All posts

AI Google Ads Claude MCP

How to Analyze and Optimize Google Ads With Claude

September 2, 2026 Agency AI Team 10 min read

META TITLE: How to Analyze Google Ads With AI Using Claude

META DESCRIPTION: Connect Claude to Google Ads through MCP, run four copy-paste prompts, and read the four failure modes that turn AI ad analysis into a bad call.

SLUG: analyze-google-ads-with-ai

TAGS: Google Ads, Claude, MCP, AI advertising, PPC analysis, Shopify

You can analyze Google Ads with AI by connecting Claude to your ad account through MCP, then asking questions in plain English.

Good output names the campaign rows behind every recommendation and admits when data is missing. Bad output invents a benchmark, then tells you to cut a campaign that is still learning.

This guide gives you the connection setup, four prompts you can paste today, and the shape of an answer worth acting on. It then covers the four failure modes behind most bad calls, and every step stays read-only until you approve.

How Claude Connects to Your Google Ads Account

MCP is the piece that makes any of this possible. The Model Context Protocol documentation defines it as "an open-source standard for connecting AI applications to external systems." Under that standard, Claude reaches data sources, tools and workflows, so it can pull information and perform tasks.

What a Connector Sees in Your Ad Account

A connector inherits your own login rights. Per Anthropic, "Claude inherits your permissions in that tool and can only see and do what you can already see and do." Give it a read-only ad account seat and you get a read-only Claude.

Custom connectors using remote MCP run on Free, Pro, Max, Team and Enterprise plans, with directory verification covering quality and compatibility only. Anthropic adds a warning worth keeping: "Only connect to trusted servers."

Which Google Ads MCP Connector to Use

Google publishes an open-source MCP server for the Google Ads API, exposing three query tools: search, get_resource_metadata and list_accessible_customers. Its README claims no read-only status. The Ads Developer Blog announcing it does, calling the initial release reporting and diagnostics that will not change your account.

Setup calls for a developer token with at least Explorer access, which puts it outside a two-click install.

Meta went the other way. Its hosted server, in the overview's wording, "lets you manage Meta ads directly from an AI agent". That same page lists what it does: "Create and edit ads, ad sets, and campaigns."

On a write path, a bad answer becomes a bad change. Our roundup of marketing MCP servers weighs the read and write options side by side.

How Claude Turns Your Question Into a GAQL Query

Behind every plain question sits a query language. Google's developer docs say GAQL allows "querying the Google Ads API for resources, their attributes, segments, metrics, and metadata". Claude drafts that query, runs it, then narrates whatever comes back.

Two things follow. An ambiguous question produces an ambiguous query, and a vague answer after it. You can also ask to see the query it wrote, and you should.

Four Prompts That Produce Usable Google Ads Analysis

Prompts for Wasted Spend and Negative Keywords

Start with money leaving the account:

> Using only the campaign rows you retrieved, list campaigns from the last 30 days where spend rose and conversion value fell. Show spend, conversions and conversion value for each. If a figure is missing, say so instead of estimating.

Then go after queries. Google's Help Center defines the search terms report as a list of terms that resulted in your ad being shown. That makes it the right input for negatives:

> Pull the search terms report for the last 30 days. Flag negative keyword candidates with zero conversions and spend above my break-even threshold. Quote the exact term and its spend for each candidate.

Both prompts share one instruction worth keeping permanently. Claude must quote the row, or it does not get to make the claim.

Prompts for Performance Max and Attribution

Performance Max needs its own question. Google calls it a goal-based campaign type that reaches all Google Ads inventory. Coverage spans YouTube, Display, Search, Discover, Gmail and Maps from one campaign.

Channel-level reporting inside that single campaign is limited, which is why asset groups are the more useful cut. Google's help pages describe them as "collections of creative assets, like images, text, logos, and videos", each built around a theme or audience:

> For each Performance Max campaign, break results down by asset group. Rank asset groups by conversion value and by cost. Name each asset group exactly as it appears in the account.

Creative decisions follow from that ranking, and our guide to creative fatigue covers the next move. Finally, force the attribution question before any ROAS answer:

> Before answering, state which attribution model and conversion window these numbers use. Then compare ROAS by campaign for the last 30 days against the previous 30.

Why bother? Because conversion counts move with the model. Attribution models "let you choose how much credit each ad interaction gets".

What Good Claude Output Looks Like

How to Make Claude Cite Campaign Data

Anthropic recommends making a response "auditable by having it cite quotes and sources for each of its claims." Applied to ads, each recommendation names the campaign, the date range and the metric values.

Two further techniques from that same page belong in your system prompt. Give permission to be unsure, since the guidance says to "Allow Claude to say 'I don't know'". Then apply what Anthropic calls external knowledge restriction, which keeps the model off its general knowledge.

That second instruction kills invented benchmarks. Without it, an answer can quote an industry CPA that came straight out of training data. Real figures for your vertical belong in our 2026 Shopify ad benchmarks instead.

How to Make Claude Calculate ROAS With Code

Claude can run JavaScript on the numbers. Anthropic describes its analysis tool as a built-in code sandbox for complex math and data work. Send an exported CSV there whenever the output is a ROAS or CPA you plan to act on.

Cheap verification is available too. One documented check is "Best-of-N verification: Run Claude through the same prompt multiple times and compare the outputs." Inconsistent answers across runs are your cue to go back to the rows.

A caveat comes from Anthropic directly. These techniques "don't eliminate them entirely", so validation stays your job on anything expensive.

Where AI Google Ads Analysis Goes Wrong

Why Yesterday's Google Ads Data Is Incomplete

Conversion reporting lags can run up to 24-48 hours, a figure Google publishes in its conversion tracking documentation. Treat conversions and conversion value from the last two days as provisional, even when spend already looks final. Default to a 30-day window when you frame the question.

How Many Conversions Target ROAS Needs

Target ROAS carries a data floor. Google requires at least 15 conversions in the past 30 days at the conversion tracking level for Search and Shopping campaigns. That is an account-level bar, so a campaign sitting well under its share of it will not produce a ROAS number worth comparing.

Target ROAS is itself an AI product. Google's documentation says the strategy uses "Google's AI to predict the value of a potential conversion" on every search. Platform AI is already bidding, and your analysis AI sits one layer above and reviews what that bidding did.

Why the Search Terms Report Hides Queries

Privacy rules strip low-volume queries out. Terms without enough query activity "are omitted from the search terms report", according to Google, to meet "our standards on data privacy". Any wasted-spend answer therefore comes from a partial list, so add negatives one at a time, checking spend on each.

How Often to Change Google Ads Campaigns

Over-optimizing has a measurable price. Google warns that frequent manual adjustments to budgets or targets "will reset the standard 7-14 day learning period" and delay optimization. Batch your changes into one weekly pass.

Recommendations from Google follow the same approve-first pattern. Optimization score is an estimate of how well your Google Ads account is set to perform, running from 0-100%.

Recommendations are applied from the Recommendations page. Check your auto-apply settings before you trust that, because Google can apply opted-in recommendation types on its own.

How to Apply a Claude Recommendation Safely

Hold one rule above the rest: analysis is automated, execution is not. Read the answer, check the rows it cited, then make the change yourself inside the platform.

A workable weekly loop runs about twenty minutes. Ask the four prompts above, keep only recommendations backed by quoted rows, and discard anything without a named campaign. Apply the survivors in a single batch, then leave the account alone through the learning period.

Break-even ROAS is what makes any of this actionable. Our break-even ROAS calculator sets the threshold. MER shows whether blended spend still works, and incrementality testing shows whether the lift was real.

Running the Workflow on Shopify With Agency AI

Shopify merchants can skip the developer setup. Agency AI's App Store listing says you can "Connect Claude or ChatGPT via MCP to analyze, plan, forecast, and execute in chat."

Their own positioning line reads: "No dashboards, no filters, no export. Ask in plain words and get the answer, the chart, and what to do about it." Both lines are the vendor's own.

Those two surfaces do not quite agree on execution, and the MCP page is the more specific one. Its FAQ answers the read-only question with a hedge: "Yes. Right now it reads and explains: analysis, plans, and clear next steps." Agency AI's stated position today, then, is analysis and recommendations, with launching, editing and pausing left to you.

Pricing for the MCP plan sits separately from the $59 per month app. Check agencyai.app/mcp for the current figure.

The positioning line above is about not having to build a report to get an answer. Agency AI states the plan still includes a free analytics dashboard. Use it whenever you want the numbers on a screen.

Listing facts for the wider app: "Agency AI: Meta, Google Ads AI", rated 5.0 from 4 reviews. It went live on 27 October 2025 and offers a 30-day free trial.

A step-by-step setup guide lives in our post on connecting Claude to Meta and Google Ads. Store-specific campaign structure is covered in Google Ads for Shopify.

Frequently Asked Questions

Can Claude change my Google Ads campaigns?

Only if the connector you installed ships write tools, and most do not. A read-only connector can still tell you to pause a campaign or add a negative keyword. That change then happens in the Google Ads interface, under your own login, and the gap between the two is where your safety lives.

How far back should I ask Claude to look?

Use 30 days as your default and 90 days when you want a seasonality read. Google's conversion reporting lag of up to 24-48 hours means the last two days will still move after you look at them. Compare like for like: a 30-day window against the previous 30, never against a partial week.

Why do Claude's ROAS numbers differ from the Google Ads interface?

Usually the attribution model, the conversion window or the date range differs. Attribution models let you choose how much credit each ad interaction gets for a conversion. Ask Claude to state the model and window before it reports any ROAS figure.

Do I need a paid Claude plan to use an ads connector?

Not for the platform capability, and that is unusual for this kind of integration. Custom connectors using remote MCP are documented on Free, Pro, Max, Team and Enterprise plans, with free users limited to one custom connector. Individual connector providers can set their own plan requirements, so check the one you intend to use.

How do I stop Claude inventing benchmark numbers?

Paste your export into the conversation and tell Claude that file is the only source it may use. External knowledge restriction is the documented name for that technique, alongside giving the model explicit permission to say it does not know. Anthropic is clear that these techniques reduce hallucinations and do not eliminate them entirely, so verify anything expensive.

Sources

modelcontextprotocol.io, the standard's own documentation site, carries the protocol definition. Anthropic supplies the connector behavior. Two support articles at support.claude.com cover remote MCP custom connectors and MCP connectors, and claude.com documents the Connectors Directory.

Prompting techniques trace to Anthropic's guidance on reducing hallucinations at platform.claude.com, and the Anthropic blog describes the analysis tool. Server capabilities are documented in the googleads/google-ads-mcp repository on GitHub and in Google's Ads Developer Blog post announcing it. Meta's ads MCP server overview on developers.facebook.com covers the write side.

Google Ads Help documents the conversion reporting lag, the learning period and the Target ROAS requirements. That same Help Center covers search terms scope, attribution models, optimization score, auto-apply recommendations and Performance Max. Query language wording is taken from developers.google.com.

Agency AI's Shopify App Store listing and its own MCP page supply the product facts. MCP plan pricing is not listed above, since agencyai.app/mcp is the live source for that figure.

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