ChatGPT with no account access is a copywriting intern: fast, articulate, and blind. ChatGPT connected to your Google Ads account over MCP is an operator: it reads the same live data the interface shows you, answers questions with the evidence attached, and — on your command — makes the change in seconds, whether that is adding negatives, pushing an RSA, or moving a budget. Almost every disappointment people report with ChatGPT for Google Ads comes from asking connected questions in intern mode. This guide covers both modes honestly, then walks through four connected workflows you can run this week.
What ChatGPT can do with no account access — and where that ceiling is
Unconnected ChatGPT is genuinely good at three things. It drafts ad copy against a brief. It explains platform mechanics — match types, bid strategies, policy areas — at whatever depth you ask. And it critiques whatever you paste: a screenshot of a campaign table, an exported search-terms CSV, a competitor's landing page.
The ceiling is that everything runs on what you carry to it. That has three costs:
- Staleness. The export you paste describes the account as it was when you exported. Every follow-up question runs against the same ageing snapshot.
- Truncation. Paste limits force you to pre-decide what matters. The wasted spend hiding in row 4,000 of the search terms report never makes it into the conversation.
- No execution. ChatGPT can tell you to add forty negative keywords. You still spend the next twenty minutes adding them, and somewhere around number twenty-five you stop.
The same jobs, seen from both sides of the connector:
| Task | Unconnected ChatGPT | Connected over MCP |
|---|---|---|
| Draft ad copy | Good — works from your brief | Better — grounded in the queries that convert |
| Explain a metric or setting | Good | The same, with your account as the worked example |
| Find wasted spend | Limited to what your export included | Reads the full account, ranks leaks by cost |
| Answer "why did CPA rise?" | Educated guesswork | Evidence from terms, change history, auctions |
| Make the change | You implement it by hand | Proposes the exact call; you approve; it executes in seconds, logged |
If your use of ChatGPT for Google Ads is a loop of export, paste, read advice, go implement — the connector is the upgrade, because it deletes all three costs at once.
Connecting ChatGPT to your account: the five-minute version
The mechanism is an MCP connector — a URL that gives ChatGPT tools for reading and writing your Google Ads data, authorised by your own Google sign-in rather than a shared password. You need three things: a Google login with access to the ad account, ChatGPT with developer mode switched on (the Free plan is enough), and the connector URL itself. Nothing installs anywhere.
- In ChatGPT, switch on developer mode (Settings, then Security and login), then go to Plugins, press +, choose Create app, then Create MCP App, and paste the AdCopilot URL into the form. The ChatGPT connection guide has every click, with screenshots.
- Press Sign in with AdCopilot and approve. The URL is public and the same for everyone; your own sign-in — and the Google account you connected to AdCopilot — is what authorises you.
- Start a new chat with @AdCopilot and ask a read question — "list my campaigns with spend this month". Tool calls appear in the conversation as they run.
Two defaults worth knowing before your first session. Reads change nothing, but writes surface an approval step — you see the exact change before it executes. And deletion is not possible at all: the destructive tools are absent from the connector server-side, which is a stronger guarantee than any instruction you could type. The details live on the security page.
The trial runs seven days on the full Pro plan for a new workspace, no card required — then it drops to the Free plan (one account), not a lockout.
Workflow 1: the Monday performance readout
The prompt:
Compare last week (Mon-Sun) with the week before for account
123-456-7890. Spend, conversions, cost per conversion, by campaign.
Flag anything that moved by a lot in either direction, and tell me
the most likely cause for each flag, citing the data you used.
What comes back is a table plus narrative — and the narrative is the point. A dashboard shows you that CPA rose; ChatGPT connected to the account can check whether search terms shifted, whether a budget cap started binding, or whether one campaign's conversions simply lagged.
The habit that makes this workflow compound: ask the follow-up. "Which search terms drove the extra spend in Campaign B?" costs one sentence, not another export. The follow-up question is the entire advantage of a conversation over a dashboard.
Make it a standing Monday slot and keep the prompt fixed — same comparison, same thresholds — so weeks stay comparable. The useful follow-ups recur: which ad groups drove the movement, whether a budget cap bound on the strongest days, what appears in change history for the period, and which of the flags deserves action rather than observation. Ten minutes on most Mondays; the eventful ones announce themselves.
Workflow 2: search terms triage and negative keywords
The search terms report is where money leaks, and it is exactly the kind of long, dull table a model reads better than a tired human. Ask:
Pull search terms for the last 30 days across all search campaigns.
Show me terms with meaningful spend and zero conversions, grouped by
theme. Recommend negative keywords, with the match type and the level
(campaign or ad group) you would add each at, and say why.
Review the list — this is where your judgement earns its keep, because "free" might be a waste theme for one business and the entire market for another. Then: "Add the first eight as phrase-match negatives at the campaign level." The client shows the exact tool call with every term listed. One approval, and the change is live under your own Google sign-in, attributed to you in change history.
This is the best first write to delegate because the blast radius is small and the payoff is immediate. There is a deeper treatment in the negative keyword automation guide.
Workflow 3: drafting and pushing RSA variants
Copy is where unconnected ChatGPT already looked good. Connected, two things improve. The drafts are grounded — "write headlines using the phrasing from my top converting search terms" produces copy in the customer's language, not the brand deck's. And the copy-paste chain disappears: the model drafts fifteen headlines and four descriptions, you edit in chat, and it pushes the ad into the ad group you name, with your approval on the write.
Ask for the new RSA to arrive paused if you want a final look in the interface before it can serve. Additions that start paused cannot spend a cent until you enable them — a cheap safety habit while you calibrate trust.
Workflow 4: budget pacing checks before month-end
Google spends to daily budgets; you answer for a monthly number. Around the 20th of the month, ask:
For each campaign: month-to-date spend, current daily run rate over
the last 7 days, and projected month-end total at that rate. Compare
against a monthly plan of [your number] and tell me which campaigns
will land over or under, and by how much.
The model does the run-rate arithmetic you would otherwise do in a spreadsheet, and because it can also read performance, it can propose where a surplus should move — toward the campaign converting cheapest with headroom, not just the loudest one. Budget changes go through the same approval step as everything else.
What to keep away from ChatGPT entirely
An honest boundary list, because a tool sold as good at everything is trustworthy at nothing:
- Strategy that depends on facts outside the account. Margins, sales capacity, cash position, the client's politics. ChatGPT will happily optimise toward the wrong goal with great confidence if you let it.
- Unverified numbers in client-facing documents. Spot-check figures against the interface before they leave the building. Same API, so they should match — when they do not, the cause is usually date range, timezone or conversion lag, and you want to find that before the client does.
- Irreversible actions. Nothing irreversible should be delegated to any AI system. This is the one item the connector enforces for you: deletion is refused server-side, in any letter case, so the category of unrecoverable mistake is closed by architecture rather than by vigilance.
Whether ChatGPT specifically — rather than AI generally — is allowed to manage accounts is a question with some nuance to it; the can-ChatGPT-manage-Google-Ads answer covers policy and practice. And once the connector is in place, the quality of your results tracks the quality of your prompts almost linearly — the prompt library is the fastest way to raise that floor.
The split to carry away: delegate the reading, the aggregation and the drafting without hesitation — that work has no blast radius, and the model is faster at all of it than anyone on your team. Keep the judgement calls, and keep them cheap: one approval click at the exact moments money moves. That division is what the connector's defaults are built around, and it is the difference between using ChatGPT for Google Ads and merely chatting about them.