ChatGPT prompts for Google Ads: what works and what does not

Prompts that work without account data

For any task that does not require current account state, ChatGPT performs well. Write brief copy to generate variants, describe your audience to get targeting angle suggestions, share your landing page copy to get headline ideas that match it. These tasks are text-in, text-out and the model is genuinely useful.

• "Write 10 Google Ads RSA headlines for a B2B SaaS tool that reduces invoice processing time. Under 30 characters each. No superlatives."
• "My Google Ads description is [X]. Write 5 variants that test a different benefit."
• "Given this landing page intro paragraph, what three pain points should my ad copy lead with?"
• "What negative keywords should I add for a campaign selling enterprise payroll software?"

Where prompts break down without data

The most common mistake is asking ChatGPT diagnostic questions about a specific account. "My ROAS dropped from 3.2 to 1.9 this week, what happened?" is a question the model cannot answer without seeing the account. It will produce a plausible list of possible causes — seasonality, Quality Score change, auction pressure, landing page issues — but it cannot tell you which one is actually true.

This is not a model failure. It is an information failure. The model is reasoning from your description of symptoms, not from the data. The result is generic advice dressed as diagnosis.

Prompts that work with a connected tool

When ChatGPT has access to live account data through a GPT Action or MCP connector, diagnostic questions become answerable. The model is now reasoning over actual campaign state, not your text description of it.

With SpendSignoff connected, the same ROAS question produces a different kind of answer: "Your ROAS dropped from 3.2 to 1.9 between May 1 and May 7. The Brand Search campaign accounts for 60% of the drop. Impressions held steady but CTR fell from 4.1% to 2.3%, suggesting ad relevance or a Quality Score change. I am queuing a draft to update the top RSA headline for your review."

That is a diagnostic, a cause hypothesis, and a proposed action — all grounded in account numbers the model actually read.

Prompt patterns for connected Google Ads work

These prompts work because they have data to work with. They are also specific: a clear metric, a threshold, a time range. Vague prompts like "optimize my account" are as useless with live data as without it — the model needs a target to reason toward.

• "Read my active campaigns and tell me which ones are pacing over budget this week."
• "Which ad groups have a CTR below 1% and more than 500 impressions in the last 14 days?"
• "Draft a 10% budget cut on any campaign with a CPA above $60 and less than 2.0 ROAS."
• "Find any ad groups with no negative keywords and suggest five to add."
• "Compare ROAS across all active campaigns and identify the top and bottom performers."