Using the Ads Agents
The five Ad Manager agents — per-campaign, portfolio and cross-platform — with one composer, the Tune menu, the drawers, long conversations, and why every change still needs you.
Five chat assistants live inside GrowthAd Manager. An Ads Agent is scoped to one campaign (Google or Meta) and knows its metrics, ad groups or ad sets, creatives and history. The Portfolio Agent (Google), the Meta Portfolio Agent and the Cross-Platform Agent reason across the campaigns you select on the matching Ad Manager tab. All five analyse and propose. None of them executes anything against your spend.
Before you start
Every agent is owner-only and needs the provider connected under SettingsIntegrations. The three portfolio-type agents need at least one campaign selected in the Campaign scope drawer — otherwise the composer stays disabled and the panel says which platform's campaigns it wants. The agents are a web-portal feature; the iOS admin shell hides their buttons.
Which agent to open
| Ads Agent | Portfolio Agent | Meta Portfolio Agent | Cross-Platform Agent | |
|---|---|---|---|---|
| Scope | One Google or Meta campaign | The Google campaigns you select | The Meta campaigns you select | The Google and Meta campaigns you select |
| Where | The floating Google Ads Agent or Meta Ads Agent button on a campaign detail page | Ad Manager → Google Ads tab → Google Portfolio Agent | Ad Manager → Meta Ads tab → Meta Portfolio Agent | Ad Manager → All tab → Cross-Platform Agent |
| Good at | Search terms, negatives, RSA copy, ad sets, creatives, placements, one campaign's CPA | Budget reallocation, comparing Google campaigns, pause candidates | Budget reallocation, frequency, placements, audience overlap, pausing or enabling several campaigns, bid strategies per ad set | Where each platform is strongest, a coordinated reallocation across both |
One sheet, one composer
Every agent opens as a sheet with the same shape: a header naming the agent and its scope, a chat column, and one composer at the bottom. Nothing sits open above the chat any more — the context that used to live there is behind drawers you open from the tool strip:
| Drawer | What it holds | Where |
|---|---|---|
| Research brief | The campaign research brief (below) with its Research runs, the project link and the live-data health strip | Google Ads Agent |
| Context & knowledge | The knowledge store, the changelog and the live Google Ads data | Google Ads Agent |
| Conversations | Your saved threads for this campaign — switch, rename, delete, export | Ads Agents |
| Action history | Every proposal, what you approved or rejected, and when | Ads Agents |
| Campaign scope | The campaigns the agent may read and reason about, grouped by client — opened from the N campaigns · N accounts chip | The three portfolio-type agents |
The empty stage shows the platform mark, one line, and at most four prompt chips. Every chip drafts into the composer; nothing sends until you press send.
The Tune menu
The composer carries the model picker and one Tune menu:
| Setting | What it changes | Default |
|---|---|---|
| Model | Which model answers. A model whose row reads Analysis only cannot propose campaign changes at all. | Default model |
| Effort | How much reasoning to spend: Low, Medium, High, X-High or Max. Higher takes longer. | X-High on a campaign, High on a portfolio |
| Web research | Whether the agent may look things up on the public web and cite what it found. Auto, On or Off. Greyed out for models that cannot research. | Auto |
| Thinking | Extended thinking on the turn. Shown as a fold above the reply. | On, on all five agents |
| Playbooks / More prompts | Drafts a saved long-form prompt, or one of the contextual prompts the four-chip stage could not show, into the composer. | — |
Analysis-only models cannot propose changes
Pick one of those and you will get good reading of the data and no actionable proposals. The composer footnote says so — on the Google Ads Agent: Analysis only — this model reviews the campaign and reports findings; switch to a Claude model to propose budget, bid or status changes. If you wanted a recommendation you can approve, switch model first. A conversation that already carries web research or another model's reasoning cannot continue on an analysis-only model: the agent asks you to start a new conversation for it or to switch back.
Playbooks and quick prompts
The Google Ads Agent and the Portfolio Agent share six playbooks that cover the recurring jobs: Complete optimization, Budget and bidding, Search terms and negatives, Ads and landing pages, Competitive research and Verify and learn. The Google Ads Agent's research modes (Full campaign buildout, Search terms & negatives, RSA and landing pages, Competitor gap, Bidding and budget, Change audit & learning) are the Research runs inside the Research brief drawer, not a Tune entry.
Quick prompts are one click each and condition-aware — Improve low CTR, Fix zero conversions and Reduce high CPA appear when they apply. The Meta Ads Agent has its own (Analyze performance, Frequency check, Cost analysis …), and each portfolio-type agent has a set for its tab.
The campaign research brief
This is the highest-value thing on the Google campaign page and the most frequently skipped. It is structured business context the agent reads before it answers anything, and the difference between a generic answer and a useful one. It lives in the Research brief drawer.
| Field | What to put in it |
|---|---|
| Core services | What you actually sell — the placeholder suggests mobile detailing, tree service, vacation rentals |
| Niche / market and Landing URLs | Where you compete and where traffic should go |
| Service areas | Where you serve |
| Competitors | Who you lose to |
| Differentiators | Why someone picks you |
| Exclusions | Services you do not offer, so the agent stops bidding on them |
| Negative themes | The poor-fit searches that waste money |
| Compliance notes | What you may not say |
| Landing page notes | Which pages to prefer and which to avoid |
| Positioning, Ideal customer profile (ICP) and Audience notes | Who you want |
| Proof points | What you can claim |
| Pricing posture | Premium, mid, value |
| Brand voice | How ad copy should sound |
| Seasonality and Common objections | When demand moves and what stops people buying |
Exclusions earns its keep immediately
Most wasted spend in a small account is traffic for something the business does not do. Filling that one field turns find wasted spend from a guess into a list.
What the agents can read
Beyond your metrics, both campaign agents read your check-in schedule for the campaign, the provider's own change history (Google) or account activity (Meta), and — with Web research on — public sources they cite. The Google Ads Agent resolves place names into real geo targets before it proposes location targeting; the Meta Ads Agent resolves them into Meta location keys and can read the account's conversion setup (pixels, custom conversions and each ad set's promoted object) before recommending a conversion-optimised change.
Long conversations
Conversations do not lose their thread. Every agent pages the whole history, summarises the oldest turns when a conversation outgrows the model's window (an Earlier entries summarized chip marks a turn that carried a summary) and keeps trimming inside a long, tool-heavy turn so the request itself is never dropped. Web research and thinking survive across turns. The thread opens on its newest messages; Load earlier at the top pages the older ones in, and on the Google Ads Agent and the Google Portfolio Agent it also counts the proposals those earlier turns carry.
New conversation starts a fresh thread on every agent. On the Ads Agents, the Meta Portfolio Agent and the Cross-Platform Agent the previous thread is retired, not deleted. On the Google Portfolio Agent it is scoped to the current campaign selection: that selection starts again from a blank thread while every other selection keeps its own history. Stop ends the turn immediately.
Approval, always
The agent proposes. You approve. There is no path where a chat message moves money on its own. Each proposal is one card: what changes as before → after rows in the account's units — money, ROAS and click-through rows also show the size of the change — one status chip, Why this change, and Approve / Reject. The portfolio reads — compare campaigns, budget allocation, the portfolio and cross-platform overviews and the platform comparison — come back as a table under the turn that called them (inside that turn's Research and tools fold when it has one) the next time the thread loads: reopen the sheet, change the selection or reload the page. They do not appear while the turn is still running. On the Ads Agents, several proposals in one turn add an Approve all row; the portfolio-type agents take each proposal on its own card or in the change-set review sheet.
No autonomous optimisation runs by default
Durable Google Ads optimisation tasks — the research, plan, approve, execute, verify loop — ship behind a flag that is off by default. With it off, chats behave like ordinary approvals and create no runnable tasks. On the Meta side, the auto-approve rules card stores rules that nothing evaluates: every action still requires a manual approval. Do not plan around unattended changes.
Approved changes go through the same staged review and the same frozen money guards as an edit you made by hand. New Meta campaigns, ad sets and ads are always created paused — you enable them after review. A change to a Google budget that other campaigns share is refused at approval time; the agent says so instead of re-proposing it. All five agents see your rejections in their recent-actions context and do not propose the same change again for thirty days unless you ask for it — if one tries, the proposal is refused and the agent is told the date you rejected it, so it explains that instead of retrying. Your own edits in the change-set review sheet are never held back by this rule.
History
Both Ads Agents keep an Action history drawer — what was proposed, what you approved or rejected, and when. The portfolio-type agents show their proposals and outcomes in the conversation itself. Use the history before asking what changed recently? — it is the record of your own decisions, not the provider's change log (ask for the change history or account activity for that).
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| The agent explains but never proposes anything actionable | An analysis-only model is selected | Switch model; web research is also greyed out for those |
| A portfolio agent's composer is disabled | No campaigns selected | Open the Campaign scope drawer and pick at least one campaign |
| One or more selected campaigns are unavailable or outside your … scope | A selected campaign is gone, on the other platform, or not yours | Re-open the Campaign scope drawer and pick again |
| The agent asks you to start a new conversation for this model | The thread carries research or another model's reasoning | Start a new conversation, or switch back to a Claude model |
| Answers are generic and miss your business | The campaign research brief is empty | Fill it in, especially Exclusions and Negative themes |
| An approved change did not reach the provider | Approval and application are separate steps | Press Review & Apply on the staged-changes bar, then Apply in the review sheet |
| A proposal was rejected as too large | It crossed a frozen money guard | Step it up over smaller edits |
| A Meta bid-strategy proposal says the ad set is not synced | The ad set is not yet a synced member of the campaign | Press Sync Meta campaigns, then ask again |
| A Meta approval stopped working overnight | Meta approvals expire after 24 hours | Ask the agent to re-propose |
| You expected a rule to act while you were away | The Meta auto-approve rules card is dormant | Every action needs a manual approval |