The expensive multi-brand AI mistake is a good change applied to the wrong account. Same login, five similarly named accounts, one tired operator at 4pm: one connection that reaches Google and Meta makes that mistake faster, not harder. I worry less about a weak keyword suggestion than a perfectly sensible budget edit landing in Brand B when the task belongs to Brand A. My checklist is deliberately boring: separate accounts, unmistakable names, one brand per chat, echoed IDs, approval before writes, and a log check afterward. The goal is not to make Claude sound certain. It is to make the destination unmistakable.

1. Before connecting: make every account unmistakable

Three matching ad-account levers in a miniature control room, with the wrong brand’s lever pulled

  1. Give every brand its own ad account. Do not ask naming conventions or chat instructions to untangle brands whose campaigns and budgets already share an account.
  2. Keep each account’s billing association explicit. Check the payment profile during setup rather than assuming a new account inherits the right one under an MCC or Business Manager.
  3. Replace names such as Account 1, Test, and New Account. Those labels describe the day somebody created the account, not the brand whose money it spends.
  4. Use a consistent name containing brand, region, platform, and ID. A readable pattern beats five variations that require an operator to remember which abbreviation means which client.
  5. Put the account ID in the name wherever practical. Brand nicknames can overlap; the ID is the part the assistant must repeat before it changes anything.
  6. Create one wiki row per brand with the account name, account ID, MCC link, payment profile, owner, and guardrails. Keep the mapping outside the ad platform too.
  7. Check the row against the actual account before using it in a Project or prompt. A neatly formatted copy of the wrong ID is still the wrong ID.
  8. Use the multi-brand account setup checklist to review account separation and payment-profile pitfalls. Finish the structure before giving an assistant permission to operate inside it.

2. Set rules per brand, not once for the whole team

  1. Write down each brand’s CPA target, ROAS floor, daily budget limit, and protected campaigns. A target that makes sense for one client does not become everybody’s target because they share a connection.
  2. Store those rules in a Claude Project with brand files and approval rules. Do not make a useful instruction depend on somebody finding an old message.
  3. Mark protected campaigns as read-only in the brand’s instructions. Name the campaigns explicitly so “leave the important ones alone” never has to pass for an operating rule.
  4. Use enforced account permissions for restrictions that must hold. A prompt can state a budget limit; it should not be mistaken for an account-level block on changing that budget.
  5. Review what the actual connector exposes before relying on a control. One gateway can reach several platforms, but that does not make every account’s permissions or available actions identical.
  6. Set write-capable tools to require approval where the connector offers that control. Review the Claude connector controls rather than assuming every connected tool asks first.
  7. Block edits to dangerous areas where account-level controls support it. If a restriction cannot be enforced there, keep the task read-only rather than treating a written warning as a lock.
  8. Start read-only across the MCC, then enable approval-gated writes on one or two accounts. Prove the account selection and approval path before widening access to the rest.
  9. Keep the rollout proportional to the risk. For an illustrative $20k/month spread across four brands, check the read path on all four and the write path on one first.
  10. Require a dry-run preview for cross-account work. A shared audit can be useful; a shared mistake repeats. List every affected account and protected campaign before approving anything.

3. At session start: name one brand and prove the ID

  1. Open a separate chat for each brand. Do not turn a clean Brand A thread into a Brand B thread because the second request looks quick.
  2. Put scope in the first line: Working on Brand B, Google Ads account 123-456, only in this chat. Include the platform rather than relying on the brand name alone.
  3. Enable only the connector the conversation needs. Claude’s per-chat connector selection applies to that conversation; repeat the check when opening another chat rather than assuming yesterday’s setup carries over.
  4. Do not confuse connector selection with account restriction. If the enabled connection reaches every client, selecting it still leaves account identification to the task and the permissions behind it.
  5. Ask Claude to read back the account name and ID before requesting recommendations. Compare both with the brand’s wiki row, not with whatever the chat last called the account.
  6. Ask for the current daily budget of the campaign being changed. Tie the readback to the proposed edit so a familiar-looking account name does not become the entire identity check.
  7. Stop if the name, ID, or current value does not match. Resolve the mismatch before discussing the merits of the recommendation; a good recommendation does not repair bad scope.
  8. Repeat the account ID in every write request. Pause keyword X in account 123-456 is clearer than pause it there, especially after several rounds of analysis.
  9. Keep approval enabled for budget edits and publishing. Do not trade repeated confirmation for conversation-wide permission just because the first few proposed changes look sensible.
  10. Open a new chat when the brand changes. Carry the relevant brand rules into that session, not a pile of context from the account you just finished.

4. Paste a starter prompt before requesting work

  1. Use the full scope prompt below at the start of each brand chat. Keep it in that brand’s Project so nobody has to reconstruct the rules from memory.
  2. Replace the placeholders with the actual brand, platform, and account ID. Leaving an example ID in a reusable prompt defeats the check before the first tool call.
  3. Read Claude’s reply before proceeding. The prompt requests proof of scope; merely pasting it does not mean the assistant has supplied that proof.
Working ONLY on [Brand B - Google Ads account 123-456] in this chat.
No other brand.

1. First, confirm the account you see: name and ID.
   For any proposed budget edit, read back the campaign and current daily budget.
2. Propose changes as a dry run first: what changes, affected accounts,
   affected campaigns, and shared assets that could be touched.
3. Wait for my explicit approval before any write.
4. After an approved change, report the account ID, timestamp,
   and before-and-after values.
5. If I mention another brand, stop and tell me to open a new chat.
  1. Keep the identity check ahead of the recommendation. Otherwise, the conversation can spend ten useful minutes improving a campaign in an account nobody meant to open.
  2. Keep the dry run separate from execution. A proposal should show what could change; approval should refer to that stated change, not to a vague instruction to improve performance.
  3. Keep the last instruction. “While you are in there, check Brand C” is exactly the shortcut this prompt is designed to interrupt, not a harmless exception to it.
  4. Treat the prompt as a session habit, not a substitute for permissions. Keep account-level restrictions and approval controls in place even when the assistant repeats every instruction correctly.

5. Before approval: inspect the destination, then the change

  1. Check the account ID in the proposed write. Do this before judging whether the bid, keyword, pause, or budget recommendation makes sense. Destination comes before optimization.
  2. Match the campaign or other target to the readback. The right account is necessary, but it is not permission to edit whichever campaign happens to look relevant.
  3. Compare the current value with the proposed value. If the assistant cannot state what changes from what, send the proposal back rather than approving an incomplete description.
  4. Check the proposal against that brand’s targets and protected campaigns. Do not borrow Brand A’s CPA target or budget limit to justify a change in Brand B.
  5. Preview the blast radius: every account, campaign, or shared asset the change could affect. This matters most when a task mentions shared negatives or several accounts at once.
  6. Inspect the actual tool input when the approval interface exposes it. A clean explanation in chat is useful, but it is not a reason to ignore a different account ID in the write request.
  7. Approve only the stated change. If the destination, amount, or affected campaigns change afterward, require a fresh preview rather than stretching the earlier approval to cover it.
  8. Keep cross-account requests explicit. For an MCC-wide audit or shared-negative task, require the account list; “all the brands” is not a sufficiently precise destination for a write.
  9. Stop on any mismatch. Do not approve now and promise to inspect the log later. The log is the final check, not permission to skip the checks before execution.

6. After execution: collect evidence for each account

  1. Require a report with the account ID, timestamp, and before-and-after values for every change. An assistant saying “done” is a status message, not enough information to verify the work.
  2. Open change history in the named account and match those fields. Check the actual destination and values rather than looking only for evidence that some change happened.
  3. Filter to one account ID at a time. A combined MCC view can make a wrong-account write look ordinary when several brands have changes happening together.
  4. Separate requested changes from completed changes. If an action did not execute, keep it out of the completed list so the next task does not start from a false account state.
  5. Use groas’s reasoned change log and one-step undo to inspect what changed and why. The record makes review easier; it does not remove the need to review by brand.
  6. Check shared assets separately. Shared negative lists and audiences can affect more than the account named in the conversation; verify their associations rather than assuming one account’s history tells the whole story.
  7. Check billing associations separately when they are part of the work. Do not treat a campaign-change review as proof that the account still points to the intended payment profile.
  8. Review according to what the connected tool can actually change. Do not assume a Google task is read-only, or a Meta task is write-enabled, merely from the platform name.
  9. Stop the next edit if a required field is missing or the log disagrees with the summary. Fix the scope and establish what happened before giving the assistant more work.

7. The item teams skip: sign off against the per-brand log

  1. Keep the assistant’s summary in the chat and the evidence in the account. They serve different purposes: one explains the task; the other shows where the change actually landed.
  2. Require the same sign-off for a routine pause as for a budget edit. Familiar tasks invite shortcuts, and a familiar-looking result can still belong to the wrong brand.
  3. Do not close the session because the numbers sound right. Match the ID, timestamp, and before-and-after values first. Confidence is not account scope, however tidy the response looks.
  4. Review change history filtered to one account ID before closing the chat. In an illustrative case, an unnoticed budget increase that adds $100 in spend per day costs $600 over six days. The recommendation can be sensible, the explanation polished, and the destination wrong. Check the log by brand.