ChatGPT cannot diagnose an ad account it has never seen. Ask why CPA rose without giving it account data, and a tidy answer about competition or creative fatigue is still a guess.
Everything in this piece follows from that missing input. An assistant can draw on patterns it learned, but those patterns are not a report from your campaigns. A plausible benchmark does not become your cost per lead because it appears next to the words “your account.”
The fix is not a cleverer prompt. It is a live connection to your numbers, followed by a demand to show which numbers support the answer. I want to build that distinction from the ground up, using one account throughout. Then I will apply it to a different case, where the same rules expose a different missing input.
A benchmark cannot explain your CPA increase
Say I am looking at an account spending about $20k a month on Google Search: one branded campaign and three non-branded service campaigns. The question is straightforward: “Why did CPA rise over the last two weeks?”
Without a connection or an uploaded report, the assistant cannot see:
- How much each campaign spent.
- Which campaign lost conversions.
- Whether clicks became more expensive.
- Which conversion actions the account counts.
- Whether conversion lag affects the comparison.
It has the question and its learned patterns. That is enough to suggest possibilities. It is not enough to choose between them.
A Databox walkthrough illustrates the problem. Given a single LinkedIn CSV, ChatGPT supplied a $50–$200 B2B CPL range, an 8–12-week saturation rule, and a 3–4-week creative cycle. Only two numbers came from the file: a $113.79 CPL and a 60% spike.
The same substitution can happen in search. Published benchmarks from 13,474 US search campaigns include a 6.64% CTR, $5.42 CPC, 8.18% conversion rate, and $66.69 cost per lead. Those numbers can provide a sanity check across Google and Microsoft Ads. They cannot explain what changed in the three service campaigns in this example.
Benchmarks describe other accounts. A diagnosis has to start with this account.

Make the assistant expose what it does not know
Confidence is not evidence. When I assess an answer to that CPA question, I look for three warning signs: no campaign names, no account-specific dates or spend, and recommendations that could fit almost any advertiser.
“Try new headlines” might eventually be sensible. Without the numbers, it is a suggestion looking for a problem.
I use this five-part check, adapted from the Databox audit routine, to separate evidence from fluent padding:
- Restate the question and list what is ambiguous. Does “CPA” mean a form submission, a qualified lead, or something else?
- List the assumptions. What has the assistant assumed about the industry, sales cycle, and deal size?
- Label the source of every number. Did it come from supplied data, a connected tool, or a general benchmark?
- Name the three missing inputs that would most change the answer.
- State what cannot be answered from the available information.
A useful assistant can admit that it lacks the evidence. That is a better result than five paragraphs of confident troubleshooting aimed at an account it has never opened.
If the answer would fit any account, do not treat it as an account diagnosis.
An export gives it evidence; a live connection supports follow-up
The simplest repair is a CSV. Export the relevant Google Ads data, give it to the assistant, and ask which campaign has the highest CPA. Now it has something real to read.
For a one-off question on the $20k account, that can work. Instead of reciting industry averages, the assistant can name a campaign and show the spend and conversions behind its calculation.
But an export is a snapshot. It does not update after download, and the assistant sees only the slice I selected. Large files can hit upload or context limits. A follow-up question may require another export.
That matters because diagnosis rarely stops at the first table. If one service campaign accounts for the increase, I want to examine it by device or search term. Then I want to compare periods. The first answer creates the next question.
A live connection over MCP lets the assistant request current account data through tools rather than wait for another pasted report. MCP is the connection mechanism, not a guarantee of good analysis. The assistant still has to request the right data and interpret it correctly.
An export supplies a snapshot. A live connection lets the investigation continue without rebuilding that snapshot by hand.

Connect groas, then verify reporting before allowing changes
With groas, I connect the ad accounts and website once, then connect groas to the assistant over MCP. The server address is:
https://mcp.groas.ai/mcp
The division of work is straightforward: the assistant is the brain; groas supplies the hands and the data. That gives the assistant a way to read connected accounts and take actions, with a record of what it does.
The setup sequence is:
- Connect the relevant accounts inside groas.
- Add groas through the assistant’s custom MCP connection setup, using the server address above.
- Complete the OAuth sign-in and authorization flow.
- Request reporting only. Do not authorize campaign changes during the first check.
- Compare the returned campaign spend and conversions with the ad platform.
The ChatGPT and Claude setup walkthrough covers the connection pattern. I also cover the Claude path in connecting Google Ads to Claude in five minutes.
My first test is deliberately boring: show campaign spend and conversions for a defined period. Then answer a follow-up reporting question. No optimization theater. If the assistant cannot retrieve numbers that match the account, I fix access before considering any action.
A successful sign-in is not the test. A correct account report is.
Use three prompts to establish evidence before interpretation
Once the connection works, prompting becomes a control rather than a substitute for data. I use three instructions in order: establish the evidence rule, request the table, and permit an honest refusal.
1. Set the evidence rule
Use only my account data. Show the numbers behind each claim. Label any assumption separately.
This keeps an account diagnosis separate from general advice. Benchmarks can still be useful when explicitly requested; they should not quietly fill gaps in the account report.
2. Ask for the table that tests the question
Show my actual CPA by campaign for the last 14 days and the preceding 14 days, with spend and conversions for each period. Separate branded from non-branded campaigns. State the conversion action used.
The comparison matters. A table covering only the latest period shows where CPA stands, not what changed.
On the hypothetical $20k account, this is where I would look for a service campaign whose conversions fell while spend held, or whose costs rose without a matching increase in conversions. Those are possibilities to test, not findings to invent.
3. Give it permission to stop
If you do not have the data needed to answer, say what is missing instead of estimating.
That last instruction matters as much as the first two. I would rather get a precise request for missing data than an invented explanation delivered with excellent formatting.
No table with spend, conversions, and comparable dates means no account-specific explanation yet. Even with the table, a pattern is a lead to investigate, not automatic proof of its cause.
Reading identifies the problem; controlled actions address it
Reading and writing are different capabilities. A connector that retrieves reports does not necessarily let the assistant change campaigns.
Google’s own MCP server is read-only: it can list customers and run GAQL queries, but it cannot pause an ad or change a bid. For the workflow I want, that leaves the manual step intact. The assistant finds something; I still go into Google Ads to act on it.
A write-capable connection through groas closes that gap. Suppose the next drill-down shows an expensive search-term cluster in one non-branded service campaign. I can ask the assistant to show the terms and their costs, propose a change under my rules, and wait for approval before acting.
The order is important:
- Show the evidence.
- Explain the proposed change.
- Get approval under the agreed rules.
- Make the change and show the log.
That is the practical point of connecting the data and actions in groas. It is not just a more conversational report. It removes the handoff between identifying work and doing it.
I want the evidence before the action and the log after it. groas keeps changes logged with their reasons and makes them undoable in one step. Autopilot stays off until it is switched on; a connected account is not permission to let everything run unattended.
Live data does not supply missing business context
The connection fixes access. It does not make every conclusion correct.
On the $20k account, the assistant might retrieve the last 14 days perfectly and still misunderstand the increase. Platform conversions do not automatically tell it that the sales team tightened qualification. Campaign totals do not explain an unreported landing-page swap. Blending branded and non-branded search, or Search and Performance Max, can hide the movement I need to inspect.
This is where I separate three questions:
- What happened in the account? Spend, clicks, conversions, and campaign-level changes belong here.
- What does the account count? The conversion action, attribution setting, and date range define the report.
- What happened in the business? Qualified pipeline, closed-won revenue, and operational changes may require additional data or context.
Before accepting a conclusion, I ask the assistant to state its definitions. Then I compare the platform metric with the downstream result that matters. If that result is not available through the connection, I have to supply it or leave the question open.
The assistant can read the data it receives. I still own the context it does not have. That is why I keep changes behind approvals at first, even when the reporting looks right.
FAQ
How do I get ChatGPT to read my ad data instead of guessing?
I give it accessible data, then require it to identify the evidence behind its claims. An uploaded report can answer a bounded question. A live groas connection supports account queries and follow-up analysis without repeated exports.
The check stays the same: ask for spend, conversions, campaign names, and dates. A confident paragraph is not a substitute for a verifiable table.
How do I connect my ad accounts to ChatGPT?
Connect the accounts inside groas, then add groas through ChatGPT’s custom MCP setup using https://mcp.groas.ai/mcp and complete OAuth authorization. Start with a reporting request and compare it with the ad platform before allowing changes.
I use the same principle in Claude: establish the connection, verify the numbers, then consider actions. Reporting comes before permission to optimize.
What does groas cost for this workflow?
groas Unlimited is $19 a month, flat, for the whole team, with unlimited teammates, ad accounts, and websites. It covers Google Ads, Microsoft and Bing Ads, Meta Ads, and the other groas features.
I still make the first account test reporting-only. Paying for access does not remove the need to verify what the assistant retrieves.
Work a new case through the same model
Now take a home services operator spending $6k a month across Google Search and Meta. The question is: “Why do leads cost more this month?”
Disconnected, the assistant might offer auction pressure, creative fatigue, and a benchmark CPL. None of those establishes what happened. I would run the same audit: name the campaigns, identify the dates, and label the source of every number.
After connecting the accounts, I would first ask for spend and platform conversions by campaign across comparable periods, keeping branded search separate from non-branded. That establishes where the reported increase sits.
Then comes the new complication: the operator cares about booked jobs, not just leads.
A connection to ad data does not, by itself, prove which leads became bookings. I would ask for cost per booked job only if booking data is available. Otherwise, the assistant should identify that missing input rather than rename platform conversions “booked jobs.” Different label, different business decision.
Suppose the supplied data then shows one non-branded search campaign holding spend while conversions fall, and one Meta campaign generating cheap leads that do not become bookings. The blended CPA would hide two different problems. I would investigate them separately, not apply one account-wide fix.
If the operator also reports a landing-page change during the period, that becomes another explanation to test. It does not become proof that bids are innocent simply because the dates overlap.
That is the model working: establish access, define the metric, inspect the change, ask for missing context, then decide what action the evidence supports.
Give the assistant a live connection, make it cite spend and conversions, and keep changes logged and undoable. Then I am not asking ChatGPT to guess about the ads. I am asking it to read them.

