Can ChatGPT Create a Google Ads Campaign End to End?

There are two versions of "ChatGPT built my campaign". In one, it wrote a nice plan and you spent the afternoon clicking it into existence. In the other, the whole campaign — structure, keywords, ad and all — exists in your account by the end of the conversation, built in seconds through tool calls you approved, then parked paused for your review. The second version is real, and this page walks its actual sequence.

The answer and the required setup

ChatGPT creates campaigns end to end once two conditions hold: MCP connector support is enabled in ChatGPT (OpenAI gates this behind developer mode, plan-dependent), and the connector you add exposes write tools. Read-only connectors analyse; a write-capable one such as AdCopilot builds. The ChatGPT connection guide covers both conditions; whether ChatGPT can "manage" ads more broadly is its own question.

The build sequence: what gets created, in what order

A campaign is a tree, and the agent builds the whole tree in dependency order — each step one or more approved tool calls:

1. Campaign — name, type (search or Performance Max), bidding, daily budget. Created paused.
2. Ad group — one per keyword theme, with a default CPC bid where manual bidding applies.
3. Keywords — with explicit match types, plus the first negative keywords.
4. Responsive search ad — up to fifteen headlines and four descriptions, drafted in chat for your edits before any tool call.
5. Assets — sitelinks, callouts, structured snippets, linked to the campaign.
6. Targeting — geo targets, ad schedule, device adjustments where wanted. Location exclusions stay a UI job.

Then the agent reports the inventory: everything created, with IDs. The campaign sits paused until you enable it in the UI — the build is a reviewable draft with real structure, not live spend.

The two decisions the AI should never make alone

An agent will propose defaults for everything, and for most of the tree its defaults are fine. Two inputs are different in kind, and a well-prompted build states both up front:

• The budget ceiling. How much this campaign may spend per day is a business decision about risk and cash, not an optimisation output. Give it as a constraint; never accept it as a suggestion.
• The conversion goal. What the campaign optimises toward — which conversion action, lead quality versus volume — decides what "working" means. An agent can describe the trade-offs; only you know which one the business wants.

Everything else — structure, match types, copy themes, scheduling — is legitimately delegable, because everything else is reversible and reviewable.

A worked example: brief to built in one conversation

A realistic exchange, compressed:

You: "In account 123-456-7890, build a search campaign for our emergency plumbing service in Leeds. Budget forty a day in account currency, Maximise conversions — and confirm the phone-call conversion action is live before you start. Keep it paused."

The agent: proposes the campaign call — you approve. Proposes one ad group per theme (emergency call-outs, boiler repairs), each with phrase and exact keywords pulled from the language on your landing page — two approvals. Drafts the RSA in chat; you cut the weakest headlines, tighten one claim; it creates the ad — one approval. Sitelinks for pricing, reviews and contact — one approval. Geo target Leeds plus surrounding postcodes — one approval.

The close: an inventory with IDs, a paused campaign in the account, and a reminder of the one thing it cannot do: enable is your click, in the UI, after review.

Common failure points — and how to avoid a half-built campaign

The failure modes are mundane and all avoidable:

• No conversion tracking. A campaign optimising toward conversions that are not measured builds fine and learns nothing. Ask the agent to verify conversion actions exist before starting.
• The abandoned tree. Decline an approval mid-build without saying why, end the chat, and you own a campaign shell with no ads. Finish the sequence, or ask for a cleanup pass — "list everything created this conversation, pause anything incomplete".
• Approval fatigue. Six approvals invite skimming, and skimmed approvals defeat their purpose. Read the budget call and the final-copy call carefully; those two carry the money.
• Enabling from the chat trigger-happy. The paused default only protects you if the review actually happens. Open the campaign in the UI, check it as you would a junior's work, then enable.

When to build in the UI instead

The honest boundary: use the interface when the work is primarily visual or exploratory — auditioning image assets, dragging through the keyword-planner's suggestions, or making the strategic bidding decisions the connector deliberately leaves in the UI. Use the agent when you know what the campaign should be and want the distance between brief and built collapsed to one conversation. Most operators converge on brief-driven builds through chat, with a UI review before enabling — which is the workflow the connector setup tutorial gets you to in an afternoon.