AI assistants are good at reading a large report and producing a confident sentence: “Lower the bid on this keyword by 20%” or “Move budget from Campaign B to Campaign A.” The sentence may be right. It is still a hypothesis about how shoppers and an auction will behave, built from data that has limits. The person who approves the change remains accountable for the money it spends. This guide sets out what to examine before agreeing, a simple flow for moving from suggestion to change, and a checklist to keep next to the approve button.
What to examine
Evidence
Ask what the recommendation is based on: which dataset, which period, which metric definitions. You should be able to trace every number back to source figures. If a suggestion cites an ACoS, check that it was calculated from total spend and total attributed sales, not from averaged rows, and that attributed sales were not mixed with seller sales.
Data freshness
Look for the “as of” time. Advertising results for the latest days are provisional because attributed sales keep arriving after the click. A recommendation built on the last three days may be reacting to data that has not finished arriving.
Sample size
Small numbers produce big swings. A keyword with no orders after a handful of clicks may simply be unlucky. A rough way to see this, assuming each click converts independently at the same rate (a simplification), is:
chance of zero orders = (1 − CVR) ^ clicks
| Target (fictional) | Clicks | Attributed orders | Observed CVR | Chance of zero orders if the true CVR were 10% |
|---|---|---|---|---|
| Keyword A | 9 | 0 | 0.0% | about 39% |
| Keyword B | 30 | 0 | 0.0% | about 4% |
| Keyword C | 60 | 0 | 0.0% | about 0.2% |
| Keyword D | 400 | 52 | 13.0% | not relevant |
Zero orders on 9 clicks happens by chance about four times in ten even for a keyword that usually converts well. Zero orders on 60 clicks is much harder to explain away. A recommendation to cut Keyword A deserves far more doubt than one to cut Keyword C.
Assumptions
Every recommendation assumes something: that click costs stay roughly level, that demand is steady, that the product stays in stock, that the price does not change. Good recommendations state their assumptions. If yours does not, write them down yourself before deciding.
Counterfactuals
Ask what would have happened without the change. A campaign that “improved” after a bid cut may have improved because a holiday ended. A keyword that looks weak may have been weak only while the product was out of stock or while a competitor ran a promotion. Compare against a period with similar conditions, not just the previous week.
Reversibility
Some changes are easy to undo; their costs are not. A bid can be changed back in seconds, but money spent while it was set wrong is gone. Pausing a campaign stops spend and also stops the data you would need to evaluate it. Prefer changes that are small, bounded in time, and simple to reverse.
Approval and rollback boundaries
Decide in advance who may approve which changes and how large they can be without a second approver. Record the original values before any change, and define a rollback trigger: a specific condition, such as spend exceeding a set amount with no attributed orders, that tells a named person to restore the original setting.
Propose → Explain → Validate → Authorize → Act → Verify
A clear sequence keeps the AI in an advisory role and the person in control.
| Step | What happens | Output |
|---|---|---|
| Propose | An assistant or analyst drafts one specific, bounded change | “Lower Keyword C bid from 1.20 to 0.95 for 14 days” |
| Explain | The proposal shows evidence, period, as-of time and assumptions | A rationale linked to source figures |
| Validate | A person checks the numbers, sample size, freshness and context | Confirmed, revised or rejected |
| Authorize | Someone with authority approves within agreed limits | A recorded approval |
| Act | The change is applied exactly as approved, nothing more | A change log with old and new values |
| Verify | After attribution has matured, results are compared with expectations | Keep, adjust or roll back |
The example values in the table are fictional. The important part is the separation: the step that proposes is never the step that authorizes. In KEYXE’s design, recommendations are drafts for people to review. There are no tools that change advertising accounts, and any future action tool would show a preview and require a person’s approval, separate from read access.
A practical checklist
- The data period and “as of” time are visible, and the most recent days are marked provisional.
- The metric definitions match yours, and ratios are calculated from totals.
- Clicks and orders are numerous enough that chance is an unlikely explanation.
- The assumptions are written down and still true today.
- Seasonality, promotions, price changes and stock-outs have been considered.
- Inventory can support extra demand if the change works (check days of cover).
- The change is specific: which setting, by how much, for how long.
- The original values are recorded and the change can be reversed.
- A rollback trigger and a review date are set.
- A person with authority has approved it, within agreed limits.
- No data is shared beyond what permissions and data-use policies allow.
- The result is reviewed after the attribution window has had time to mature.
Common mistakes
- Approving a change because the explanation sounds confident rather than because the evidence holds.
- Reacting to the last few days before late attributions arrive.
- Making several changes at once, so no single result can be read.
- Skipping the verify step, which turns every decision into an untested guess.
- Letting the same system propose and approve.
Try it
Use the MCP workspace of the synthetic demo to see how a tool response carries its period, “as of” time and warnings, then apply the checklist to a sample campaign in the advertising workspace. The Ads Metrics Calculator at the free tools helps recheck any ratio. Related guides: What Is MCP? and Inventory Data Freshness. Terms are defined in the glossary.