Google Ads MCP: A practical setup and reporting guide

What is Google Ads MCP?

Google Ads MCP is a bridge between an MCP-compatible AI assistant and the Google Ads API. Instead of supplying a new spreadsheet for every question, you let the assistant request approved account data through a server. The server determines which tools exist and how requests reach Google Ads.

Think of the assistant as the analyst, the server as the interface and Google Ads as the source system. A helpful answer still needs the correct account, reporting definitions and business question. A connection makes retrieval easier; it does not supply the reasoning the account review requires.

Google maintains a public googleads/google-ads-mcp repository. Verify the publisher and release before installing anything with a similar name. Community servers and hosted products may expose different tools, require different credentials or retain data under different terms.

• Google’s official MCP integration guide

Is the Google Ads MCP server read-only?

The answer is version-dependent. At this guide’s review on 6 November 2026, Google’s integration page describes a read-only release. The upstream repository README also lists asset creation and label mutations with a proposal-and-confirmation process. Treat that discrepancy as a reason to inspect the release and exposed tools, not to assume every installation is read-only.

For a reporting deployment, disable write categories and use an appropriately restricted account identity. The current upstream README documents disabling the assets and labels namespaces. Confirm the resulting tool list in the host before opening client data.

An instruction such as “do not make changes” is useful context, but it is not a permission boundary. Tool configuration and source-system access should enforce the intended limit. Keep any future write workflow separate, with a reviewable proposal and a named approver.

namespaces:
  customers: true
  search: true
  metadata: true
  assets: false
  labels: false

• Upstream tools, namespaces and mutation confirmation workflow

Set up the account, server and host in that order

First, choose the Google Ads customer account the analysis should cover and confirm that the intended identity can access it. If access goes through a manager account, distinguish the manager’s login customer ID from the client account being queried.

Next, follow the current server release’s Google Cloud and authentication instructions. Google’s integration guide uses a project ID and Google credentials; it says the latest server no longer requires the former developer token after the September 2026 change. Older examples may still include that variable. Match instructions to the release rather than mixing old and new configurations.

Finally, choose the transport your host supports: a local stdio process or an authenticated Streamable HTTP service. A local process configuration cannot simply be pasted into a web connector’s URL field. Keep credential files on the approved host and out of prompts, source control and shared screenshots.

1. Select one customer account and record the manager relationship if applicable.
2. Review the upstream release, runtime dependencies and credential requirements.
3. Apply the reporting-only tool configuration and least-privilege account access.
4. Configure the host for the server’s actual transport and authenticate.
5. List available tools and reconcile a small account lookup before expanding the task.

• Follow the host-specific Claude setup guide
• Follow the ChatGPT setup and export guide

Build a first report you can reconcile

Start with a narrow request: campaign-level spend, clicks and the defined conversion measure for one account over a closed date range. Ask the assistant to show the query, filters and returned rows before interpreting them.

Google Ads Query Language, or GAQL, selects fields from a reporting resource. The illustrative query below requests campaign totals. It contains no account ID because the customer is selected when the request is sent. Validate field compatibility with the API version used by the server.

Use the same dates, account currency, time zone and conversion columns in the UI comparison. Cost fields expressed in micros need unit conversion before presentation. Check totals before adding segments; segmentation changes row granularity and can create misleading totals if rows are combined carelessly.

SELECT
  campaign.id,
  campaign.name,
  metrics.clicks,
  metrics.cost_micros,
  metrics.conversions
FROM campaign
WHERE segments.date BETWEEN '2026-10-01' AND '2026-10-31'
ORDER BY metrics.cost_micros DESC
LIMIT 50

• Google Ads Query Language overview
• Google Ads reporting example: cost_micros conversion

Use live data to narrow the diagnosis

Once retrieval is dependable, give the assistant a decision to support. Which campaigns have higher ad spend per qualified opportunity? Which search groups deserve a closer intent review? Which changes coincide with a tracking discrepancy?

Some answers require another system. If Google Ads only receives form submissions, the server cannot reveal which people became customers unless those outcomes exist in accessible data. Join approved CRM information using stable identifiers or reconcile an aggregate cohort report separately.

Require a distinction between observation and explanation. “Recorded conversions fell” describes the returned data. “The audience became worse” is a hypothesis that needs additional evidence. Keep the useful uncertainty rather than asking the model to fill every gap.

• Define CPL, CAC and mature lead cohorts

Troubleshoot the failing layer instead of reconnecting everything

If tools are missing, inspect server startup, the host configuration and tool filters. If authentication fails, inspect the approved credential flow. If the client account is unavailable, confirm the identity’s access and manager context. These are different faults.

If the query fails, inspect the resource, fields and API version. If it returns no rows, check dates, status filters and whether the account actually has matching data. An empty result is not evidence of a broken integration by itself.

Keep a small reference report for regression checks after updates. Record the server version, permitted tools and a known query. Recheck those boundaries when upgrading because a new release can change capability as well as fix bugs.

Frequently asked questions