The Perpetua MCP answers questions about your advertising data in plain English, which makes it feel like you can ask anything however you like. You mostly can — but how you word a question affects how reliable the answer is, more than you'd expect. This article covers the habits that get you accurate numbers, and the handful of question types that tend to mislead.
None of this is hard. It's five habits, and after a week they're automatic.
Quick links
Name the geo company and the marketplace
A geo company is one advertising account on a single marketplace, within Perpetua. Acme Inc US and Acme Inc CA are two separate geo companies, even for the same brand — which is exactly why "my account" is a risky way to start a question.
As a reminder, to navigate through your different Perpetua accounts, click on the Account Name > Select or Search for the account in question > Click on the specific region flag, as each with have its own unique geo ID:
To determine the geo ID of a specific Perpetua geo account, look for in it in the URL of the Perpetua page you are currently in. In the example below, the geo ID would be 966:
Say what you mean, and say it in full. The Perpetua MCP matches on names, and short fragments can match something you didn't intend. "RY" might find the account you meant, or it might find three others. Full names or geo IDs, every time.
Instead of asking "How's my account doing?" try asking "Show me spend, sales and ACoS by week for the last 8 complete weeks in [full geo company name] on Amazon US."
The second question takes ten more seconds to type and saves you from re-running the whole thing.
Use complete time periods
Today's data is still coming through. Include the current day in a comparison and a half-finished day of numbers make it like you are losing traction — and it'll look that way every time you run it, which is a great way to lose faith in a tool that's working fine. So ask for complete periods and exclude today. "The last 8 complete weeks" beats "the last 2 months."
The bigger trap is the attribution window, and it might even fool people who know their data well.
When someone clicks your ad, they often don't buy straight away. As long as they purchase within the attribution window, that sale is credited back to the original click. Spend is recorded immediately; the sale can land days or weeks later and gets backdated to the click date.
So a recent date range is never finished. Ask about the same week twice, a fortnight apart, and you'll get two different ACoS figures — both correct. Compare spend and sales over the same recent seven days and you'll see spend with very little behind it and go looking for a problem that doesn't exist.
Two things make this bite harder than you'd expect:
Perpetua's attribution window on Sponsored Products is longer than Amazon's. That's deliberate — a longer window means a keyword that brings in a sale weeks after the click gets credit for it instead of having its bid cut. Better bidding, but recent numbers keep filling in for longer.
So Perpetua and Amazon won't agree on a recent window. A gap between the two isn't necessarily an error. The further back you look, the more they converge.
The exact window varies by ad type, platform, and whether you're a seller or a vendor — our attribution window help article has the breakdown.
Therefore, when comparing spend against sales, ask the Perpetua MCP for the attribution window that applies, then pick a period that ended at least that long ago. For a genuinely recent view, ask for spend on its own — that one's settled.
Ask what changed, not why
The Perpetua MCP can see your data. It cannot see Perpetua's engine reasoning — the logic behind a bid change simply isn't in what it has access to.
Here's why that matters: ask a "why" question and you won't get a blank stare. You'll get a fluent, plausible-sounding answer, because the assistant is good at building explanations. That explanation isn't coming from Perpetua, and there's no easy way to tell from looking at it.
So point the question at what you can actually verify.
Instead of asking "Why did my spend jump last week?" try asking "Where did spend increase most in [geo company] between [first complete week] and [second complete week] — broken down by ad unit, then by target?"
That gets you the where, which is usually what you needed anyway. For the why, ask your CSM — they can see things the Perpetua MCP can't.
Write each question so it stands on its own
The Perpetua MCP doesn't carry context from one question to the next the way a conversation does. A follow-up like "now show me last month" has nothing to attach itself to.
Claude will usually fill the gap for you, and usually gets it right — but "usually" is doing a lot of work in that sentence. The safe habit is to restate the geo company, the metric and the time period in every question.
Yes, it feels repetitive. It's also the single easiest way to avoid a wrong answer, and it costs you nothing but a bit of copy-paste.
Asking across several geo companies
If your Perpetua login reaches more than one geo company, you can ask about all of them in a single conversation. For agency teams this is often the moment the Perpetua MCP earns its place — a Monday-morning scan across every brand you manage, in one question.
It's also where one small misunderstanding turns into a whole list of wrong numbers. Two habits keep you safe:
Say explicitly that you want all of them. "Every geo company I can access" is unambiguous. "My accounts" isn't.
Ask for the underlying numbers alongside any ranking. If you ask which geo companies are struggling most, ask for the actual figures too. A ranked list with no numbers behind it is almost impossible to sanity-check — and very easy to forward to a client before anyone has.
IMPORTANT: Be careful comparing geo companies across different marketplaces. Currencies differ, and cross-marketplace totals don't add up the way you'd assume unless you ask for them to be handled properly.
Set the ground rules once
If you use the Perpetua MCP regularly, paste this at the top of a conversation and you won't have to repeat yourself all week:
When answering from Perpetua data: name the geo company and marketplace behind each figure. Use complete time periods and exclude today. Tell me the attribution lag before comparing spend against sales. If a field I asked for isn't available, say so rather than substituting something related. If I ask why something happened, tell me what changed and say that you can't see the engine's reasoning. For questions covering more than one geo company, show me the per-geo-company numbers as well as the summary.
It's long, but you only paste it once per conversation, and it turns most of this article into something you don't have to think about.
Always check before it reaches a client
The Perpetua MCP is early software, and a wrong answer can still look completely right — formatted, tabled, confident, and wrong. That's the failure mode worth guarding against, because it doesn't announce itself.
Treat what comes out the way you'd treat your own export: useful straight away for your own thinking, checked against the Perpetua app before anyone else sees it.
And if a figure doesn't match the app, that's a bug and we'd genuinely like to hear about it. Tell your CSM, and include the exact question you asked alongside what you got back — that pair is what makes it fixable.