AI Doesn't Pick the Winner: How the Best-Offer Flag Is Calculated and Why the Final Word Is Yours

The question behind the badge

When offers start arriving for a natural stone demand on MarbleMap, one of them may carry a small flag: AI pick, best offer. In B2B, that kind of badge earns suspicion before it earns trust — and it should. Any platform that puts an algorithm between a buyer and a serious purchasing decision owes that buyer two answers: how exactly is this calculated, and what happens if I ignore it?

This article answers both. It is not a how-to on comparing offers. It is about the harder question underneath: how much should you trust the machine?

A formula you can read in one sentence

Here is the entire logic of the flag: it goes to the offer with the lowest grand total on a shared euro basis, among the offers that priced every line of your demand. No hidden weighting, no learned preferences, no scoring model that shifts under your feet.

Two mechanical details make that sentence honest. First, currency. Suppliers quote in their own currency and every offer keeps it, but the system converts each total to an approximate EUR equivalent so that quotes from different countries can sit on one axis. Without a shared basis, lowest would mean nothing.

Second, withdrawn offers are excluded from every calculation. An offer a supplier has pulled back cannot win the flag, cannot serve as the runner-up, and cannot distort the numbers you see. The comparison reflects only offers that are actually on the table.

A formula this narrow is a deliberate choice. A rule you can verify in your head is worth more than a clever one you have to take on faith.

Why the cheapest-looking offer is sometimes excluded on purpose

The most consequential rule is coverage. On itemized demands, only full-coverage offers — the ones that priced every single line — compete for the flag. An offer that skips lines looks cheap for the wrong reason: its total is lower because it answered a smaller question, not because it is more competitive. Rewarding it would teach every supplier to quote selectively and quietly break the comparison for everyone.

So coverage is labeled in the open. A full-coverage tag means all lines were priced; a partial tag such as three out of five tells you exactly what is missing, and unpriced lines appear as a plain dash in the comparison table. The same discipline runs through the rest of the tooling: the AI top five shortlist draws only from full-coverage offers, and when you consider splitting a demand across suppliers, the single-supplier benchmark is the cheapest full-coverage offer. The savings you see are real, not an artifact of missing lines.

The plan boundaries are stated just as openly: the AI recommendation and the top five shortlist are Pro features; the comparison table for itemized demands belongs to Pro as well, with an upgrade card in its place on the free plan; and the board for splitting a demand across suppliers opens on the paid buyer plans, Pro and Business.

Partial offers are not punished for existing. They are kept out of the like-for-like comparison, nothing more. They stay visible, they can be filtered and sorted, and they can still win the individual lines they actually quoted. On general requests without line items, the distinction disappears entirely and offers compare on total price alone.

What the card shows, and what it refuses to decide

Transparency continues on the offer card itself. The flagged offer carries a one-line explanation of why it was chosen, factor tags, and a concrete figure: how much cheaper it is than the runner-up. The system shows its work instead of asking for faith.

Just as telling is what the card does not have. There is no accept-for-me button, no countdown, no automation waiting to close the deal. The flagged offer presents exactly two actions, review and accept, and both belong to you. No offer is ever accepted on its own, no demand ever closes on its own, and nothing becomes binding until the buyer chooses to accept.

That restraint is not modesty for its own sake. The formula counts money and coverage — the things a machine can count honestly. It does not know that your project schedule makes a short lead time worth a premium, that certain payment terms matter more than a three percent gap, or that you have years of history with one of the suppliers. Those judgments stay with you, which is why you can ignore the flag entirely, sort by lead time, rating or coverage instead, and even award different lines to different suppliers. The system recommends; the buyer decides.

Why a commission-free platform can afford a boring algorithm

Trusting an algorithm ultimately means trusting the incentives of whoever wrote it. A marketplace that takes a percentage of every transaction has a structural interest in nudging you toward closing, and toward closing bigger. Its recommendation engine does not need to be malicious to drift; the business model leans on it.

MarbleMap takes no commission. Its revenue comes from subscriptions alone; the platform takes no share of any sale, and payment never flows through it. Whether you accept the flagged offer, a different one, or none at all, the platform earns exactly the same. Add sealed offers that competitors cannot see and verified companies on the supplier side, and the flag has nothing to gain from being wrong. A boring, auditable formula is not a limitation of the system — it is the natural product of a business model that does not need to steer you.

The final word is yours

If you buy natural stone, open a demand on MarbleMap — buyers start free — and watch the comparison work for you instead of on you. If you supply stone, demands from vetted buyers are waiting for your sealed offer, hidden from your competitors. Either way, the machine does the counting and you do the deciding, at marblemap.co.