Google Ads MCP and AI Agents: What Actually Works

The short answer

“AI agent for Google Ads” means three different things, and the difference decides whether anything in your account can change.

Which means the honest headline is one most articles avoid: nothing official from Google will touch your campaigns. If you want an agent that pauses an ad group or moves a bid, you are leaving Google’s own tooling to get it.

That is not a complaint. It is a design decision, and once you know about it you stop looking for a setting that does not exist.

Google’s own MCP server

Google publishes an open-source MCP server for the Google Ads API. Its documentation says this, and it is worth quoting exactly rather than paraphrasing:

“This implementation is strictly read-only. It cannot modify bids, pause campaigns, or create new assets.”

Three tools, and that is the whole surface:

• list_accessible_customers — the Google Ads accounts your login can reach, with their IDs and names.
• search — runs a Google Ads Query Language request to pull metrics, budgets and status.
• get_resource_metadata — describes a resource type, for example what fields a campaign has.

Almost everything useful happens through the middle one. GAQL is Google’s query language for ads data, roughly the way SQL is for a database. Normally you would have to write it. With this connected, you describe what you want in ordinary words and the AI writes the query.

It runs on your machine, not Google’s

This is the part that surprises people coming from Meta. Meta hosts its server, so you paste an address and sign in. Google does not host this for you. You install it, in Python, and it talks to your AI app over standard input and output on your own computer. There is also a documented option to deploy it to Google Cloud Run and reach it over HTTP if you want it hosted.

Authentication is OAuth 2.0 or a service account. The code lives in Google’s google-ads-mcp repository.

The practical consequence: this is a developer-shaped setup, not a two-minute one. If that puts you off, the honest alternative is the next section, not a workaround.

Ask Advisor, inside Google Ads

On 10 August 2026 Google announced Ask Advisor, which it describes as “our in-product AI agent across Google’s marketing platforms.” No setup, because it is already in the interface.

What it does today: insight cards on your homepage personalised to your business, a prompt box for asking your own questions of your data, and dashboards built from a text prompt, which Google lists as coming soon.

Two things to be clear about. First, it explains and recommends; it does not go and make the changes. You still act. Second, at announcement it was in beta and limited to English-language accounts, so whether you can see it at all depends on your account.

For most advertisers this is the sensible starting point, precisely because there is nothing to install.

The ones that can actually change things

Search for a Google Ads AI agent and most results are third-party servers and tools. Some can write to your account, which is exactly what Google’s own server refuses to do.

We are not going to rank them. We have not audited them, and a list of tools we have not tested is not worth your time. What is worth your time is knowing what you are deciding.

• Who holds the access. With Google’s server the connection is between you and Google. With a third party, a company sits in the middle of your ad account.
• Whether it can write. That is the reason to use one, and the reason to be careful. Google’s server cannot make a mistake that costs money. A writing agent can.
• Whether it asks first. Whether you get a confirmation before something changes is decided by the tool, not by the protocol. We explain why in how the AI actually works.

On a client account, this is a decision to make in writing, not by installing something on a Tuesday.

Google and Meta, side by side

If you have already connected Meta, this comparison will save you some confusion, because the two are built on opposite assumptions.

Neither is better. They answer different questions. Google’s says let the AI read everything and change nothing. Meta’s says let the AI work, and we will give you rules to govern it. If you want the Meta side, that is the setup walkthrough.

What it is genuinely good at

Read-only sounds like a limitation until you use it. Most of the hours in a Google Ads account go into looking things up, not clicking things.

• Questions the interface makes tedious. Which campaigns spent more than last month while converting less. You would build that report. Here you ask for it.
• Cross-account comparisons. It can list every account your login reaches, so “compare these four clients” is one request rather than four exports.
• GAQL without learning GAQL. The AI writes the query. You never see it unless you ask.
• Explaining the number after it fetches it. Not just what impression share was, but which part of the account moved it.

And there is a real safety dividend. Because the server cannot write, the worst outcome of a misunderstood request is a wrong answer, not a wrong campaign. That makes it a much easier thing to hand to someone still learning the account.

Questions worth actually asking it

“Read-only” is easy to nod at and hard to picture, so here is the concrete version. These are shaped for the one tool that does the work, search, which means anything the Google Ads Query Language can return is fair game.

The pattern that gets a useful answer every time: name the account, bound the dates, say what to compare against, and say what shape you want the answer in.

• Spend that stopped converting. “For account 123-456-7890, list campaigns where cost rose more than 20% this month versus last while conversions fell. Show cost, conversions and cost per conversion for both periods.”
• Where the budget actually goes. “Rank ad groups in campaign X by spend for the last 30 days, and next to each, its share of total conversions. I am looking for spend concentrated where results are not.”
• The account inventory nobody has time to do. “List every campaign across all accounts I can access, with status, budget and last 7 days of spend.” This is the one that uses list_accessible_customers, and it is the single thing that takes longest to do by hand.
• Search terms against the negatives you already have. “Search terms in the last 30 days with more than 10 clicks and zero conversions for campaign X.”
• What is quietly limited. “Which campaigns are budget-limited right now, and how much spend did each lose to that in the last 14 days?”
• A structure check before you inherit an account. “Describe this account: campaign count by type and status, how many ad groups per campaign, and which campaigns have no active ads.”

Two habits that make the answers more trustworthy:

• Ask for the query. “Show me the GAQL you ran” turns an answer you have to trust into one you can check, and it is how you learn what the data actually contains.
• Ask what the number excludes. Date range, account, campaign status and conversion window all silently shape a figure. A number that will drive a decision is worth one follow-up question.

Stated plainly: the prompts above are ours. Google documents the three tools and the read-only boundary; it does not publish question templates.

What it will not do for you

• It will not optimise anything. No bid changes, no pausing, no budget moves. It can tell you what you should probably do. Then you go and do it.
• It will not build campaigns. Creating assets is explicitly excluded.
• It will not fix a suspension. Account-level enforcement is a different problem entirely.
• It does not remove the need to know Google Ads. It answers what you ask. Asking the wrong question gets you a confident, accurate, useless answer.
• It is not a set-and-forget system. Nothing here runs on a schedule. You ask, it answers.

The gap between what people expect from “AI agent” and what the official tooling does is wide, and it is worth closing before you spend an afternoon looking for the button.

FAQ

What is the best AI agent for Google Ads?

It depends what you need it to do. For asking questions of your data with no setup, Ask Advisor inside Google Ads. For deeper analysis through Claude or another AI app, Google’s official MCP server, which is read-only. For anything that actually changes campaigns, a third-party tool, with the trade-offs that brings. There is no official Google option that edits your account.

Can an AI agent run my Google Ads for me?

Not through anything Google publishes. Google’s own MCP server states plainly that it cannot modify bids, pause campaigns or create assets, and Ask Advisor advises rather than executes. Only third-party tools write to your account.

How do I connect Claude to Google Ads?

Through Google’s open-source MCP server. Unlike Meta’s, it is not hosted for you: you install it in Python and it runs locally, or you deploy it to Google Cloud Run and connect over HTTP. Authentication is OAuth 2.0 or a service account.

Is the Google Ads MCP server free?

The server is open source and published by Google in its google-ads-mcp repository. You still need Google Ads API access, and if you deploy it to Cloud Run you pay for that hosting.

Why is Google’s version read-only when Meta’s is not?

Google does not explain the reasoning in the documentation, so we will not invent one. What is documented is the outcome: Google’s server reads, Meta’s reads and writes, and Meta added portfolio-level rules to govern what agents may do.

What can it actually tell me?

Anything the Google Ads Query Language can return: spend, impressions, clicks, conversions, budgets, campaign and ad group status, and the structure of the account. It can also list every account your login can reach, which makes multi-account questions quick.

Do I need to know GAQL?

No. Writing the query is the job you are handing over. It helps to know what data exists, because that shapes what you think to ask.

Is it safe to connect this to a client account?

Safer than the writing alternatives, and that is the point of the design. Because the server cannot modify bids, pause campaigns or create assets, the worst outcome of a misunderstood request is a wrong answer rather than a wrong campaign. What it still has is read access to that client’s data, so the access itself is worth agreeing in writing even when nothing can be changed.

Can I make it run a report every week automatically?

Not by itself. Nothing here runs on a schedule; you ask and it answers. Scheduling would mean building something around the server rather than using it as documented, which puts you back in developer territory.

Should I use Ask Advisor or the MCP server?

Ask Advisor if you want answers with nothing to install, accepting that it was in beta and limited to English-language accounts at announcement. The MCP server if you want to ask through your own AI app, compare several accounts in one request, or work across data the interface makes tedious — and you are willing to install and authenticate it. Neither changes your campaigns, so it is a question of convenience rather than risk.

Sources

• Google Ads API — MCP server documentation. Source of the read-only quote, the three tool names and what each does, OAuth 2.0 and service account authentication, the local and Cloud Run deployment options, and the repository link.
• googleads/google-ads-mcp on GitHub. Google’s published source for the server.
• Google — Try new AI tools from Google Ads and Analytics (10 August 2026). Source of the Ask Advisor description, the insight cards and prompt box, dashboards being listed as coming soon, and the beta and English-language limitation.
• The Meta comparison comes from our own sourced write-ups: how the AI actually works and the Meta setup walkthrough, both of which quote Meta’s documentation directly.

Checked on 17 August 2026, and the read-only wording was re-read on Google’s page immediately before publishing, because that single sentence is what this article rests on. Two things we have deliberately not stated: a launch date for the server, which we saw only in third-party write-ups and not in Google’s own documentation, and any ranking of third-party tools, which we have not tested. Both products are moving, so check the source pages before acting.

If you would rather someone ran the account than the tooling, we manage Google Ads, or get in touch.