The quality of an AI-assisted analysis depends on the quality of the question. Asking “analyse this account” gives generic answers. Specific questions, with context, data and answer format, produce useful analyses, which then need checking. It is a skill that can be trained.
The essentials in a definition
Prompts for account analysis are questions and instructions given to an AI assistant to interpret reports, find anomalies and propose hypotheses. They work best when they include context (objectives, changes, period), the data and the desired format. Account data can be read through the API, and AI guidance and NIST advise validation and protection of sensitive data.
The problem with vague questions
A question such as “what can I improve?” returns advice that suits any account. The AI does not know the business, the objective or the constraints, and fills the gaps with assumptions. The result seems sensible but not very useful.
Worse: it may present assumptions as facts.
Anatomy of a good request
State the role (“act as an account analyst”), the context (business, objective, budget, period), the data (summary or table), the concrete question (“which campaigns spend more than the maximum acquisition cost?”) and the format (short list, table, steps). Ask it to flag assumptions and uncertainties.
The more concrete the request, the more verifiable the answer.
Writing analysis requests that work
- Give role, context and data
- Ask one concrete question
- Ask for assumptions and uncertainties
- Verify against the original data
Examples of useful questions
“Compare these two weeks and state the three biggest changes and possible causes.” “Which search terms should be excluded, and why?” “What test hypotheses do you suggest, ranked by impact and effort?” “What information is missing to draw conclusions?”
Always ask for justification and references to the data supplied.
A good question is already half the analysis.
Verify and protect
Check numbers and conclusions against the original data. Be wary of unsupported statements. Do not share personal or confidential data with tools without checking privacy. Keep the requests that work as templates for the team.
Use AI as an analysis assistant, not as the decision-maker.
Decision table
| Prompt element | Example | Effect |
|---|---|---|
| Context | Objective, period, recent changes | More relevant answers |
| Data | Table with metrics per campaign | Analysis based on facts |
| Question | “Which campaigns deviated most from forecast and why?” | Clear direction |
| Format | List of hypotheses ordered by impact | Usable answer |
Common mistakes to avoid
- Giving confidential data to unapproved tools. Anonymise and use only authorised tools.
- Accepting causes without checking. Treat answers as hypotheses.
- Vague questions. “What is going on?” yields generic answers.
How to measure the result
Record how many suggested hypotheses are confirmed in the data and time saved; adjust the template prompt based on the hit rate.
Limits and exceptions
Answers can be wrong or seem convincing without being so, and assistants do not know the full business context. There are confidentiality risks when sharing data. In important decisions, always validate against the original data and have a review process. Prefer tools with adequate privacy guarantees.
Illustrative example
Illustrative example: a manager creates a template prompt with the account objective, the month’s changes and a table of metrics per campaign, and asks for three hypotheses ordered by impact and the checks for each. They validate two of the hypotheses in the data and discard the third. Analysis that took hours is ready in minutes, with stricter verification.
In practice: checkpoints
Include the period, the metric definitions and the decision you must make in each prompt.
Frequently asked questions
Can I trust AI conclusions?
With verification. Check numbers and assumptions against the data.
What data can I share?
Avoid personal and confidential data. Use aggregated summaries and validate the tool.
How do I keep the best requests?
In a shared library of templates.
What data can I give the AI?
Those the tool and company rules allow; avoid personal data and anonymise when necessary.
How do I improve the answers?
Give context, structured data, a specific question and the intended format; ask for the justification.
Good analysis prompts include context, data, a specific question and the desired format; answers are hypotheses to validate, with data protected.
Sources and recommended reading
- Google Search Central: generative AI content guidance
- NIST: AI Risk Management Framework
- Google Ads API: introduction
Last editorial review: October 2026. Informational content; always confirm current conditions in the official sources before deciding.