Find the spend that is not coming back.

The question nobody has time to ask every week

You have four tabs open and three of them disagree about what your ROAS was yesterday. Somewhere in there is an ad set that has spent six hundred dollars this month and produced two purchases, and a search term that has been quietly buying clicks from people looking for a free version of your product.

This is not a hard analysis. It is a tedious one, and tedious loses to whatever is on fire. So the leak stays. The value of asking an assistant is not that it is smarter than you about your account. It is that it will actually do the sweep, on a Tuesday, in ninety seconds, because you asked it in the same window where you were already working.

How Muze finds it

On Meta, list_account_ads returns the account's whole working set, every ad with its insight data attached, and takes an optional date range to narrow the window. get_ad_insights drills into one ad's stored per-period rows. compare_meta_performance runs the period-over-period comparison, defaulting to month-to-date against the same span last month, and get_meta_last30days gives a fixed rolling window when you want a stable denominator.

On Google, list_google_ads_data returns every entity for a customer with insights attached, including Performance Max asset groups, and inspect_google_campaign_ads goes one campaign deep. compare_google_performance mirrors the Meta comparison, which matters more than it sounds: the same shape of answer on both platforms is what lets an assistant reason across them without doing arithmetic it will get wrong.

On Amazon, waste has a specific shape and Muze treats it that way. list_amazon_search_terms returns pending recommendations that harvest converting terms as keywords and block the wasteful ones as negatives, each row carrying a row_id so you can apply or dismiss them one at a time.

Then you decide what to do about it

Finding waste and cutting it are separate operations, on purpose. Nothing on the read side of this feature changes your account, which means you can run the sweep as often as you like without thinking about it.

When you do want to act, the tools are explicit. On Meta: pause_ad, pause_ad_set and pause_meta_campaign, each classed as destructive, and launch_plan_pause_ads for a batch. pause_ad will not run without confirm=True, and its own description tells the assistant to ask you first. On Google: pause_google_campaign, and apply_google_campaign_action for budget, target CPA and target ROAS changes, whose description states plainly that spend changes take effect immediately. On Amazon: apply_amazon_search_term applies one recommendation to the live account, create_amazon_negative_keywords batch-creates exclusions, update_amazon_keyword_bid adjusts a bid, and update_amazon_campaign_budget or set_amazon_campaign_state handles the campaign level.

What counts as waste is your call

Muze does not know your margin, so it cannot tell you what a bad ROAS is. A 2.0 is excellent for one business and a slow bleed for another. The honest version of this feature is that it surfaces the spend and the return, and you or your assistant apply the threshold.

If you want the threshold to be arithmetic rather than instinct, break-even ROAS is one divided by your gross margin. At a 40% margin you need to clear 2.5 before an ad is contributing anything. Tell your assistant the number once and the sweep becomes a filter rather than a list.

Two limits worth knowing before you trust a total. Amazon figures on this connection are ad-attributed only, because Muze holds the Amazon Ads campaign management scope and not SP-API, so there is no marketplace revenue figure. And cross-platform totals are summed without currency conversion, which only means anything if your accounts share a currency.

What to type

These are prompts, not commands. Any MCP client will do: ChatGPT, Claude, Cursor or anything else that speaks the protocol.

• “Which Meta ads spent more than $100 in the last 30 days with no purchases?”
• “Compare this month against last month on Google and tell me which campaigns got worse.”
• “Show me Amazon search terms that spent with zero orders, then add the worst as negatives after I review the list.”
• “Across Meta, Google and Amazon, where is my worst return this week?”
• “Pause the five worst ads in this ad set, and ask me before each one.”

The tools behind it

The whole detection half is read-only. Every tool that stops or changes spend is separately classed, and the destructive ones will not fire without explicit confirmation.

• list_account_ads (read): Every Meta ad with insight data, optional date range.
• compare_meta_performance (read): Meta period over period, this span against the last.
• list_google_ads_data (read): All Google entities with insights, PMax asset groups included.
• compare_google_performance (read): The Google mirror of the Meta comparison.
• list_amazon_search_terms (read): Pending harvest and negative recommendations, each with a row_id.
• pause_ad (destructive, confirmation required): Stops one Meta ad spending. Requires confirm=True.
• apply_google_campaign_action (write): Budget, target CPA and target ROAS changes. Effective immediately.
• create_amazon_negative_keywords (write): Batch exclusions, created enabled so they bite straight away.

Platforms covered. Meta Ads, Google Ads and Amazon Ads, with Shopify revenue available for context.

Nothing spends without you

• Reads are free of side effects. You can sweep the account as often as you want.
• pause_ad, pause_ad_set, pause_meta_campaign, pause_google_campaign and set_amazon_campaign_state are classed as destructive and require confirmation.
• Negative keywords on Amazon are created enabled, because an exclusion that is paused does nothing. Muze says so rather than hiding it.
• Google negative keyword management is not exposed over MCP today. On Google the levers are pausing, budgets and bidding targets.

Frequently asked questions

Can AI find wasted ad spend in my account?

Yes. Muze gives an AI assistant read access to every ad, campaign and search term across Meta, Google and Amazon, so it can list what spent, what returned, and what got worse compared with the previous period. It finds the candidates. You decide what counts as waste.

Will it pause things automatically?

Not unless you ask, and even then it asks again. Pausing tools are classed as destructive and require explicit confirmation, and pause_ad will not execute without confirm=True. The only exception is autopilot, which is off until you deliberately turn it on.

Does it work on Amazon search terms?

Yes, and that is where Amazon waste usually hides. list_amazon_search_terms returns recommendations to harvest converting terms as keywords and block the rest as negatives. You apply them one at a time with apply_amazon_search_term or dismiss them.

Can it add negative keywords on Google?

Not through the MCP today. Amazon negatives have dedicated tools, but on Google the write tools cover pausing, budgets and bidding targets. We would rather say that than let you find out mid-conversation.

How often should I run this?

The reads are cheap and change nothing, so weekly is reasonable for most accounts and daily is not excessive if you are spending seriously. get_blended_performance in particular is a non-LLM read and is documented as safe to call often.

What does it cost?

The free tier includes 25 read-only tool calls per month, which is enough to run a sweep and look at the results. Write actions such as pausing or adding negatives require a paid plan. Muze never takes a percentage of ad spend.

Read next

Other features

Or see every Muze feature.