CompSniper
Original research

We tested 100 eBay sold-comp searches. 37.9% of priced rows were classified out.

Across 100 predeclared product searches, CompSniper parsed 19,220 priced rows and retained 11,942 after title-based relevance cleaning. The product median changed by at least 10% in 34 searches and by at least 50% in 11. This measures classifier behavior on one fixed sample, not human-labeled accuracy or future sale prices.

Collected August 31, 2026 from ebay.com. Published and reviewed by Marc Andrew, Owner of CompSniper.

100
products
19,220
raw priced rows
11,942
cleaned rows
7,278
rows removed
34
medians moved 10%+
18.3%
Best Offer marked
Main finding

Cleaning often changed the sample more than the median

The weighted row-removal rate was 37.87%, while the median absolute product-level price change was 4.2%. That distinction matters. Removing many obvious accessories may leave the center price nearly unchanged, while a smaller number of badly priced mismatches can move it substantially.

Direction of median changes
73
increased
21
decreased
6
unchanged
Absolute median change

Most searches moved modestly, but the tail was large

Under 1%28 products
1% to under 5%25 products
5% to under 10%13 products
10% to under 25%19 products
25% to under 50%4 products
50% or more11 products
Category comparison

Removal rates varied by product type

Each category contained exactly 20 products. The percentages below are weighted by priced rows, while the median price-change column gives equal weight to each product.

Smartphones and tablets45.99%
Gaming and computing26.28%
Cameras and lenses30.31%
Audio and music40.6%
Collectibles and luxury47.05%
CategoryRawCleanRemovedMedian abs. change
Smartphones and tablets4,1682,25145.99%4.74%
Gaming and computing4,2843,15826.28%3.21%
Cameras and lenses3,6162,52030.31%0.52%
Audio and music3,5072,08340.6%7.03%
Collectibles and luxury3,6451,93047.05%8.84%
Largest movements

The tail is why row-level evidence matters

These are the 15 largest absolute median changes in this sample. A large movement is a reason to inspect the retained and removed listings, not proof that every classification was correct.

ProductCategoryRowsRaw medianCleaned medianChange
GoPro Hero 12 BlackCameras and lenses219 to 61$59.99$230.00+283.4%
Lenovo ThinkPad X1 Carbon Gen 11Gaming and computing222 to 14$293.50$602.49+105.28%
Canon RF 70-200mm f2.8 L IS USMCameras and lenses224 to 46$1,104.50$2,097.00+89.86%
NVIDIA GeForce RTX 4090Gaming and computing233 to 51$1,399.00$2,599.99+85.85%
Sony Walkman TPS-L2Collectibles and luxury114 to 45$328.55$573.00+74.4%
Yamaha HS8 pairAudio and music234 to 39$246.36$420.00+70.48%
Sony FE 24-70mm f2.8 GM IICameras and lenses229 to 62$1,061.10$1,714.50+61.58%
Samsung Galaxy Tab S9 128GBSmartphones and tablets209 to 19$211.99$340.00+60.38%
Panasonic Lumix GH6 bodyCameras and lenses138 to 20$571.00$882.49+54.55%
Google Pixel 7 Pro 128GBSmartphones and tablets226 to 69$424.09$199.99-52.84%
LEGO Titanic 10294Collectibles and luxury179 to 107$400.00$600.00+50%
Line 6 Helix FloorAudio and music201 to 23$554.99$795.00+43.25%
Bose QuietComfort Ultra HeadphonesAudio and music228 to 100$165.20$226.25+36.96%
Chanel Classic Flap Medium BlackCollectibles and luxury226 to 136$3,000.00$4,025.00+34.17%
Audient iD14 MKIIAudio and music91 to 27$179.99$225.00+25.01%
Study protocol

Fixed before the first request

  1. 1. Products: 100 named products across five categories, 20 per category, selected before collection.
  2. 2. Query: One keyword per product on ebay.com, page one, ended-recently sort, any condition.
  3. 3. Window: Sold dates from June 2 through August 31, 2026.
  4. 4. Sample: 240 requested results. Pre-cleaning counts can be slightly higher because eBay may return more rows before the response cap is applied.
  5. 5. Comparison: `rawMedian` and `rawSampleCount` were captured before title classification from the same fetched page. Cleaned statistics were computed afterward.
  6. 6. Execution: Two workers, one-second start pacing, bounded temporary retries. All 100 requests succeeded with unique request IDs.
Limitations

What this study can and cannot establish

No human-labeled ground truth

The study measures production classifier behavior. It does not calculate precision, recall, or prove that every removed row was irrelevant.

One marketplace and one run

All searches used ebay.com on August 31, 2026. Other dates, marketplaces, pages, conditions, and query wording can produce different samples.

Bounded search pages

A result page is not the full population of eBay sales. Reported headline counts can be much larger than the parsed sample.

Best Offer uncertainty

2,184 cleaned rows carried an accepted Best Offer marker. The accepted amount is not disclosed and may differ from the visible price.

Title-based classification

Cleaning used listing titles. Details visible only in descriptions or images may not affect classification.

Not a price forecast

A cleaned median summarizes this historical sample. It is not a formal appraisal or a guarantee of a future transaction price.

Predeclared sample

All 100 products

Smartphones and tablets

  1. 01Apple iPhone 15 Pro 128GB
  2. 02Apple iPhone 14 Pro 128GB
  3. 03Apple iPhone 13 128GB
  4. 04Samsung Galaxy S24 Ultra 256GB
  5. 05Samsung Galaxy S23 Ultra 256GB
  6. 06Google Pixel 8 Pro 128GB
  7. 07Google Pixel 7 Pro 128GB
  8. 08OnePlus 12 256GB
  9. 09Apple iPad Pro 11 4th generation 128GB
  10. 10Apple iPad Air 5th generation 64GB
  11. 11Samsung Galaxy Tab S9 128GB
  12. 12Apple Watch Ultra 2
  13. 13Apple Watch Series 9 45mm
  14. 14Samsung Galaxy Watch 6 Classic 47mm
  15. 15Sony Xperia 1 V 256GB
  16. 16Motorola Razr Plus 2023
  17. 17Samsung Galaxy Z Fold 5 256GB
  18. 18Samsung Galaxy Z Flip 5 256GB
  19. 19Kindle Paperwhite 11th generation
  20. 20reMarkable 2 tablet

Gaming and computing

  1. 01Sony PlayStation 5 Slim Disc Edition
  2. 02Microsoft Xbox Series X 1TB
  3. 03Nintendo Switch OLED
  4. 04Steam Deck OLED 512GB
  5. 05ASUS ROG Ally Z1 Extreme
  6. 06NVIDIA GeForce RTX 4090
  7. 07NVIDIA GeForce RTX 4080 Super
  8. 08AMD Radeon RX 7900 XTX
  9. 09Apple MacBook Pro 14 M3 Pro 18GB 512GB
  10. 10Apple MacBook Air M2 13 8GB 256GB
  11. 11Dell XPS 15 9530
  12. 12Lenovo ThinkPad X1 Carbon Gen 11
  13. 13ASUS ROG Zephyrus G14 2023
  14. 14Intel Core i9-14900K
  15. 15AMD Ryzen 9 7950X3D
  16. 16Samsung 990 Pro 2TB SSD
  17. 17Meta Quest 3 128GB
  18. 18PlayStation Portal Remote Player
  19. 19Nintendo 3DS XL
  20. 20Analogue Pocket

Cameras and lenses

  1. 01Sony A7 IV body
  2. 02Sony A7 III body
  3. 03Canon EOS R6 Mark II body
  4. 04Canon EOS R5 body
  5. 05Nikon Z6 II body
  6. 06Fujifilm X-T5 body
  7. 07Fujifilm X100VI
  8. 08Panasonic Lumix GH6 body
  9. 09DJI Osmo Pocket 3 Creator Combo
  10. 10GoPro Hero 12 Black
  11. 11Sony FE 24-70mm f2.8 GM II
  12. 12Sony FE 70-200mm f2.8 GM II
  13. 13Canon RF 24-70mm f2.8 L IS USM
  14. 14Canon RF 70-200mm f2.8 L IS USM
  15. 15Nikon Z 24-70mm f2.8 S
  16. 16Sigma 24-70mm f2.8 DG DN Art Sony E
  17. 17Tamron 28-75mm f2.8 G2 Sony E
  18. 18Fujifilm XF 16-55mm f2.8 R LM WR
  19. 19Leica Q3
  20. 20DJI Mini 4 Pro Fly More Combo

Audio and music

  1. 01Shure SM7B
  2. 02Shure MV7
  3. 03Rode NT1 5th Generation
  4. 04Neumann TLM 103
  5. 05Audio-Technica AT2020
  6. 06Sennheiser HD 650
  7. 07Sennheiser HD 800 S
  8. 08Sony WH-1000XM5
  9. 09Bose QuietComfort Ultra Headphones
  10. 10Apple AirPods Max
  11. 11Focusrite Scarlett 2i2 4th Gen
  12. 12Universal Audio Apollo Twin X Duo
  13. 13Audient iD14 MKII
  14. 14Yamaha HS8 pair
  15. 15KRK Rokit 5 G4 pair
  16. 16Fender Player Stratocaster
  17. 17Gibson Les Paul Standard 60s
  18. 18Boss Katana 50 MKII
  19. 19Line 6 Helix Floor
  20. 20Akai MPC Live II

Collectibles and luxury

  1. 01Rolex Submariner 126610LN
  2. 02Omega Speedmaster Moonwatch 3861
  3. 03Tudor Black Bay 58 79030N
  4. 04Seiko Prospex SPB143
  5. 05Air Jordan 1 Chicago Lost and Found size 10
  6. 06Nike SB Dunk Low Travis Scott size 10
  7. 07Adidas Yeezy Boost 350 V2 Zebra size 10
  8. 08Pokemon Charizard Base Set 4/102 PSA 9
  9. 09Pokemon Pikachu Van Gogh PSA 10
  10. 102023 Topps Chrome Victor Wembanyama #1 PSA 10
  11. 112018 Panini Prizm Luka Doncic #280 PSA 10
  12. 121986 Fleer Michael Jordan #57 PSA 8
  13. 13LEGO Millennium Falcon 75192
  14. 14LEGO Titanic 10294
  15. 15Nintendo Game Boy Advance SP AGS-101
  16. 16Sony Walkman TPS-L2
  17. 17Louis Vuitton Neverfull MM Monogram
  18. 18Gucci GG Marmont Small Shoulder Bag
  19. 19Chanel Classic Flap Medium Black
  20. 20Hermes Birkin 30 Togo
Questions

Study FAQ

What did the 100-product study measure?

It measured how CompSniper's title-based relevance cleaning changed the priced sample and median for 100 predeclared product searches. Each request reported the raw sample count and median from the same upstream page, then returned the cleaned rows and deterministic price summary.

Did relevance cleaning always increase the median?

No. The median increased for 73 products, decreased for 21, and was unchanged for 6. Cleaning removes rows based on product relevance, not whether their prices are high or low.

Does this study prove that every removed listing was wrong?

No. This was not a human-labeled precision and recall evaluation. It reports what the production relevance classifier removed and how the summary changed. The public product-level CSV does not claim that every classification decision was correct.

How were accepted Best Offer listings handled?

The cleaned sample contained 2,184 rows marked as accepted Best Offer, or 18.29% of cleaned priced rows. eBay does not disclose the accepted amount, so those markers are reported as a limitation rather than treated as exact transaction prices.

Inspect the data or run your own sample

Download the product-level CSV, inspect the reproducible collector, or use the free API tier to test your own queries with raw and cleaned evidence.