All posts

MCP Google Ads AI Shopify

Google Ads MCP: What the Official Server Does and What It Cannot Do

September 2, 2026 Agency AI Team 9 min read

META TITLE: Google Ads MCP: What It Does and Cannot Do

META DESCRIPTION: Google's official Google Ads MCP server is read-only. Here is what its three tools do, what it cannot change, and how merchants actually use it.

SLUG: google-ads-mcp

TAGS: google ads mcp, mcp server, google ads api, gaql, shopify advertising, ai marketing

Google Ads MCP connects an AI assistant directly to a Google Ads account through the Model Context Protocol. Google's official server arrived on 7 October 2025, and its first release is read-only for reporting and diagnostics.

Good use looks like fast diagnosis, with the answer to a cost per conversion question coming from your own data.

Bad use looks like expecting the assistant to pause a campaign, because Google's own documentation says it cannot. This article covers the three tools the server ships, its setup and quota limits, and the handoff to a human.

What the Official Google Ads MCP Server Is

MCP is an open-source standard for connecting AI applications to external systems. Under that standard, Google open-sourced an official Google Ads API MCP Server. The announcement went out on the Google Ads Developer Blog on 7 October 2025.

Google published it under the Apache-2.0 license, and the code lives in a public GitHub repository. You install it with pipx, pointed straight at that repo. Nothing is hosted for you, which matters more than it first sounds.

If you are surveying the wider category, our roundup of the best MCP servers for marketing covers the neighbors.

The Three Tools the Server Ships

list_accessible_customers returns the Google Ads customer IDs and account names accessible to the authenticated user.
search executes Google Ads Query Language requests to fetch resource metrics, budgets and status.
get_resource_metadata retrieves metadata about a Google Ads API resource type, for example campaign.

Search does the heavy lifting here. Google Ads Query Language queries the API for resources, their attributes, segments, metrics and metadata. Anything expressible in GAQL is therefore something the assistant can pull and reason over.

The Four Read-Only Resources and the Config File

discovery-document, the Google Ads API discovery document.
metrics, the available metrics for reporting.
segments, the available segments for reporting.
release-notes, the latest API release notes.

Those four stop the assistant from guessing field names. A tools_config.yaml file enables or disables individual tools. Disabling the other two tools is a config change, not a code change.

Google Ads MCP Setup Requirements

The server uses standard input/output as its transport, so it runs as a local process on your own machine. No hosted remote endpoint exists to point Claude or ChatGPT at. Someone has to install it, authenticate it, and keep it running.

Before configuration, you need a developer token plus OAuth credentials from the Google Ads Developer console.

A developer token is a 22-character alphanumeric string that lets an app connect to the Google Ads API. Obtaining one means holding a Google Ads manager account and filing an API Center application.

Merchants without a technical person usually stall right here.

What the Official Google Ads MCP Server Cannot Do

Can Google Ads MCP Pause Campaigns or Change Bids

Google's developer documentation is blunt about scope. Verbatim: "This implementation is strictly read-only. It cannot modify bids, pause campaigns, or create new assets."

The launch blog says the same thing. That first release is read-only, usable for reporting and diagnostics. It makes no changes to your account.

Any page promising campaign launches through Google's own MCP server is describing something else. Third-party servers can wrap the write endpoints of the Google Ads API. Google's server does not.

Google Ads API Daily Operation Limits

New tokens are issued at Test Account Access first. Explorer access is the next level up, and the first that can touch a live account. It caps production accounts at 2,880 API operations per day.

Basic access lifts that ceiling to 15,000 operations per day, and it requires an application review. Conversational analysis burns through either allowance faster than a nightly report does. Every follow-up question is another round of queries.

Exceeding the daily limit returns a RESOURCE_EXHAUSTED error. Rejected requests still count against the user's daily operation quota, so a badly formed query loop costs you twice.

What Data Google Ads MCP Sends to the Model

Google's own repository carries a warning worth reading twice. Word for word: "The MCP Server will expose your data to the Agent or LLM that you connect to it." Ad spend, campaign names, product names and conversion values all travel across that connection.

Deciding which assistant sees that is a governance question, not a technical one. Settle it before the first query runs.

How to Analyze Google Ads Through MCP

Diagnostic Questions to Ask Google Ads MCP

Reporting through a chat connector is a poor use of it. Pulling a 30-day spend table into a conversation just recreates a dashboard with worse formatting. Diagnosis is where a read-only connection earns its keep.

Useful prompts sound like questions an analyst would ask out loud. Which campaigns lost conversion value week over week, and what changed in their budgets? Which asset groups carry spend but almost no conversions?

Answers to those come out of one or two well-formed GAQL calls.

Merchants running Shopify catalogs have a second layer to check, since campaign structure and feed health interact. Our guide to running Google Ads for a Shopify store covers what usually breaks first.

Why Optimization Score Is Not a Performance Target

Optimization score is Google's own estimate of how well your account is set to perform. It runs on a 0 to 100% scale.

The score covers fewer campaign types than most merchants assume. Only active Search, Display, Video Action, App, Performance Max, Demand Gen and Shopping campaigns are included.

An account reaches a 100% score by applying or dismissing every recommendation. Dismissing counts as much as applying does.

That makes the number a housekeeping metric rather than a performance one. It is a poor target for an AI assistant.

Who Applies the Change After the Analysis

Read-only analysis ends with a decision, and a human still has to act on it. Two patterns work in practice.

One route is to carry the recommendation into the Google Ads interface yourself. Another is a product with an explicit apply step.

Whichever route you pick, the economics have to be settled first. Judge every proposed change against your break-even ROAS, and sanity-check the account total against blended MER. Without those two numbers, an assistant will happily optimize toward a metric that loses you money.

Google Ads MCP Compared With Meta's Ads MCP

Meta took the opposite approach. Its ads MCP server is remote hosted by Meta at mcp.facebook.com/ads, and it does support write actions.

The developer announcement went out on 16 July 2026, with campaign create, edit and delete in scope. Both of those are vendor claims, taken from Meta's own developer documentation and blog.

Running both platforms through one assistant therefore means two different trust models. One connector can only read your account. The other can change things.

Our walkthrough on connecting Claude to Meta and Google Ads covers the practical setup. For the strategic split, see our Google Ads versus Facebook Ads comparison.

Merchants spending their own money often prefer that nothing moves without a click. A connector that can only read is a smaller blast radius by design.

Agency AI as a Google Ads MCP Option

Agency AI describes itself as the Meta and Google Ads MCP for Shopify. Its connector links ad accounts and store data to Claude or ChatGPT.

The company's MCP page positions the product around answers in chat rather than report building. That page puts it this way: "No dashboards, no filters, no export. Ask in plain words and get the answer, the chart, and what to do about it."

There is no separate MCP fee. Its own page says the connector comes included with an Agency AI account at no added cost. It also states that the connector cannot spend money or change your ads, which matches the read-only framing above.

All of that is vendor material from the company. Agency AI also says its agentic engine applies 90-plus rules across Meta and Google as a single funnel. Treat that as a company statement rather than documented fact.

The two platforms then handle recommendations differently. Google recommendations carry an Apply action, and the change is made in Google Ads. Meta recommendation cards carry no Apply button, and nothing on Meta is ever auto-applied.

Its Shopify App Store listing is named Agency AI: Meta, Google Ads AI. That listing launched 27 October 2025 and is rated 5.0 from 4 reviews. Essentials costs $59 per month or $492 per year, with a 30-day free trial.

A Claude and ChatGPT MCP connector is named on that listing as a shipped feature. Anyone comparing options should also read our AdKit alternatives rundown.

Published results exist for a single store. Lolahemp's Meta cost per purchase was $143.75 from 1 August to 30 September 2025, on 33 purchases. It fell to $63.56 from 10 October to 6 November 2025, on 95 purchases.

That case study page itself flags that the windows are not the same length. Neither screenshot is a full account export.

Frequently Asked Questions

Can the official Google Ads MCP server pause a campaign?

No. Google's developer documentation states the implementation is strictly read-only. It cannot modify bids, pause campaigns, or create new assets. Any change still has to be made by a person or by a separate tool with write access.

What tools does the official Google Ads MCP server include?

Three of them ship today. list_accessible_customers returns the customer IDs and account names your login can reach. search runs Google Ads Query Language requests for metrics, budgets and status. get_resource_metadata returns metadata for a resource type such as campaign.

Do I need a developer token to use Google Ads MCP?

Yes. Running the official server requires a developer token plus OAuth credentials from the Google Ads Developer console. A developer token is a 22-character alphanumeric string. Getting one means holding a Google Ads manager account and filing an API Center application.

How many Google Ads API operations do I get per day?

Explorer access allows 2,880 operations per day for production accounts. Basic access raises that to 15,000 operations per day, and it requires an application review. Going over returns a RESOURCE_EXHAUSTED error, and rejected requests still count against the daily quota.

Is optimization score a good target for an AI assistant?

No. An account reaches 100% by applying or dismissing every recommendation, so the score measures housekeeping rather than performance. It also covers only active Search, Display, Video Action, App, Performance Max, Demand Gen and Shopping campaigns.

Is Meta's ads MCP more capable than Google's?

Meta's server is remote hosted and supports write actions, according to Meta's own documentation. Google's official server is read-only. Capability is not the same as safety. Read-only is the easier thing to hand an assistant on a live spending account.

Sources

Release date, read-only scope and the reporting-and-diagnostics framing come from the Google Ads Developer Blog post of 7 October 2025. Google's developer documentation for the MCP server supplies the tool names, the resource list and stdio transport. The googleads/google-ads-mcp repository on GitHub covers credential requirements, the Apache-2.0 license and the pipx install.

Developer token length and eligibility are set out in Google's own developer token guide. GAQL scope comes from Google's query language overview. Access levels and daily operation limits sit on Google's access levels page. That page also names Test Account Access as the starting point.

Its quotas best-practices page documents the RESOURCE_EXHAUSTED behavior. Google Ads Help defines optimization score, its campaign coverage and the 100% mechanic. MCP's own definition sits on modelcontextprotocol.io.

Meta's hosted, write-capable server and its 16 July 2026 developer announcement come from Meta's own developer documentation and developer blog. Both are vendor claims. Agency AI product statements are drawn from its MCP page, Shopify App Store listing and published Lolahemp case study.

The 90-plus rule engine is a statement the company made directly to us, and it is not on those pages. All of it is vendor material.

Free to install. Launch campaigns and pay nothing for your first 30 days.