Connect Claude and ChatGPT to the same groas MCP server, ask the same account question, and put their recommendations side by side before approving a change. The point is not to crown a better assistant. It is to find the decision that still needs a human.

That is the test I would run before a budget shift I would have to explain to someone else. Same account, same question, same date range. Independent answers first. Comparison second.

My expectation: the assistants will agree on most numbers and split on some recommendations. Different reasoning applied to the same evidence should expose assumptions that a single polished answer hides. Where they disagree is where I want the operator looking before anything goes live.

That is a hypothesis, not a result. This protocol is how to check it on your account.

The question: does a second assistant expose a decision worth reviewing?

A second assistant is useful only if it contributes something beyond another confident paragraph. I would judge this test by whether it exposes a query mismatch, an unsupported number, or a recommendation whose assumptions deserve inspection.

The tempting shortcut is to ask one assistant twice. That gives you another answer, but not an independent model’s reading. For this protocol, I want two differently trained models working separately, not one model revising its own homework.

There are two limits to keep in view:

  • Agreement is not proof. Models can share blind spots and produce the same confident mistake.
  • Disagreement is not proof of useful judgment. One assistant may simply have pulled the wrong campaigns or misunderstood the question.

So the side-by-side is the instrument, not the verdict. Before comparing reasoning, establish that both assistants actually used the same evidence. One MCP server gives them access to the same account. It does not force them to request the same rows.

I would call the useful output a disagreement register: a short record of where the answers split, why they split, and what needs checking before approval. Not a leaderboard. A place to put unresolved decisions instead of letting fluent prose settle them.

Setup: one groas server, two independent chats

MCP provides the connection between each assistant and the tools it can use. Its architecture documentation describes the separation that matters here: the host connects to the server, but the protocol does not dictate how the host’s model reasons with the context it receives.

That is the arrangement I want: shared access to the account, separate reasoning about it. Connect both assistants to the same groas server, then check that each can retrieve the account data needed for the question. Client configuration differs; do not treat a successful connection in one assistant as evidence that the other is ready.

For this test, ask both assistants to analyze and recommend, not execute. Keep the approval gate outside both conversations. The exercise loses its point if one assistant starts moving budgets while the other is still explaining its choice of date range.

The groas side costs $19 a month flat for the whole team, with unlimited teammates, ad accounts, and websites, and every feature included. Connecting the second assistant does not add another groas seat or account charge. That is the cheap part of the second opinion. The attention still comes from you.

If the account is not connected yet, start with the walkthrough for connecting Google Ads to Claude in five minutes. Then establish the second connection before starting the comparison.

Controls: fix the question before either assistant answers

Everything below assumes one account and one question. I would schedule the check weekly while an account is changing, and before a consequential budget decision. Keep the scope narrow enough to inspect the evidence yourself.

1. Write one question with an explicit window

Paste the same prompt into both chats. Not “How are my campaigns doing?” That invites two different audits, followed by an afternoon pretending they are comparable.

Use a question with a checkable answer. For example:

Which campaign’s cost per acquisition moved most between 1 and 30 November? Use the year specified for this test. Show the figures behind the comparison, explain what the account evidence supports, and recommend a change—or explain why you would make none. Return the data and tool calls first, then the explanation, then the recommendation. Do not make changes.

Before sending it, put the actual year into the prompt. Fix the account, campaign scope, conversion actions, and comparison window wherever the question leaves room for interpretation. If “moved most” means the largest percentage change rather than the largest absolute change, say so.

A precise question is a control, not prompt-writing theater. If the assistants answer different questions, the test measures your ambiguity rather than their judgment.

2. Keep the first answers blind

Open separate chats. Neither assistant sees the other’s answer, and neither gets a summary of what the other “thinks.”

Do not paste the first recommendation into the second chat and ask whether it looks right. That tests how the second assistant reviews a supplied answer, not how it independently reads the account. Those are different jobs.

Save both initial answers before challenging either one. Otherwise, the assistant’s revised position can quietly replace the position you meant to compare.

Independent reads first. Human comparison second.

3. Ask for evidence in the same order

Require the same output structure from both assistants:

  1. The data retrieved, including scope and dates.
  2. The figures used in the answer.
  3. The explanation those figures support.
  4. The proposed action, its size, and the metric it targets.
  5. Any assumption that could change the recommendation.

The structure matters because writing style is distracting. A crisp answer can hide a missing conversion action. A clumsy one can contain the better-supported recommendation.

I want the evidence exposed before the sales pitch arrives.

Compare retrieval and numbers before reading the recommendation

Start with what each assistant actually asked for. Inspect the tool calls and returned data available in each conversation. Did both use the same campaign scope, date range, attribution setting, and conversion actions?

If one queried three campaigns and the other queried the account, stop there. They are not yet disagreeing about strategy. They are answering different questions.

Shared access is not the same as matched evidence. Correct the query mismatch before treating the answers as a comparison of reasoning.

Then check the figures. A number belongs in the disagreement register if its definition, scope, or window differs, even when the prose around it sounds identical. A CPA quoted to the cent from different campaign scopes is not agreement. It is matching typography.

Use a compact comparison sheet:

CheckClaudeChatGPTWhat needs resolving?
Scope and datesRecord what it retrievedRecord what it retrievedAny mismatch
Spend, conversions, CPARecord the supporting figuresRecord the supporting figuresDefinition or value differences
ExplanationRecord the claimed causeRecord the claimed causeUnsupported assumptions
RecommendationRecord action and sizeRecord action and sizeDecision requiring review

Expect most numbers to match once the retrieval is aligned. Treat that as an expectation to test, not something to wave through. Confirm the figures that will support the decision against the underlying account data, even when both assistants quote them.

Also keep the records straight: the tool output supports what an assistant read; groas’s action log supports what changed. A readable change log is valuable, but it is not a substitute for inspecting the evidence behind an analysis.

Compare recommendations only after the evidence matches

Now read what each assistant wants to do. Put the proposals into the same shape:

  • Action: what changes?
  • Size: by how much?
  • Target: which metric is the change supposed to improve?
  • Reason: what in the retrieved evidence supports it?
  • Assumption: what must be true for the recommendation to make sense?

Same action with different reasoning deserves attention. The assistants may reach the same destination through incompatible assumptions. Agreement on “reduce the budget” does not settle whether either explanation survives inspection.

Different actions on the same verified numbers are the high-value case. One assistant might recommend pausing an ad group; the other might recommend cutting the bid 20% and waiting. That is the split this protocol is designed to expose.

It does not automatically mean the account is genuinely ambiguous. One recommendation may still be weak. The useful question is which recommendation survives a check of its assumptions—not which assistant sounds more certain.

My expectation is that these splits will cluster around decisions the evidence does not settle cleanly. If they instead cluster around sloppy retrieval or unsupported claims, that changes the lesson: fix the evidence workflow before asking either assistant for more strategic freedom.

Sort each split before deciding what to do

Two cut-paper figures compare two columns of account data in an open notebook, circling the rows where the columns differ

I would sort the comparison into four buckets. The bucket determines the next check; confidence of delivery does not.

  1. Different data. One assistant used a narrower window, different campaigns, or different conversion actions. This is a query gap. Clarify the prompt and rerun the side that departed from the agreed scope. Do not spend time debating recommendations built on different inputs.
  2. Same data, different reading. This is the judgment check. Put both proposals against the evidence: which conversion action does each treat as primary, what do the last 30 days of that action show, and what assumption separates “pause” from “reduce and wait”? Decide whether a holdout is appropriate before scaling either proposal. Do not let agreement on the numbers conceal disagreement about what those numbers mean.
  3. A number without support. One answer states a CPA, spend figure, or date that the retrieved evidence does not support. Ask for the exact source row or calculation. If the assistant cannot substantiate it, discard that run’s recommendation. Not most of it. All of it. A recommendation built on an unsupported input needs a fresh run, not a nicer explanation.
  4. Agreement, including a specific figure. This is not a disagreement bucket, but it belongs in the register because it is easy to stop checking here. Verify the decision-critical figure against the platform itself before it supports a change or goes into a deck. Two matching answers do not create a source.

After recording the independent answers, challenge the contested point. Present it to the other assistant without attributing it, and ask for the supporting or contradicting evidence. Keep this second-stage response separate from the original result: the blind portion of the test is over.

Watch how the assistant responds. Does it immediately concede? Repeat its position more forcefully? Or retrieve evidence and explain why that evidence supports its reading?

I would trust the last response further than the first two. I would still verify the row. A model defending itself eloquently is not the same thing as a recommendation surviving inspection.

Decide what the result changes before repeating the test

The output should be an approval decision, not a transcript collection. For each proposed change, record whether you approve it, reject it, or hold it for a specific check. Keep the reason beside the decision.

A useful second opinion changes what you inspect or approve. If it produces another page of agreeable commentary and nothing else, it has not earned a permanent place in the workflow.

Keep the weekly cadence while the account is moving: a new offer, fresh conversion tracking, or a stretch where CPA drifts and the cause is unclear. When both assistants match on retrieval and numbers for three or four weeks on the questions you actually ask, consider dropping the second run to monthly and keeping it for consequential budget decisions.

Do not infer a universal winner from that pattern. Matching on routine questions tells you something about those questions. It does not establish that either assistant should now make every strategic call without review.

The point is to learn where the extra inspection earns its attention cost. Keep the second opinion where it catches something worth catching.

The same discipline applies if you later stop approving every change by hand. groas logs each action with its reason and lets you undo it in one step. Autopilot stays off until you switch it on, and it operates under the rules you set. Those controls make changes inspectable and reversible; they do not make a weak reason good.

If the assistants disagree usefully, keep the cross-check around the decisions where it exposes assumptions. If they mostly reveal query mismatches, tighten the setup. If they consistently agree and the evidence checks out, reduce the cadence rather than paying an attention tax forever.

I would not use this protocol to answer “Which assistant is best for PPC?” I would use it to answer the narrower question that matters before approval: which recommendation is supported by this account, and what remains unresolved?

Run it before your next budget decision. Let the disagreement do the work of a second opinion, and keep the human where the human is worth something: at the split.