Picking the right resale platform is a routing problem, not a price problem

A friend of mine buys vintage Levi's at estate sales. She's good at it. But twice now I've watched her list a $90 jacket on Facebook Marketplace because that's the app she had open, when the same jacket clears $140 on Depop in two days. The item was never the problem. The destination was.

I've been building a thrift-pricing tool, and for a while I treated "where should this sell" as an afterthought — a string I appended to the price estimate. eBay for most things, Depop for streetwear, done. That was wrong, and the way it was wrong is interesting enough to write up.

The real question isn't "which platform has the highest price." It's "which platform gives the best expected outcome for this item, accounting for how fast it sells, what the fees take, and how much the seller has to do." Those four variables pull in different directions, and the naive version optimizes only the first one.

Here's the concrete failure. eBay almost always shows the highest sold prices for collectibles and electronics, so a top-price heuristic routes nearly everything there. But eBay's final value fee is ~13%, listings can sit for weeks, and for a $25 pair of jeans the effort-to-payout ratio is terrible. Depop and Poshmark skew toward fashion buyers who'll pay more for the same garment and buy faster, even though their fees are similar or worse. Facebook Marketplace has no selling fee and instant local cash, but only for things people will drive to pick up — furniture, appliances, bikes. A single "highest price wins" rule gets the furniture case exactly backwards.

So I reframed it as a routing problem. Each platform is a candidate destination with a score, and the score is a function of the item's attributes, not a fixed ranking.

The inputs that actually matter, roughly in order of signal:

• Category and subcategory. Fashion vs electronics vs furniture vs collectibles is the single biggest splitter. Get this right and you've made 70% of the decision.
• Price band. Sub-$30 items punish slow, high-fee platforms because the absolute fee is small but the time cost dominates. High-value items justify the wait.
• Shippability. A dresser can't go on Poshmark in any practical sense. Weight and dimensions quietly veto whole platforms.
• Brand and "search demand." A known brand (Carhartt, Le Creuset, Sony) has buyers actively searching on specific platforms. Generic items don't, and do better with browse-driven local sale.

The model I landed on isn't machine learning — it's an explicit scoring function, and I want to defend that choice because it's unfashionable. I have a small table of platform profiles (typical fee %, median days-to-sell by category, audience skew, shippability requirement). For a given item I compute an expected-net-per-week-of-effort number per platform and rank them. Something close to:

score(platform, item) =
  estimatedPrice(platform, item.category)
    * (1 - platform.feeRate)
    * demandMultiplier(platform, item.brand)
  / max(platform.medianDaysToSell[item.category], 1)
  - shippingFriction(platform, item)

It's crude. The estimatedPrice per platform is itself an estimate stacked on an estimate, and medianDaysToSell is a table I assembled from public data and seller anecdotes, not a live feed. I'm not going to pretend that's rigorous. But the explicit version has one property an ML model wouldn't give me cheaply: when it routes a couch to eBay, I can read the table and see that shippability friction should have vetoed it, and I can fix the rule in one line. A black box that's confidently wrong on furniture is much worse than a transparent rule that's wrong in a way I can inspect.

The vision model feeds this. Item identification comes from a multimodal model (GPT-4o class) that returns category, probable brand, and condition cues from the photo. That output is the input to the routing function — and it's also where the whole thing is most fragile. If the model calls a Coach bag "generic brown handbag," the demand multiplier collapses and it routes to local sale instead of Poshmark, costing the seller real money. The routing logic is only as good as the identification feeding it, and identification is the part I trust least. It will misidentify things. I surface the detected brand prominently so a human can catch it before listing, rather than hiding the assumption inside a confident-looking recommendation.

The tradeoff I'm still not satisfied with: the fee and days-to-sell tables are static. Real fee structures change, promoted-listing dynamics shift, and a category that sold in 3 days last year takes 9 now. The honest version of this tool pulls live comparable-sale velocity per platform. I haven't built that — the data access is uneven across platforms, and some of them actively don't want to be queried — so I'm shipping with periodically-updated tables and being upfront that they're approximations. It's the weakest part of the system and I'd rather say so than dress it up.

What surprised me is how much value sits in just not defaulting to the open app. Even an imperfect routing score beats "whatever I had open," because the baseline isn't a smart choice — it's habit. My friend doesn't need a perfect model. She needs something to say "this one's a Depop jacket, not a Facebook jacket" before she lists it wrong.

If you want to see the version that's running, it's at https://thriftflipper.app — photo an item, get the value range and the suggested platform. It's rough on identification and the platform tables are approximations, but it works for the decision it's meant to help with.