. Robots check: The collection script requested each public robots.txt and continued only when the sitemap and representative VDP path were not disallowed for the general crawler group.
3. VDP frame: Eligible URLs came from public dealer sitemaps, contained a VIN-shaped identifier, and were explicitly marked used or pre-owned in the path.
4. Deterministic selection: Eligible URLs within each domain were ranked using a published SHA-256 rule and a versioned sampling salt. The first eligible rows were processed until the fixed domain quota was met.
5. Inclusion: A page had to be live and yield at least three unique VIN-linked vehicle images. Failed, removed, blocked, placeholder-only, or insufficient-image pages were recorded internally and replaced by the next ranked page from the same domain.
6. Image review: The hero plus images sampled from the beginning, middle, and end of the gallery were normalized to JPEG and reviewed at temperature zero using confidential-primary-reviewer. The number of reviewed images is present in every public row.
7. Deterministic count: Gallery completeness used the number of unique source image paths after thumbnail and resize variants were deduplicated.
8. Missing data: No score was imputed. A source that could not satisfy inclusion criteria was excluded and replaced within its existing domain quota.
9. Anonymization: Public rows retain an anonymous dealer ID, broad segment, region, URL hash, body style, evidence fields, and scores. They omit dealer/customer identities and all source identifiers that would expose a VIN, stock number, page title, URL, state, or image.
10. QA: 12 rows were selected across all four score bands and re-reviewed with confidential-qa-reviewer. Exact and within-one-point agreement are published in the QA download. This is a second-model sensitivity check, not a human reliability study.

The complete executable procedure is in scripts/build-vdp-photo-benchmark.mjs. The downloadable methodology records the sampling salt, thresholds, reviewer configuration, replacement rules, and file lineage.

Anonymized representative examples

Raw dealer photos are not republished. The examples below are de-identified row summaries selected mechanically from the lowest, nearest-to-median, and highest total scores. They illustrate how the rubric behaves without naming or visually exposing a dealership, vehicle listing, customer, VIN, stock number, plate, or location.

Limitations and scope boundaries

• Directional observational sample, not a probability sample of all US dealerships or all live used vehicles.
• Dealer-owned VDPs were sampled from nine public sitemaps across eight states using a deterministic URL hash rank.
• The sample is intentionally balanced between independent and franchise listings, so aggregate percentages are not market-share weighted.
• Most sampled websites used the same inventory website/image-delivery platform, which may cluster gallery and presentation behavior.
• Scores use the published hero plus representative images sampled from the beginning, middle, and end of each gallery; not every image received multimodal review.
• Gallery completeness is based on detected image count; sampled visual evidence may miss a proof photo elsewhere in the full gallery.
• Visible edit safety records observable warning signs only. It cannot prove an image is unedited or that every vehicle detail matches an unavailable source photo.
• Model-assisted scores can contain classification error. A documented stratified second-pass QA review is published separately; no formal human inter-rater study was performed.
• The benchmark does not measure or imply leads, appointments, sales, time-to-sale, price realization, or causal business impact.
• Listings change and may be removed after the collection date.

The sample was balanced for this audit rather than weighted to the actual US dealer population. That makes independent-versus-franchise coverage easier to inspect, but it means the overall percentages should not be projected to the industry. Website-platform concentration may also make gallery behavior more similar than it would be in a wider platform sample.

The v1.1 replication broadens platform coverage, but it is still purposive and unweighted: franchise rows outnumber independent rows, one platform contributes 41 of 111 included VDPs, and gallery extraction behaves differently across platforms. The replication reduces one v1 limitation without creating a representative industry sample.

This work evaluates observable photo merchandising only. It does not evaluate vehicle quality, dealer ethics, pricing, availability, customer experience, lead volume, appointments, sales, days-to-sale, or return on investment. It cannot establish that improving a score changes any commercial outcome.

How dealers can use the framework

Start with the Dealer Inventory Photo Evidence Scorecard, then use the blank dealer photo audit template to apply the same six dimensions to your own inventory. Keep the sample rule stable, preserve the source date, and compare workflow changes over time without turning an internal score into an unsupported sales claim.

For the practical workflow behind the first image, use the VDP hero-image previewer and the first-nine VDP photo checklist. If a source photo is sharp and truthful but the environment is distracting, test a copy with the car background remover. Retake blurry, cropped, stale, or inaccurate source photos instead of trying to repair them with AI.

FAQ

Is the July VDP study a test of CarPixAI's current workflow?

No. The July 28, 2026 v1 sample and the v1.1 replication score live dealer listing galleries with a frozen 12-point rubric. The CarPixAI workflow qualification is a separate product test dated September 15, 2026.

Is this a representative automotive industry study?

No. The frozen v1 contains 108 purposively sampled public VDPs, and the separate v1.1 replication contains 111 of 122 fixed candidates. Neither is a probability sample, and neither should be generalized to every US dealership.

Can the findings be reproduced?

The public methodology, rubric, row-level calculations, URL hashes, sampling rule, and executable script are available. Live listings can change or disappear, and dealer URLs are withheld from public files to avoid naming or ranking stores, so exact visual replay requires the internal provenance manifest.

Did CarPixAI compare edited images with source photos?

No. The visible-edit dimension identifies observable warning signs in published images. It does not establish whether an image was edited or whether every detail matches an unavailable original.

Does a higher photo score cause more leads or sales?

No causal outcome was measured. The benchmark only describes observable photo practices on the collection date.