Why ChatGPT's Google Ads advice is usually useless
Ask ChatGPT "why is my cost per conversion going up?" and you will get a competent, generic answer: check your search terms, review your bidding strategy, look at auction insights, consider your landing page. All true. None of it tells you what is happening in your account this week.
The reason is simple. A plain chat has no access to your data. It is reasoning about Google Ads in the abstract, from documentation and forum posts, not about the campaign that quietly ate 40% of your budget on Tuesday.
The usual workaround is copy-paste: export a report, screenshot a dashboard, paste a table into the chat. That works once. It does not survive contact with a real week, because by the time you have assembled the context, you could have read the numbers yourself. And a pasted table strips exactly the things that matter — which conversion action is which, whether a campaign is budget-limited, what changed three days ago.
What changes when ChatGPT can read the account
Connecting ChatGPT to Google Ads through a connector changes the shape of the work. Instead of describing your account, you ask about it:
- "Which campaigns are pacing behind and why?"
- "Show me last week's spend broken down by campaign, with the change versus the week before."
- "What search terms cost me money and produced nothing in the last 30 days?"
Each answer is read live from the account at the moment you ask. Nothing is cached, nothing is summarised from last month's export. And because the same connection can write as well as read, the follow-up — "add those as negatives" — is one message rather than an hour of clicking.
If you have not set this up yet, the step-by-step is here: connect ChatGPT to Google Ads. It takes about five minutes and needs no code. There is also a two-minute video of the whole thing on Google Ads and ChatGPT, connected, which compares it against the other ways of getting your data into a chat.
The four jobs it genuinely does well
Not everything in paid search should be handed to a chat. These four are where the leverage actually is.
1. Reporting you would otherwise skip
The reports nobody builds are the ones that would catch problems early: week-over-week pacing, spend concentration, the gap between a campaign's budget and what it actually spent. These are trivial to ask for and tedious to build. Ask in plain language, get the numbers, move on.
The useful trick is asking for the *comparison*, not the number. "How much did I spend last week" is a fact. "How much did I spend last week versus the week before, by campaign, and which changes are outside normal variance" is a diagnosis.
2. Search-term cleanup
Every account leaks money through irrelevant queries. Finding them means reading a long search-terms report with tired eyes, and the report gets longer every week.
This is the single highest-return task to delegate. Ask for search terms with spend above a threshold and zero conversions over a meaningful window, get a list, decide what is genuinely irrelevant, and have the negatives added. The judgement stays yours — some expensive non-converting queries are exactly the traffic you want, and only you know that.
3. Campaign builds as drafts
Building a search campaign is mostly structure: campaign, ad groups, keywords with the right match types, responsive search ads with enough distinct headlines, sitelinks. It is knowable work, and it is slow.
Asked to build one, a connected assistant produces the whole structure — paused. That distinction is everything. A build is a draft you review, not a launch. You read the keywords, cut the ones that are wrong, fix the ad copy that sounds like a robot wrote it, then turn it on yourself.
4. Budget and bid changes, proposed not applied
"Cut the budget on anything above a €40 cost per lead" is a sentence. Executing it across a dozen campaigns is twenty minutes of careful clicking, and careful clicking is where mistakes live.
The right shape here is a proposal: the assistant tells you which campaigns match, what it would change, and what the change implies. You approve. Nothing that moves money should happen because a chat inferred that you wanted it.
The jobs it should not own
Being honest about the boundary is what makes the rest trustworthy.
Strategy and positioning. What you sell, to whom, at what margin, and why anyone should choose you — none of that is in the ad account, so none of it can be read from the ad account.
Conversion tracking design. Deciding what counts as a conversion and what it is worth is a business decision with consequences everywhere downstream. An assistant can tell you your tracking looks broken. It should not decide what you count.
Anything with legal or brand risk. Claims in ad copy, competitor comparisons, regulated-industry language. Generated copy is a first draft; someone accountable reads it.
Big structural bets. Migrating an account to a different campaign type, or changing bidding strategy across the board, deserves a human plan and a rollback path.
A first-week routine that actually sticks
Automation projects die from ambition. Start with a routine small enough to survive a busy week.
Monday, five minutes. Ask for last week's spend and results by campaign, with the week-over-week change, and anything pacing unusually. Read it. Do nothing else.
Wednesday, ten minutes. Ask for wasted search terms over the last 14 days. Approve the negatives you agree with. Reject the rest.
Friday, ten minutes. Ask what changed in the account this week and whether it did what you expected. This is the habit that catches your own mistakes.
That is 25 minutes a week, and it is more consistent account hygiene than most accounts get.
Once the routine holds, add builds: next time you would have booked an afternoon to launch a campaign, ask for the draft first and spend the afternoon editing instead of assembling.
The guardrails that make this safe
Handing account access to a chat should make you nervous. The reasons it is safe are specific, and worth checking in whatever tool you use:
- New campaigns are created paused. A build cannot start spending on its own.
- Changes that move money are proposed first. Budgets, bids, new audiences — you approve before anything applies.
- No platform passwords. Access comes from Google's own sign-in, the same way any approved marketing tool connects, and you revoke it from your Google Account whenever you want.
- Reading is always safe. A report request cannot alter a campaign, so the whole diagnostic half of this workflow carries no risk at all.
If a tool cannot tell you clearly which of these it does, that is the answer to whether you should connect it.
Where to start
Connect the account, then run the Monday question for two weeks before you change anything. You will learn more about your own account from two weeks of consistent week-over-week comparisons than from a month of dashboard staring.
The setup guide is here: connect ChatGPT to Google Ads. If you use Claude instead, the Claude version is fewer steps. Both are covered in the documentation, along with what the connection can and cannot do.