A weekly creative test readout grouped by hook and format

Two sources, both read-only for this job.

• Meta. There is no built-in Meta Ads block. Instead, an admin connects Meta’s official ads MCP server under Settings → Tool connections. MCP is a standard way to plug a service’s tools into an AI agent. Once it is on the Relay’s allowlist, the Relay can list ad accounts and read campaign and ad insights using the agency’s connected account.
• Google Ads. The Google Ads block reads customers, campaigns, ad groups and ad performance, and runs custom queries. It cannot change budgets, bids or status.

The Relay joins both to the tags table by ad ID, then groups spend, click-through, cost per purchase and frequency by hook and by format.

How does it decide what is a winner?

It follows the agency’s testing rules, stored in the Knowledge Base, not its own judgement. Ours say:

• No call below the spend floor. A hook with less than an agreed spend per variant is "too early", whatever its numbers say.
• Compare like with like. Video against video, prospecting against prospecting.
• Fatigue is a trend, not a number. An ad is flagged when frequency passes the agreed level and click-through has fallen for three weeks in a row.
• Respect brand limits. Some clients do not allow price-led hooks or before-and-after claims in certain categories. Proposals stay inside those limits.

What does the Monday readout look like?

The readout goes to the account lead’s chat with the Relay, with a notification. It leads with ideas, then formats, then tired ads, then three proposed tests. The example above is illustrative. A good readout is short enough to read before standup.

If the agency also sends the readout to the client, the Relay drafts the email, and email always waits for an approval. Keep the recipient policy on internal only to start, so drafts go to the account lead, who forwards them.

What happens to budget changes?

Every call a Relay makes is classed as read, write, destructive or spend. Reads run. Anything that moves money, such as raising an ad set budget through Meta’s tools, is a spend call and waits for a person to approve the exact change, unless the Relay has delegated limits and the change fits inside them. Approved calls run once and are written to the audit log.

Google Ads changes cannot be made from the Relay at all, because that block only reads. The readout says "proposed" and a person makes the change in Google Ads.

We start every new client in Observe, where writes are blocked, and keep it there for the first month. The proposals still appear, so the team can judge them without risk.

What did we learn in the first month?

• Tagging discipline decides everything. A week with sloppy tags gave a confident and wrong readout. We now block launches until the tags row exists.
• "Too early" is a useful answer. The spend floor stopped the team killing hooks after two days.
• Proposals got better with context. Adding the client’s brand limits and past test results to the Knowledge Base cut proposals that the team would never run.
• Keep the budget move small. Approving a 20% shift is an easy decision. Approving a rebuild is not, so the rules ask for small, reversible moves.

For pacing spend against a monthly plan, see ad budget pacing across Google and Meta. For running Meta campaigns from Chat, see Meta ads campaign management.