Performance Max and AI: what you can still control

What Performance Max actually automates

A Performance Max campaign spans Search, Shopping, YouTube, Display, Discover, Gmail and Maps from a single campaign, and Google decides which of those an impression goes to, what the ad looks like when it gets there, who sees it and what to bid. You supply a budget, a goal, a set of assets, some audience hints and — for retail — a product feed.

That is the trade, and it is a real one: it works, it reaches inventory that would take four campaign types to reach separately, and it removes most of the levers people used to pull. The frustration with PMax is not that it performs badly. It is that when it does, the account gives you fewer places to look. As set out on the automation page, choosing PMax is choosing automation — the question is what is left over.

What you can actually see

Reliably visible, in the interface and through the API:

• Campaign-level performance. Cost, conversions, conversion value, impressions and clicks, segmented by time, device and geography like any campaign.
• Asset group performance. Which asset group is carrying the campaign, which is dormant, and which never got out of the starting blocks.
• Asset-level ratings. Individual assets are labelled — low, good, best, or still learning. These are relative judgements within an asset group rather than statistics you can bank on, and a “low” asset in a strong group is not the same finding as a low asset in a weak one.
• Search categories and themes. Grouped query insight rather than a raw search term list. Less than a Search campaign gives you, more than nothing, and steadily improving.
• Serving status and diagnostics. Why a campaign or asset group is not running, which is the first thing to check when spend goes flat.

Structurally harder or impossible to get:

• A per-query report. There is no PMax equivalent of the Search campaign's search term view.
• Channel-level control. You cannot bid differently on YouTube than on Search inside a PMax campaign, whatever reporting exists. There is no lever, only a view.
• Conversion credit per asset. You learn that an asset group converted; you do not learn that the third headline did it.

What you can actually control

The last column is the honest bit: which of these AdCopilot's connector can reach, and which remain interface work.

What an AI agent can genuinely help with

Four things, in descending order of how confident we are that they are worth your time.

1. Inventory and coverage. Which asset groups are running on the bare minimum, which have no video and are therefore being given one Google assembled, which assets are rated low, which asset groups are paused and forgotten. This is pure reading, it takes one pass, and it is the audit nobody does because clicking through asset groups one at a time is miserable. In API terms it is a single query:

   SELECT asset_group.name,
          asset_group.status,
          asset_group_asset.field_type,
          asset_group_asset.performance_label
   FROM   asset_group_asset
   WHERE  campaign.advertising_channel_type = 'PERFORMANCE_MAX'

2. Drafting assets at volume. Thirty characters for a headline, ninety for a long headline, written fifteen different ways for four audiences. Writing inside a hard constraint is something a language model does well and most people do reluctantly. Deciding which drafts sound like your company is the part that stays human.
3. Comparing PMax with everything else. Read the PMax campaign and the brand Search campaign over the same window and ask what happened to brand impression share when PMax started. The answer usually settles an argument that has been running for weeks.
4. Proposing exclusions. Search category insight plus your own knowledge of who never buys produces a list of negatives worth adding — the same job described on the negative keywords page, with the caveat that PMax gives the agent less to read.

Building a PMax campaign through a connector

This is the part most Google Ads connectors do not do at all, so it is worth being specific about the sequence. Assets exist before the things that reference them, which is why it looks back to front:

1. Create the assets. Text assets for headlines, long headlines and descriptions; image assets for the landscape, square and logo slots; a YouTube video asset if you have one. Each returns a resource name.
2. Create the campaign. create_performance_max_campaign takes a budget, a business name asset, a logo asset, a bid strategy of maximise conversions or maximise conversion value with an optional target, and a flag to opt out of final URL expansion. It creates the campaign paused unless you say otherwise.
3. Create the asset group. create_asset_group links the assets to the campaign in one batch, enforcing the minimums: three headlines, two descriptions, one long headline, a landscape image, a square image and a logo.
4. Fill it out. add_assets_to_asset_group adds more later — extra headlines, a second video, a landscape logo.
5. Finish in the interface. Audience signals, search themes, listing groups for a retail feed and brand exclusions are not in the tool set. Then set the campaign live yourself.

What an agent cannot do here

• See inside the bidding. Nobody outside Google can. An agent that claims to be optimising PMax bids is optimising a target, which you could have set yourself.
• Give you levers that do not exist. No channel-level budgets, no placement bidding, no query-level control. A connector exposes the API; it cannot exceed it.
• Remove an asset. remove_asset_from_asset_group is one of the four destructive tools never exposed to a hosted connector. This one is not only about safety: removing an asset discards its learning, and adding it back does not restore it. Pausing an asset group and building a new one is usually the better move anyway.
• Judge creative. A model can tell you an asset is rated low. It cannot tell you the photograph is off-brand, that the claim needs legal sign-off, or that the headline reads like a competitor.
• Rescue a campaign with a measurement problem. PMax optimises hard toward whatever you told it to value. If the conversion setup counts the wrong things, PMax will find you more of the wrong things faster than any other campaign type. Start with the audit, not with the asset groups.

The questions worth asking about a PMax campaign

Paste these once your assistant is connected. Every one of them is a read.

• “List every asset group in this account with its status and how many headlines, descriptions, images and videos it has.”
• “Which assets are rated low, and in which asset groups?”
• “Compare this PMax campaign with my brand Search campaign over the last 90 days — cost, conversions and conversion value each.”
• “What search categories is the PMax campaign appearing for, and which of them look irrelevant?”
• “Is anything preventing this campaign or asset group from serving?”
• “What changed on this campaign in the last 30 days, and did conversion value move with it?”

All of those run on the read-only tier — three tools, none of which can change anything — which is what the free one-week pilot gives you. Writing comes later, and the full tool list is on the MCP server page.

Frequently asked questions