How to evaluate an AI Amazon PPC tool: a nine-point checklist.

The fastest way to evaluate an AI Amazon PPC tool: ignore the demo and ask for the decision trace. If a vendor cannot show you, bid by bid, why it changed a specific bid last Tuesday, you are not buying intelligence — you are buying a black box with a Stripe key. Below is the nine-point checklist we wish more sellers used before signing.

More than 70% of Amazon sellers now run ads, and a growing share of those accounts are managed by some flavour of "AI." The label has stopped meaning anything. These nine questions separate a real profit model from an autobidder wearing a wig.

1. Can it explain a single bid?

Ask: "Show me why this bid moved." A defensible tool produces a trace — the inputs, the model's estimate, the alternatives ruled out, the final number. If the answer is "the algorithm decided," you cannot defend that spend to a CFO, and neither can the vendor. See what a trace should contain on the transparency page.

2. Does it bid on profit or on a proxy?

ACoS and ROAS are proxies. Total net profit is the target. A tool that optimises ACoS will happily push bids up on keywords that look efficient and erode margin. Ask which objective the bidder actually maximises — and read why TACoS beats ACoS if the answer is fuzzy.

3. Is it inventory-aware?

A bidder that does not read your SP-API days-of-cover will keep bidding to first page on an ASIN that is about to stock out — then make you re-buy the rank you lost. Ask whether bids throttle as inventory thins. We cover why this is the single most expensive blind spot in inventory-aware bidding.

4. How does it behave on day one?

Most tools need weeks of conversion history before they bid sensibly — a cold start that costs you real money. Ask what happens in hour one. A Bayesian warm-start produces reasonable bids immediately and tightens as data arrives, rather than flailing through a learning period on your budget.

5. Can you watch it before it spends?

This is the question that filters the field fastest. Can the tool run in full simulation on your real account, changing nothing, so you can audit its decisions before any money moves? Most vendors demand trust before evidence. The honest ones invert it — see how Simulation Mode works.

6. What does it cost when you scale?

A percentage-of-spend fee means the tool earns more every time you spend more — its incentive and yours point in opposite directions. Pacvue charges around 3% of spend; agencies often take 10–20%. A flat fee is paid the same whether your spend goes up or down. We explain why we will never charge a percentage in this post.

7. Was it built for your marketplaces?

A tool built for the US and shipped to Europe with a translation pack mis-bids on DE, FR, IT, and the rest, because keyword economics and shopper behaviour differ per marketplace. If you sell across borders, ask whether the bid math is per-marketplace or one model translated. Background: EU-native is not a translation layer.

8. What are the safety limits?

An autonomous bidder needs circuit breakers — per-marketplace spend caps, anomaly detection, rollback. Ask what stops a runaway. If the answer is "it won't do that," that is not a safety system, it is a hope. See the seven-layer architecture for what a real one looks like.

9. Who owns the data, and where does it live?

For EU sellers especially: where is the account data processed, and is the vendor a GDPR-compliant processor with a DPA? This is a procurement question that kills deals late if you leave it to the end. Ask it first.

The shortcut

If you only have time for one question, make it number one. Everything else — profit objective, inventory awareness, safety, fair pricing — tends to follow from whether a vendor built a system that can explain itself. Tools that can, usually did the rest of the work too. Tools that can't, usually didn't.

Compare the named incumbents side by side on the comparison page, or apply to watch Mirox decide on your account before you commit to anything.