From 1 to 4 photos (and an optional part number or vehicle note), KICKBACK TRACE identifies the part, explains what it is for, lists the cars it fits, grades its condition, and gathers real sales and listings so the seller gets a fair price range with links to every piece of evidence.
For your reference: TRACE gives guidance, not a guaranteed identification or an exact price.
Behind one upload, TRACE does the work a seasoned parts dealer would spend an hour on. Here is what it does, in plain words.
It studies every photo you upload: labels, stamps, logos, shape and finish. Faded or partly hidden numbers still get read, and several angles of the same part are combined into one answer.
Brand, model, part number, and what it is for: performance, off-road, or a factory replacement. No more guessing from a blurry stamp.
It checks catalogs, stores and owner forums to list the vehicles the part fits, so buyers know right away whether it works for them.
It searches the live web for real sales and listings of the same part, plus its new price, and brings back each one with a link.
It rates the part from S to D and says why: wear, rust, cracks, missing pieces. If a photo is too blurry to judge, it asks for a retake instead of guessing.
It turns the market data and the condition into a price range and a suggested price. Every fact carries a confidence label and its source, so nothing is a black box.
And it gets smarter with use: every analysis is added to TRACE's own parts database, so the next search for a similar part starts with real evidence.
Gold boxes are where AI does the thinking. White boxes are search, data and math that keep the AI honest.
The AI never invents a price: prices come from real listings and code, and every claim is checked against its source.
Every report puts the facts a seller (and a buyer) actually asks about in one place, each with its source.
Reads labels, stamps and logos in the photo, then matches them against stores and catalogs online.
What the part does and how people use it: performance, off-road, OEM replacement and more.
Years, makes, models and engines it fits, each tied to the page that says so.
A grade from S (like new) to D (heavy wear), with the reasons it saw: rust, dirt, scratches, cracks.
Used examples with price, condition, date and a link, so you can see what similar parts sold or are listed for.
The new-part price for reference, a low / median / high range from real examples, and a suggestion based on condition.
Tone of every report: advisory, “for your reference.” Prices are a guide for setting a fair asking price, never an exact value.
Each number, vehicle, price and description carries one of four labels. Nothing TRACE found is hidden or deleted; weaker items simply get a weaker label.
The source page quotes the value word for word and the decision engine rates the source a full match. A part number also has to be the one read on the photo (or typed by the seller).
There is a quoted source and it partly supports the item, or the number is real but not the one on the photo (for example a replacement number).
There is a link, but no quote that proves the value, or the engine judged the source a poor match. Useful context.
No source link at all. Kept on the report so a person can check it.
This is the real flow in the current code (branch feat/quality-v2, checked Oct 16, 2026). Steps marked “in parallel” run at the same time to save time.
1–4 photos of one part. Optional: the part number and vehicle the seller already knows. Typed hints are treated as data that may be wrong; the photos win if they clearly disagree.
A quick, lenient look at blur, whether the part is in frame, glare or dirt, and whether the label is readable. It stops only unusable photos, before any research money is spent:
Vision AI reads the label text, part numbers and brand. In parallel, a separate step grades condition S–D with reasons. An image search runs only when no part number could be read (it is skipped otherwise to save cost).
A research AI with live web search gets the photos (up to 4, resized) plus any typed hints, and answers one plain question: what is this part, its brand and number, how it is used for tuning, which vehicles it fits, and its used price level, each with source links. Up to 5 web searches across stores, forums and used listings. Kickback’s own site is excluded from this search (Kickback data comes in separately, below).
The answer and its quoted sources are copied into report fields (numbers, uses, vehicles, new prices, used listings). The instruction is to copy every item and drop none; rows without a link are kept with an empty source.
If vehicles, prices, part numbers or the use description came back empty, one extra search asks only for those missing fields, and the results are merged in. In the latest test this happened in about 1 of 5 analyses.
The number read from the label is searched across parts stores and listings. When a label number was read two different ways, a search plus the decision engine picks the right one. Result pages from this search are opened to check the number appears on them.
A decision engine scores each quoted source (0, 1 or 2) against the part. Then every item gets its label:
Note: today this check uses the quoted text returned with the search, not a fresh visit to the page. See “Readiness” below.
Similar items from Kickback’s own marketplace (a read-only copy of its public listings) are matched and added to the price evidence. The price guide then shows:
Kickback matches are listing prices on Kickback, not completed sales. Wording to confirm
Part identity, uses, vehicles, condition grade, price guide and a short plain-English summary. Every item shows its label and source link, plus a checklist of what the buyer should confirm (for example, match the number on the part itself).
TRACE does real work for every part: it pays AI services to read the photos, runs live web searches (charged per search), reads the text of the search result pages (charged by the amount of text), searches the part number, and runs the verification step. None of that is free, so every analysis has a small cost.
This is a planned budget allocation of the $0.08, not a measurement. It deliberately gives the smaller services more room. In the 18-part dev test (Oct 16, 2026) the research AI was about 74% of the measured $0.058; most of that is the web searches themselves and the page text they return.
| Analyses per month | Budget estimate | Basis |
|---|---|---|
| 1,000 | ≈ $80 | $0.08 budget × 1,000 |
| 1,500 | ≈ $120 | $0.08 budget × 1,500 |
| 5,000 | ≈ $400 | $0.08 budget × 5,000 |
Budget estimates at $0.08 per analysis; the measured value was $0.058 (Oct 16, 2026), so real spend should usually be lower. Real cost changes with the number of photos, how many searches a part needs, retries and free monthly allowances. Hosting and storage are not included. One test with full-size photos cost about $0.07 per analysis.
TRACE works end to end today on our dev server. To be straightforward, here is what we still want to improve before and after the US launch.
Today a “Confirmed” item relies on the quote returned with the web search. Next: open the source page itself and check the value is really there.
The same photos can give slightly different results on different runs. We are measuring how much and making results steadier.
Current numbers come from 15–18 test parts. We want a much bigger, more varied set before quoting quality figures.
The engine already uses a typed part number and vehicle; the upload screen does not have that input field yet.
The new lenient check needs tuning on more real photos so it rejects only truly unusable ones and gives the most helpful tips.
In the latest 18-part test round (Oct 16, 2026), no part number was wrongly marked Confirmed, and no run failed with an error.
Ready for demo / pilot use. Not yet for full production. Across three test runs, no part number was wrongly marked Confirmed, and every finding is kept and labeled rather than hidden. Before full production we will add: direct opening of every source page before confirming, more stable results across repeated runs, a larger 50-100 part test set, and seller input fields on screen.