Google Ads with ChatGPT: A guide to analysis and MCP setup

Choose between an export, a plugin and a developer integration

A business owner asking ChatGPT to review ads may need a simple exported table. An analyst asking recurring account questions may benefit from an approved MCP connection. A developer building a reporting application has a different task again.

Name the route before following setup instructions. ChatGPT web, desktop clients and an application using the OpenAI API do not share one universal configuration screen. A working integration in one surface is not proof that the same tools exist in another.

Choose the smallest setup that can answer the business question. If the decision needs one closed-period comparison, a reliable export can be more practical than maintaining a server. If the team repeatedly retrieves the same approved fields, a connection may justify its overhead.

Prepare an export ChatGPT can interpret correctly

Export the fields needed for the question: campaign ID and name, dates, spend, clicks and the specific conversion measures being compared. Add a short data dictionary explaining units, currency, time zone and what each outcome means.

Remove private customer details that the task does not need. Use aggregated CRM outcomes where they are sufficient and permitted. Avoid mixing campaign-level and event-level rows in one total without explaining the join and granularity.

OpenAI’s file guidance supports supplying source files with clear structure and review criteria. Treat that as a way to make the task inspectable. Attaching a spreadsheet does not establish that every number or causal explanation in the answer will be correct.

Question: Which campaign groups deserve investigation before more spend?
Account: the single approved customer account in this export.
Period: state exact start and end dates and account time zone.
Units: state currency and whether spend is currency units or micros.
Conversions: define form submissions, qualified leads and customers separately.
Sales maturity: identify cohorts whose outcomes are still incomplete.
Output: calculations, supporting rows, uncertainties and next checks.
Do not invent missing data or treat attributed conversions as incremental sales.

• Official OpenAI documentation: working with files

Connect a Google Ads MCP server to ChatGPT web

The current official OpenAI instructions describe adding a custom MCP server through ChatGPT Plugins. Select the plus button, choose Add custom MCP server, enter the reviewed server URL or supported tunnel connection and configure authentication. Create and install the resulting plugin before using it in a conversation.

Workspace permissions and security restrictions can affect availability. The server must implement the transport and authentication supported by that surface. A URL pointing to a repository or a local stdio launch command is not a remote MCP endpoint.

Authenticate only to a server the business has approved. Review its publisher, permitted tools and account boundaries. ChatGPT supporting read and write tools does not mean every Google Ads server exposes the same actions. A connected server is not automatically a native Google-built integration.

• Official OpenAI documentation: add a custom MCP server
• Review Google Ads server capabilities before connecting

Keep desktop MCP and API setup separate from web plugins

OpenAI’s MCP documentation distinguishes host-configured local stdio and Streamable HTTP servers from tools packaged for web use. Follow the setup for the actual desktop or Codex host rather than assuming that an arbitrary local server will appear in ChatGPT web.

Developers can also connect MCP tools through the OpenAI API. That route requires application-managed authentication, tool exposure and approval behaviour. It is an implementation project rather than a box to tick inside a normal chat.

If a guide tells you to paste an API key, OAuth secret or Google credential file into a prompt, stop and check the supported connection mechanism. Credentials belong in the approved integration configuration. The report-analysis conversation should contain the data and context needed for the task.

• Official OpenAI documentation: MCP by client surface
• Official OpenAI documentation: MCP tools in the API

Ask questions that lead to a reviewable decision

For performance reporting, request the calculation and supporting rows before the narrative. A weighted account CPL is total relevant spend divided by total leads, not the simple average of campaign CPLs. Ratios need the correct denominator.

For search-term review, supply the services, locations and documented customer-fit rules. Ask ChatGPT to return the query, evidence of mismatch, uncertainty and proposed next check. Avoid auto-excluding every term without a conversion in a short window.

For ad-copy development, give the approved offer and facts the business can support. Ask for several distinct buying reasons with the claim behind each. Verify specifications and qualifiers before using the copy in an account. A useful draft should expose what still needs confirmation.

Use only the attached report or approved reporting tools.
First show the account, period, filters and row count.
Calculate total spend, leads and weighted CPL from the source rows.
For each material change, separate observation from hypothesis.
State whether CRM customer outcomes are available.
Return three prioritized next checks with supporting evidence.
Do not make account changes or infer profit from lead volume.

• Use a search-intent map to review keywords

Verify the arithmetic and the business conclusion

Consider a hypothetical report with $600 spent on ten leads and $400 spent on 40 leads. The campaign CPLs are $60 and $10. Their simple average is $35, but the combined CPL is $1,000 divided by 50: $20. Recalculate the metric before accepting a polished summary.

Then ask whether the metric answers the decision. Cheap inquiries may not become customers. Recent leads may still be in sales. Attributed revenue may include purchases that would have occurred without the ad. ChatGPT cannot resolve those issues from a report that does not contain the relevant evidence.

Turn an accepted finding into a bounded action with an owner and a review date. Keep the source export or query alongside it. The benefit is a clearer account review, not a transfer of accountability to the assistant.

• Distinguish lead cost from customer cost
• Choose which paid ads tasks AI should support

Frequently asked questions