How to prepare campaign data for AI agent decisions (2026 guide)

Before an AI agent can recommend a campaign change, it needs the information you’d use to make that decision yourself. This guide shows performance marketers and agencies what to provide and how to check the reasoning behind each proposal.

Your paid campaigns are bringing in fewer leads, and your sales team wants answers. You ask an AI agent to investigate. It spots a rising CPA in Google Ads and recommends moving budget to Meta Ads, where conversions look cheaper.

But what if those Meta Ads conversions are low-quality leads? Or you changed Google Ads bidding three days ago, and the results haven’t caught up yet?

An agent can give you a convincing recommendation without having enough information to make a sound one. Performance figures only tell part of the story. It also needs to know what counts as a valuable conversion, what’s running now, and which recent changes could explain the results.

Here’s how to prepare campaign data for AI agent decisions: which inputs each decision needs, how to account for conversion delays, and when the agent should hold off. You’ll also get an example proposal format and a practical way to test its judgment on past decisions before approving live changes.

Key takeaways

• Match the evidence to the decision. A budget increase needs different inputs from a creative refresh.
• Provide current settings, stable campaign IDs, and a dated change record alongside performance history.
• Use a reporting period that excludes days still collecting conversions, for example 14 complete days ending seven days ago, and reassess about 21 days after a change.
• Make “no recommendation yet” an acceptable answer when data is missing or unreliable, or when a recent change hasn’t had time to show results. Poor marketing data quality leads to poor recommendations.
• Require a proposal that shows the evidence, settings, spend impact, and uncertainty. Test this format on past decisions before allowing live changes.
• Verdict: Give an AI agent the evidence you’d need to make the same decision, and require it to justify any proposed change before you approve it.

What data does an agent need for each campaign decision?

Start with the decision, then work back to the evidence. A budget increase needs results that justify more spend; a creative refresh needs detail on which ads are losing effectiveness.

For each decision, list what you’d check yourself before acting. Use those checks to decide what to give the agent.

Your weekly report is a starting point, but it may not contain everything the agent needs. Include the settings relevant to the decision and flag results still awaiting conversions. For a creative review, provide ad-level results: campaign totals can hide which ads are slipping.

Make sure the agent also understands what each metric means. Don’t assume it knows how each platform defines its PPC metrics. A “conversion” in Meta Ads may count a different action or use a different attribution window from one in Google Ads. Giving both fields the same name doesn’t make them equivalent. Define what each metric counts before asking the agent to compare results.

To give the agent access to these inputs, you can use Supermetrics, the marketing intelligence platform that connects advertising, analytics and CRM data to reporting tools and AI assistants.

Supermetrics for Claude and Supermetrics for ChatGPT let you query connected sources such as Google Ads, Meta Ads, LinkedIn Ads, and GA4 (Google Analytics 4), without uploading a new CSV for every investigation. Here’s how to connect Claude or ChatGPT to your marketing data.

Supermetrics retrieves the latest data each source has reported. You still need to account for incomplete results and supply the business context behind the decision.

Why does an agent need access to current campaign settings?

Performance history shows how a campaign has performed. Current settings show how it’s configured now. The agent needs both to recommend a change that makes sense.

Say you raised a campaign’s budget on Monday. If the agent only sees the previous two weeks’ results, it could recommend another increase before you’ve had time to assess the first. The current budget and a dated record of the change help it avoid that mistake.

Give the agent current settings alongside performance data for every campaign it reviews:

• Account ID and campaign ID
• Campaign status, such as enabled or paused
• Daily or lifetime budget
• Bid strategy, plus any target CPA (cost per acquisition) or target ROAS (return on ad spend)
• Targeting settings relevant to the decision, such as locations, audiences or ad schedules

Use stable IDs to identify campaigns. Names can change from “Brand_US” to “Brand_NA,” while the ID stays the same. Treat campaign names as labels and tie every proposal to the account and campaign IDs. Consistent campaign naming conventions still help people read the reports.

Available settings vary by platform and campaign type. Check which fields the agent can retrieve through Supermetrics, then provide any missing information it needs for the decision. It should use the actual settings, rather than guess them from past performance.

When is performance data ready for a decision?

Use a reporting period that allows for delayed conversions, with enough activity to support the decision. A falling click-through rate may warrant a creative review before there are enough conversions to justify more budget.

How should the agent account for conversion lag and incomplete periods?

For CPA or ROAS decisions, exclude recent days if conversions are still arriving. Label provisional results and use comparable reporting periods.

Someone who clicked yesterday might buy from you next week. However, if that revenue is attributed to the click date, yesterday’s ROAS will change when the conversion appears.

Several factors can affect how long results take to settle:

• Attribution windows: a longer window allows later conversions to be credited to earlier ad interactions.
• Reporting delays: offline or CRM imports may arrive in batches rather than immediately.
• Sales cycles: B2B attribution has to cover clicks that turn into qualified opportunities weeks or months later.
• Adjustments: refunds and cancellations can change recorded revenue after an order is placed.

Exclude the current day if it is incomplete. Don’t compare half of this week with a full previous week.

Make the reporting rules explicit:

1. Choose a reporting period for each task. For example, review 14 complete days of results ending seven days ago. This gives you a full two weeks of data while excluding the most recent seven days to allow for delayed conversions.
2. Base the excluded days on your own conversion and reporting delays. Seven days is an example, not a default for every account.
3. Label any results that are still provisional.
4. Ask the agent to state the reporting period in every recommendation.

What minimum evidence thresholds should you set?

Set a minimum amount of evidence for each decision, based on its cost and how much performance fluctuates. Investigating a drop can happen immediately; committing more budget needs stronger evidence.

For example, you could require:

• At least 50 conversions in a reporting period that allows for conversion delays, the threshold in the example proposal below, before recommending a budget increase.
• Enough spend and time to distinguish sustained underperformance from a short-term fluctuation before pausing.
• A minimum number of impressions per ad, plus evidence of sustained decline, before recommending creative refreshes.
• Enough campaign history to identify which elements are worth adapting to another platform.

For low-volume accounts, use a longer window or ask for a summary without a recommendation until the threshold is met.

Each proposal should name the threshold and show the figures that meet it.

When should the agent hold off?

The agent should hold off if the required data is missing, sources disagree, campaigns haven’t met minimum evidence thresholds, or budget and bidding changes are too recent to assess properly.

Make “no recommendation yet” an acceptable answer. Tell the agent to hold off when:

• Required data is missing because a connection failed or a tracking gap leaves the results unreliable.
• Sources disagree in a way that different conversion definitions, attribution settings, or reporting dates don’t explain.
• The campaign hasn’t met the minimum evidence threshold for the decision.
• A budget or bidding change is too recent to assess under your reporting rules.
• The campaign or ad set is still in the ad platform’s learning phase after a significant edit, so results don’t yet reflect settled performance.

When the agent holds off, it should explain what’s missing and what to check next. For example: “Google Ads conversions rose, but CRM-qualified leads fell. Before recommending a budget change, check which conversion actions Google Ads counts and whether the CRM results cover the same period.”

How do you record campaign changes and keep the agent’s inputs current?

A dated change record helps the agent interpret performance and avoid repeating or reversing a recent edit. Keep it alongside the current campaign settings.

If ROAS improved during a promotion, the agent needs to know that before treating the lift as a reason to increase the budget.

Record when these changes happened and which campaigns they affected:

• Budget changes, with old and new values
• Targeting and bid strategy changes
• Creative launches and removals
• Tracking changes, such as a new conversion action or a tag fix
• Promotions, product launches, outages and seasonal events

Analyze paid ad performance before and after a change, allowing for conversion delays. This helps the agent investigate, but the comparison alone won’t prove that the edit caused the result.

Ask the agent to check this record before every proposal. Reversing a budget cut made two days ago should require new evidence.

Supply context such as promotions and outages yourself. A custom data import can bring structured CSV or spreadsheet records into Supermetrics, but you should also check that the agent can access them.

Once you start approving live changes, use the same record to track what actually happened. An approved edit can fail, or someone can change the setting again. After each action:

1. Check the ad platform to confirm the setting matches the approved proposal.
2. Retrieve the updated settings before the agent’s next review.
3. Record what changed and when, with a link to the proposal.
4. Set a reassessment date that allows for new results and delayed conversions. Keep checking for problems while you wait.

How do you turn business context into campaign decision rules?

Give the agent specific targets and constraints, with an owner who keeps them current. “Keep acquisition costs efficient” leaves too much open: does acquisition mean a form submission or a qualified demo, and what cost is acceptable?

Turn that business context into clear instructions:

• Define the outcome that matters, such as qualified demo requests rather than all form submissions.
• Set the target for that outcome, such as a $120 CPA for a particular product line.
• Explain exceptions, including when a product launch can justify a higher CPA.
• Specify constraints, such as stock availability or a minimum margin.
• Identify campaigns the agent should protect from cuts, such as brand or always-on campaigns.

Record an owner and review date for each rule. Stock constraints may change daily, while a CPA target might remain valid for a quarter.

Shopify and HubSpot data can help the agent assess orders or pipeline alongside ad results. Define how those outcomes relate to the campaigns before asking it to compare them.

Instructions guide recommendations, while permissions and approval settings control edits. “Don’t recommend cutting brand spend below $200 a day” doesn’t create an enforced budget floor. In Supermetrics, choose which accounts have write access and which changes require approval, then check the floor when reviewing proposals. Our guide to managing ad campaigns with AI agents explains those controls in more detail.

What should an agent’s proposal include before you approve it?

A proposal should identify the account and campaign, show the current and proposed settings, and explain the evidence. Include the estimated spend impact and its limits, unresolved questions, and when to reassess.

Campaign spend forecasting can help you sanity-check the spend estimate. Use the example below as a review format. The account, IDs, figures, thresholds, and reporting rules are illustrative. The example includes unresolved questions to show what you should check before approving.

Search impression share is the share of eligible impressions your ads received. The 21% lost to budget suggests the campaign missed opportunities because its budget was limited. It doesn’t tell you how profitable those additional impressions would be.

In this example, ROAS remains above target after excluding the promotion, but the proposal doesn’t give the conversion count for those remaining days. Ask for that figure before approving the increase. Then check the potential spend impact and confirm that reassessment allows for 14 complete days after the change, plus the seven-day conversion delay.

How do you test an agent on historical campaign decisions?

Test past decisions using only the information available at the time. Check whether the recommendation follows from the evidence, including cases where waiting was the right response.

For each test:

1. Choose a past decision with a documented outcome, such as a budget increase on a search campaign.
2. Provide the settings, change record, and business rules from that time. Include only the performance data available then.
3. Ask for a proposal using the format above.
4. Check that the proposal uses the correct reporting period, follows your decision rules, and shows whether the evidence threshold is met. It should also flag uncertainty or explain why it recommends waiting.

Historical reports may now include conversions that hadn’t arrived when the decision was made. Use a saved snapshot from that time, or reconstruct the data available then if your records allow it. If you can’t, treat this as a review of the agent’s reasoning. You won’t be testing what it would have recommended with the information available at the time.

Review later results separately. They can help you assess the decision that was actually taken, but won’t tell you whether a different action proposed by the agent would have worked.

Include cases where keeping the campaign unchanged was the right decision, and cases where there wasn’t enough evidence to act. Check that the agent can justify both responses. Our 7-step guide covers testing the agent on known campaign problems as part of a wider rollout.

How does Supermetrics support campaign management agents?

Supermetrics supports campaign agents through connected data and campaign tools. It also helps teams monitor accounts and review performance. Here’s how each option fits into the workflow:

• Claude, ChatGPT, or Gemini Enterprise: Investigate campaign problems in the AI chat you already use. Ask why CPA rose, then let the agent query connected sources and prepare a recommendation.
• Supermetrics Studio: Review performance in a dashboard and make supported campaign changes from the same workspace.
• Supermetrics Campaign Monitor (alpha): Flag problems in connected Google Ads and Meta Ads accounts with automatic Slack or email alerts. Use those signals to decide what to investigate next.
• Supermetrics MCP server: Connect another compatible AI tool to your marketing data and supported campaign actions through MCP (Model Context Protocol). For scheduled investigations, build a workflow around these tools that determines when the agent runs and what it investigates.
• Build on Supermetrics: Build your own campaign agent or add campaign intelligence to a product using Supermetrics’ APIs and developer tools. Supermetrics handles the data connections; your team designs the agent’s behavior.

For supported campaign actions, you choose which accounts have write access and when approval is required. New campaigns are created paused, and changes are logged in Campaign history. Campaign changes currently work on Google Ads, Meta Ads, Microsoft Advertising, TikTok Ads, LinkedIn Ads, ChatGPT Ads and Snapchat Ads, so check which actions each supports before enabling write access.

Whichever setup you choose, supply the business targets and evidence thresholds the agent should use.

Give your campaign agent the context to make a sound recommendation

Start with one campaign decision. Give the agent the evidence you’d review yourself, then ask it to explain the proposed change. You should be able to trace its reasoning from the inputs to the recommendation.

Test it on past decisions before moving to live proposals. Our guide to agentic campaign management covers what to delegate next. Use the results to decide where the agent helps and where it still needs your judgment.

Start a free 14-day trial to connect your marketing data to Claude or ChatGPT through Supermetrics. Begin with one account and use that data to investigate a campaign decision before enabling live changes.

Frequently asked questions