compsapi.com / AI junk filter

AI junk filter for Reverb, Mercari and Poshmark sold prices

Our AI model reads every sale and labels what actually sold: the item itself, or a lot, a part, an accessory, an empty box, a mystery pack or a replica. On Reverb, fender stratocaster reads about $400 without it and about $900 with it.

Try it: the same search with the filter off and on

Try
enter an item above[ LIVE ]

Pick a marketplace and search. You get the median with every sale counted and the median with the AI junk filter on, side by side. Two free searches a day, no account needed.

The switch in the summary bar turns the filter off and on without a second search. Rows the model labelled as junk carry a tag saying what it thinks they are.

What it does

A search for a guitar returns pickups. A search for a console returns games, cases and empty boxes. A median over all of that describes nothing you could buy. Our AI model reads each sale and gives it one label:

LabelWhat was soldCounted as junk
itemOne normal item, the thing the search was forno
lotSeveral items sold together: bundles, bulk lots, complete setsyes
partsA component or a spare: a pickup, a neck, a replacement patchyes
accessorySomething that goes with the item: a case, a strap, lacesyes
box_onlyEmpty packagingyes
mysteryMystery boxes and packs where the buyer does not know the contentsyes
replicaCustoms, reprints and items described as not originalyes
digitalCodes and downloadsyes
otherAnything else that is not a single normal itemyes

Nothing is deleted. Every sale keeps its label, so you can check any call the model made.

How to turn it on

It is opt in, on every plan including the free one. A plain call returns every sale. Two parameters, on /v1/sold and /v1/comps, for eBay, for every marketplace source and for source=all:

  • exclude=ai_junk leaves the labelled junk out of the results and out of the median.
  • ai=1 keeps every row and adds the label to each, so you can filter your own way.
curl -G https://api.compsapi.com/v1/sold \
  -H "Authorization: Bearer YOUR_KEY" \
  --data-urlencode "source=reverb" \
  --data-urlencode "q=fender stratocaster" \
  --data-urlencode "exclude=ai_junk" \
  --data-urlencode "ai=1"

With ai=1 each row carries:

"ai": {"kind": "item", "junk": false, "confidence": 0.987, "category": "music_gear", "model": "2026-11-17"}

A very recent sale can come back before it has a label. The request still succeeds with that row kept, and the response says how many rows were left unlabelled.

The test

Each of the 996 test listings was labelled twice independently, the handful of disagreements were settled by review, and none of these listings were used to train the filter.

Test listingsModel made the same call
The 600 marketplace listings, 51 of them junk97.7%
eBay, measured separately: 690 listings from 20 searches the filter had never seen, 277 of them junkabout 92%

That leaves about 2 calls in 100 on the marketplaces where the model and the label differ, and about 8 listings in 100 on eBay. They are the borderline ones. eBay is measured separately, on searches the filter had never seen. These figures are about the filter only: the sold data comes back exactly as recorded with or without it.

Accuracy by marketplace

What these numbers measure

These percentages measure one thing: how often the AI filter agrees with our labelled test listings about what is junk. They are not a measure of the sold data. Every sale and price comes back exactly as recorded whether you use the filter or not, and with the filter off nothing is removed.

The 600 marketplace listings in the test, split by where they sold. On them the model wrongly removed 8 real items and wrongly kept 6 junk rows. A marketplace gets a percentage only where the sample held enough listings and some real junk.

MarketplaceTest listingsJunk in the sampleSame callMistakes
Mercari961693.8%3 real removed, 3 junk kept
Poshmark63598.4%1 real removed, 0 junk kept
Reverb621095.2%1 real removed, 2 junk kept
Pristine Auction5611All 11 junk listings caught, no real sale removed
Vinted49493.9%3 real removed, 0 junk kept
Goldin40597.5%0 real removed, 1 junk kept
TCGplayer660Junk is rare on this site; in testing the filter removed no real sales.
Fanatics Collect480Junk is rare on this site; in testing the filter removed no real sales.
Grailed410Junk is rare on this site; in testing the filter removed no real sales.
StockX, GOAT, PSA, MySlabsfewer than 40 each0Too few test listings to publish a figure yet.

Mercari and Vinted are the weakest, at about 94%. Pristine Auction and Reverb are the strong cases that had real junk in the sample.

By kind of junk

Across all 996 test listings: Lots 25 of 26 removed, parts 13 of 13 removed, mystery boxes and packs 30 of 32 removed, listings that are not a product, such as services 18 of 18 removed, accessories 35 of 41 removed. Empty boxes, replicas and digital items were too few in the test to quote a figure. Of 549 real single items on the marketplaces, 8 were wrongly removed (1.5%).

Where it is strongest

It is strongest on lots, parts, mystery boxes and listings that are not a product, and it leaves real sales alone on sneaker, streetwear and card sites.

Where to be careful

  • Very short or vague titles, one or two words, are sometimes left out as unclear.
  • Multipacks sold as one product (a 2 pack, a 6 pack, a 2 piece set, a figure 3 pack) are sometimes treated as lots.
  • Camera "body only" listings are sometimes treated as parts.
  • Small add-ons in toy and collectible categories and some instrument add-ons are sometimes kept as items.
  • It is not a counterfeit detector: replicas and "inspired" lookalikes are mostly not caught.

Guitar components sold under their own name (pickups, tuners, bridges, knobs, strings, pedals) count as items. Necks, bodies, pickguards and anything pulled from a guitar count as parts. A bundle of the main item with its own extras (a console with its dock and case) counts as the item, not a lot.

A very recent sale can come back before it has a label, so a first call does not always filter every row. Use ai=1 when you want to see each call for yourself.

Accuracy on eBay

The filter is the same two opt in parameters on eBay: exclude=ai_junk and ai=1 on /v1/sold and /v1/comps. A call without them answers exactly as before. On eBay, /v1/comps with exclude=ai_junk keeps the title rules filter on as well, so it is applied as junk,ai_junk.

On eBay the filter agreed with our labels on about 92% of 690 listings. They came from 20 eBay searches the filter had never seen, and each listing was labelled twice independently and settled by review. It wrongly removed 24 of 413 real items and kept 30 of 277 junk listings, so on eBay the model and the label differ on about 8 listings in 100. In our test searches, about four in ten sold eBay listings were not a single normal item.

These figures measure how well the filter sorts junk from real items. They are not a measure of the sold data, which comes back exactly as recorded with or without the filter.

The same eBay search with and without the filter

Each row is one eBay search three ways: a plain call, with the title rules filter (exclude=junk) and with the AI filter (exclude=ai_junk). Measured on 18 November 2026. Each eBay figure is the median over the 240 newest sales of the search, not the whole window. Figures are rounded.

eBay searchMedian withoutWith the title rules filterWith the AI filterSales removed by the AI filter
rolex submarinerabout $40about $44about $10,600190 of 239
canon eos r5about $65about $63about $1,810167 of 239
omega seamasterabout $220about $265about $1,900116 of 239
dyson airwrapabout $185about $190about $32095 of 239
pokemon charizard base setabout $46about $115about $250135 of 239
lego millennium falcon 75192about $535about $560about $67578 of 239

On a search for a watch or a camera, most of the newest sales are parts and accessories, so the plain median says little about the item. On the Rolex search 49 sales were left after the filter and about 39 of them were watches. The rest were magazines, posters and signs that match the search words.

What it gets wrong on eBay

  • The AI filter removes listings that are not one normal item: parts, accessories, empty boxes, lots, mystery boxes, replicas, digital codes and services. It does not check that what is left is the exact product you searched for. A search for a watch can still return a magazine or a low priced bracelet, a search for a console can still return a controller, and a search for a drone returns mostly controllers.
  • A device sold together with games, a second controller, a headset or other separate products is treated as a lot and removed, even when the device is the main thing in the listing.
  • A game console listed with its own controller, cables, dock or case is meant to be kept and usually is. Some are still removed as bundles: in our checks 2 of 23 higher priced removed listings for Xbox Series X, 2 of 21 for Nintendo Switch OLED, 1 of 18 for PS5 and 5 of 68 for PlayStation 4 Slim were single consoles of this kind. For game consoles, compare the filtered result with the unfiltered one before relying on it.
  • For drones, listings for the drone alone, controllers sold alone and the maker's retail combos are kept. A drone sold by a private seller together with goggles, controller, batteries and case in one listing is sometimes removed as a lot (3 of 17 higher priced removed listings for DJI Avata 2, 2 of 28 for DJI Mavic 3). Spare battery kits and goggles sold alone are kept on some listings and removed on others.
  • Some real lots are kept. A console with games or with two controllers can stay in, most often when the listing says so in a language other than English or without a word like "bundle". In our checks this was 0 to 3 of the 28 most expensive kept listings per console search, and 6 of 71 kept listings at or above the median for PlayStation 4 Slim. The effect on the median is small, but individual kept rows can be lots.
  • Sales that reach us without a title are always removed, because there is nothing to judge. This was 45 of 4,058 sales in our last check.
  • The filter works on one page of results at a time. On a search where most sales are parts or accessories, only a few sales may be left on the page. Check stats.count before using the median.
  • About one real item per page is removed because of unusual wording in the title, for example a camera listed with a mount adapter, a graded card with a certificate number, or a guitar whose title lists its body and neck woods.
  • If labelling does not finish in time, the answer comes back unfiltered and meta.ai.status is not ok. Check that field. In our checks this did not happen on any of the last 70 pages.

Before and after: Fender Stratocaster on Reverb

Measured on 18 November 2026: fender stratocaster on Reverb, last 30 days. Figures are rounded.

CallMedianMiddle half of prices
A plain callabout $400$80 to $1,000
With exclude=ai_junkabout $900$500 to $1,370

On Reverb, fender stratocaster reads about $400 without it and about $900 with it. The model removed about 1,130 of about 2,830 matching sales as parts, accessories or lots. The filter removes lots, pulled parts, accessories, empty boxes and replicas. A pickup set or a pedal sold as a product in its own right counts as an item and stays in, so the cheapest rows after filtering are still worth a look. Listings that include extra games, controllers or other separate products are left out, so the price reflects the item itself. A normal item sold with its case or charger is kept.

What the filter changes

Each row is the same search twice, with and without exclude=ai_junk, over the last 30 days. Measured on 18 November 2026. Figures are rounded.

MarketplaceSearchMedian withoutMedian with the AI filterSales removed
Reverbfender stratocasterabout $400about $900about 1,130 of about 2,830
Reverbgibson les paulabout $1,190about $1,650about 550 of about 1,930
Poshmarkdyson airwrapabout $220about $280about 110 of about 255
Mercarinintendo switch oledabout $165about $200about 37 of about 123
Mercariairpods proabout $55about $60about 75 of about 375
Poshmarklouis vuitton neverfullabout $600about $620about 9 of about 260

The filter moves the median most where a search pulls in parts, accessories and bundles: guitars, consoles, hair tools.

The filter can remove most of a result when the search is dominated by one kind of junk: most Mickey Mantle sales on Pristine this month were mystery box spots.

These are single searches on one day. Run the same requests with your key and you will get today's figures.

The title rules filter, compared

exclude=junk is the older filter. It works from words in the title, such as lot, bundle and replica. In the Stratocaster measurement it removed about a dozen of the matching sales. Words cannot tell a Stratocaster from a Stratocaster pickguard, because both titles say Stratocaster. The AI model reads the whole title together with the price.

The rules filter is unchanged and stays the default on /v1/comps. The two work together as exclude=junk,ai_junk, on eBay and on the marketplace sources.

What it is not

It is a data quality filter. It says whether a sale is a clean comparison for the thing you searched. It makes no statement about whether an item is genuine: that is the separate authenticity object on each row, which passes on what the marketplace states. And it is a model, so it is wrong sometimes. The numbers above say how often.

Looking for how to use CompsAPI from ChatGPT, Claude or another assistant? That is the AI assistants page.

Questions people ask

What does the AI junk filter do? Our AI model reads each sale and labels what was sold: a single item, or a lot, a part, an accessory, a box only, a mystery pack, a replica or a digital item. With exclude=ai_junk the labelled junk is left out of the results and the median. Nothing is deleted, and ai=1 shows the label on every row.

How accurate is it? In testing the model made the same call as our test labels on 97.7% of 600 marketplace listings and on about 92% of listings on eBay searches it had never seen. Each of the 996 test listings was labelled twice independently, the handful of disagreements were settled by review, and none of these listings were used to train the filter. It is weakest on Mercari and Vinted, at about 94%. These figures measure how well the filter sorts junk from real items. They are not a measure of the sold data, which comes back exactly as recorded with or without the filter.

How do I turn it on? Add exclude=ai_junk to a sold or comps request to filter, or ai=1 to get the label on each row and decide yourself. It works on eBay and all 12 marketplaces. It is opt in and available on every plan, including the free one. A plain call returns every sale.

Does it replace exclude=junk? No. exclude=junk is the title rules filter and stays as it is. The two can be combined as exclude=junk,ai_junk.

Part of the November 2026 release.

Price from what things
actually sold for

100 free requests a month · every filter on · no card

eBay plus 12 more marketplaces on the same key

Get a free key