How Our AI Pricing Works (And Where It Falls Short)

We built Clutter to Cash because I spent too many evenings pricing used goods by hand, and it seemed silly that every person selling a camera lens or a pair of jeans had to re-do the same research from scratch. This guide is the honest, technical version of how the tool works under the hood: what data we use, what the AI is actually doing, and the specific situations where the tool gets things wrong. If you're going to trust a price we give you, you should know where it came from.

What happens when you hit "Analyze"

When you upload photos and submit, three things happen in sequence. We'll walk through each one in turn.

1. Identification. Your photos are sent to a vision-capable AI model (Google's Gemini Flash is our first pass). The model is asked what the item is, what brand and model (if visible), and what condition the photos suggest. It returns a structured answer with a confidence score.
2. Pricing research. Once we have a confident identification, we query the eBay Browse API for currently-listed items matching that identification. Where completed-sale data is available for the category, we fold that in.
3. Listing generation. A text model uses the identification, condition notes, and pricing data to draft an eBay-formatted title, description, and price range. The same copy pastes cleanly into other marketplaces with light tweaks. You see the draft. You copy it. You leave.

Stage 1: how identification actually works

The vision model sees your photos the way a human would, except it's been trained on billions of product images and can often recognize a specific model number, label, or brand mark from a photo without you having to name it. For the 80% of cases where the photo is clear and the item is mainstream, Gemini Flash returns a confident identification on the first pass in under two seconds.

For the remaining 20% (blurry photos, missing labels, obscure items, counterfeit goods) we fall back to a two-step enrichment. First we run the primary photo through Google Lens to pull visually similar products from the open web. Then we re-ask the model with that extra context. If the confidence score is still below threshold, we escalate to Gemini Pro, which is slower and more expensive but meaningfully more capable at the edges of the distribution.

The key design decision here is the confidence threshold itself. If we ship a confident answer on a low-quality signal, we send you a wrong price and waste your time. If we threshold too conservatively, we reject too many legitimate queries. Our current threshold lands around 0.7 on the model's self-reported confidence, which in practice means about 1 in 20 uploads comes back with a "we couldn't confidently identify this" message. That rate is tuned deliberately.

Stage 2: where the price numbers actually come from

A lot of pricing tools for resale quietly make up numbers. We don't, and it matters that we don't. The anchor today is the eBay Browse API: a fresh query against the currently-listed catalog for items matching the identified model and category. eBay has the deepest public catalog of U.S. resale items, which is why it's the anchor. Where completed-sale data for the category is available, we fold it in alongside the active listings.

Active listing prices skew higher than what items actually move for — an asking price is what the seller hoped to get, not what a buyer agreed to pay. eBay's own seller documentation acknowledges this gap, and on most categories it can be substantial. The tool knows that and weights toward the middle of the comparable cluster, flags outliers, and surfaces a range rather than a single point estimate so you can see the spread.

We don't blend in live data from other marketplaces today (Facebook Marketplace, Mercari, Poshmark) — cross-marketplace coverage is on the roadmap. The eBay catalog is broad enough that for most U.S. resale categories it's a reasonable single anchor. For unusual categories where eBay isn't the natural marketplace (vintage furniture, niche collectibles), the tool will flag a wider range or tell you it can't find enough comparable listings.

Stage 3: listing generation

The final step is the listing copy. A language model takes the identification, your condition notes, and the comp-derived price and drafts an eBay-formatted title plus description following the structure described in our listings guide. The same copy pastes cleanly into Facebook Marketplace, Mercari, or other platforms with light formatting tweaks; the draft itself is targeted at eBay's structure today. The model is instructed to be keyword-dense in the title, specific about flaws in the description, and to never invent features or condition claims that weren't in the input.

You should treat the generated listing as a strong first draft, not a finished product. Edit it. Add context the AI couldn't know: where you bought it, why you're selling, how long you've had it. These small human touches measurably increase buyer trust and conversion, and the AI is explicitly instructed not to fabricate them because getting them wrong is worse than leaving them blank.

Where the tool gets it wrong

Being honest about the failure modes matters more than hiding them. Here are the specific situations where you should not trust our output as-is.

• Luxury authentication. The AI cannot tell a genuine Louis Vuitton bag from a good counterfeit. For items where authenticity materially changes the price (designer bags, high-end watches, collectible sneakers) use a dedicated authentication service before you list.
• Items with no meaningful sold-comp history. One-of-a-kind vintage pieces, handmade goods, regional or discontinued products, and items released in the last two weeks don't have enough data to anchor a price. In these cases we surface a wider range and flag it explicitly.
• Local market variance. A snowblower in Buffalo is worth more than a snowblower in San Diego. We don't adjust for your location. If the item is heavy enough that shipping is prohibitive and sales are local-only, our national-average price may overstate what you can actually get.
• Ambiguous identifications. The AI sometimes picks the more famous variant of a product when the real answer is the less famous one. A 2023 Toyota Camry and a 2013 Toyota Camry look similar in some photos, but their prices are completely different. Always spot-check the year and exact model against your item.

Our data practices

Your photos are normalized in your browser, uploaded to Cloudflare R2 under an anonymous, session-scoped path, and handed to the AI providers for analysis. They're retained for up to six months and deleted automatically by an R2 lifecycle rule. During that retention window the inputs are used only to improve the tool itself: better identification, better pricing, better drafts. They're never sold, never shared outside the providers in the request loop, and never tied to a person (no account exists to tie them to). Our full data practices are documented in the privacy policy, and the AI providers we use — Google Gemini for identification (with a Google Lens fallback) and OpenAI for the listing draft — are disclosed in the disclosures page.

That's the whole pipeline. If you've got questions about a specific result the tool gave you (or a case where you think we got it wrong) write to info@icandothat.ai. Feedback on mispriced categories directly shapes which categories we improve next.