Fewer wrong matches, more correct sales
| API | Wrong matches | Correct sales found | Median off by (typical search) | Rows checked |
|---|---|---|---|---|
| CompSniper | 5.4% | 64.3% | 3.1% | 3,052 |
| SoldComps | 26.0% | 62.0% | 6.2% | 3,783 |
| Soldgraph | 27.6% | 60.8% | 5.1% | 3,791 |
| Trawl | 25.7% | 54.5% | 8.2% | 3,517 |
| CompsAPI | 31.6% | 54.5% | 10.0% | 3,859 |
| CompSniper, next release (in testing) | 0.5% | Measured on wrong matches | - | Same 40 searches |
Wrong matches: the average, across searches, of the share of returned rows that were not the searched product. Correct sales found: the average share of all correct sales any API returned for a search that this API also returned. Median off by: how far the API's median price was from the median of the correct sales, for the typical search.
Wrong-match rate by product type
CompSniper had the lowest wrong-match rate in every category. The gap was widest on electronics, where a plain keyword search pulls in other models, storage sizes, parts and accessories.
| Category | CompSniper | SoldComps | Soldgraph | Trawl | CompsAPI |
|---|---|---|---|---|---|
| Trading cards (20) | 9.4% | 18.5% | 19.9% | 16.0% | 18.8% |
| Electronics (8) | 0.0% | 44.1% | 44.1% | 42.6% | 52.9% |
| Sneakers, watches and fashion (5) | 4.7% | 22.3% | 28.5% | 36.8% | 42.3% |
| Collectibles and home (7) | 0.3% | 29.2% | 29.9% | 26.5% | 36.0% |
Same search, very different results
The share of returned rows that were not the searched product, for five of the 40 searches:
| Search | CompSniper | The other four | What the wrong rows were |
|---|---|---|---|
| iphone 13 pro 256gb | 0% | 69% to 77% | Pro Max models, other storage sizes, parts and cases |
| airpods pro 2 | 0% | 70% to 90% | single earbuds, charging cases only and earlier models |
| rtx 4090 | 0% | 58% to 91% | other graphics cards, laptops and prebuilt PCs |
| charizard 4/102 base set holo psa 9 | 0% | 36% to 46% | Celebrations reprints, other grades and other graders |
| lego 75192 millennium falcon | 2% | 23% to 43% | minifigures, parts, knockoff sets and incomplete builds |
A median built from a mix of products is the price of the mix. Across the 40 searches, CompSniper's median was typically 3.1% from the median of the correct sales; the other APIs were typically 5.1% to 10.0% off.
One protocol, written before the first call
- Same searches, same moment. The 40 searches were sent to all five APIs on ebay.com within the same few minutes on October 18, 2026, so every API saw the same eBay sales.
- Defaults only. Each API's standard sold-listings endpoint, a plain keyword, newest sales first, and no filters or exclude words. CompSniper used its default cleanup.
- The first 100 rows of each response were checked.
- Definitions first. For each search, the right product was written down before any API was called (for example, “iPhone 13 Pro, not Pro Max, 256GB, working”).
- Blind review. All 7,285 unique sales, matched across APIs by eBay item ID, were labeled correct, wrong or unclear by AI reviewers working from those definitions. Reviewers saw only the title, price and definition, never which API returned the row. Unclear rows are left out of the rates.
- Consistent rules. Lots and multiples count as wrong unless the search asks for one; other grades and graders count as wrong when the search names a grade; parallels and other languages of the right card count as correct unless the search rules them out.
What we searched for
Trading cards (20)
- charizard 4/102 base set holo psa 9
- 2024 topps chrome caleb williams 202
- pikachu celebrations 005/025
- 2018 prizm luka doncic 280 psa 10
- umbreon vmax 215/203
- 2003 topps chrome lebron james 111
- black lotus
- luffy op05-119
- 2023 bowman chrome paul skenes auto
- mewtwo 150/165
- shohei ohtani 2018 topps update us1
- 1986 fleer michael jordan 57 psa 8
- blue eyes white dragon lob-001
- 2020 prizm justin herbert 325 silver
- lugia neo genesis 9/111
- victor wembanyama 2023 prizm 136
- charizard ex 199/165
- 2000 bowman chrome tom brady 236
- ragavan mh2
- psa 10 pikachu illustrator
Electronics (8)
- iphone 13 pro 256gb
- sony a7 iii body
- nintendo switch oled
- airpods pro 2
- rtx 4090
- macbook air m2 13
- steam deck oled 1tb
- canon ae-1 program
Sneakers, watches and fashion (5)
- nike dunk low panda size 10
- jordan 4 bred reimagined 10.5
- rolex submariner 116610
- louis vuitton neverfull mm
- arcteryx beta ar jacket mens
Collectibles and home (7)
- lego 75192 millennium falcon
- hot wheels 2020 custom mustang
- funko pop 01
- yeti tundra 45
- dewalt dcf887
- morgan silver dollar 1921
- vintage pyrex
The next cleanup release
The next version of CompSniper's cleanup is in testing. Run on these same 40 searches three times, it returned 0.4%, 0.6% and 0.4% wrong matches (0.5% on average), down from 5.4% on today's live version. When it ships, we will rerun the full benchmark on the live product and update this page.
Questions about this benchmark
Which eBay sold listings API returns the most accurate results?
In this October 2026 benchmark, CompSniper returned the fewest wrong matches: 5.4% of its rows on today's live version, and 0.5% with its next cleanup release in testing, were not the searched product, against 26.0% for SoldComps, 27.6% for Soldgraph, 25.7% for Trawl and 31.6% for CompsAPI. It also found the largest share of the correct sales (64.3%).
What counts as a wrong match?
A returned sale that is not the product searched for: an accessory or part sold on its own, a different model, size, storage, grade or card number, a lot or bundle of several items, a reprint, proxy or replica, or an incomplete set. The rule for each search was written down before any API was called.
How were the APIs tested?
The same 40 plain keyword searches were sent to all five APIs on ebay.com within the same few minutes on October 18, 2026, with no filters or exclude words, using each API's standard sold-listings endpoint and its default settings. The first 100 rows of each response were checked.
How were the rows checked?
All 7,285 unique sales (matched across APIs by eBay item ID) were labeled correct, wrong or unclear against the written definition for their search. Reviewers saw only the title, price and definition, never which API had returned the row. Unclear rows were left out of the rates.
Why do wrong matches matter for pricing?
A median or average built from mixed results is the price of the mix, not of the item. A search for an iPhone 13 Pro 256GB that also returns Pro Max models, other storage sizes and cases can move the median far from what that phone actually sells for.