Skip to main content
AI Marketing

How to use Claude with an MCP server to optimize Google Ads

Connect Claude to Google Ads with an MCP server: three setup options, safe permissions, 35 optimization prompts, GAQL queries, and a repeatable workflow.

25 min read

Using Claude with an MCP server to optimize Google Ads means connecting Claude, Anthropic's AI assistant, directly to your Google Ads account through the Model Context Protocol (MCP), so it can pull live performance data, find wasted spend, explain why results changed, and, with your approval, make changes like adding negative keywords. Instead of exporting CSVs and pasting them into a chat, you ask a question in plain English and Claude queries the account itself.

This guide covers everything you need to do it well: what MCP is, the three ways to connect Claude to Google Ads (including the hosted option Markifact and Google's own open-source server), how to set permissions so an AI never makes a change you didn't approve, 35 prompts organized by optimization task, the GAQL queries worth knowing, and the mistakes we see most often. At Sloss Digital we built our own MCP server that connects Claude to Google Ads for the accounts we manage, so this guide comes from daily use, not theory.

In this guide:

What does it mean to connect Claude to Google Ads with MCP?

On its own, Claude can only work with what you paste into the chat. Connected through MCP, Claude gets a set of tools it can call on its own, such as "run this Google Ads report" or "list the accounts I can access." When you ask "Which search terms wasted the most money last month?", Claude:

  1. Picks the right tool and writes the query (for Google Ads, usually a GAQL query).
  2. Runs it against your live account through the Google Ads API.
  3. Reads the results, does the math, and answers in plain English, showing the data behind its answer.
  4. If you ask for changes, proposes them and waits for your approval before anything is written to the account.

The practical difference is speed and depth. A search terms review that takes an hour of exporting, filtering, and pivoting becomes a two-minute conversation, and you can follow up immediately: "Now check whether any of those terms converted in the last 90 days."

What is an MCP server?

The Model Context Protocol (MCP) is an open standard that Anthropic introduced in November 2024 for connecting AI assistants to outside tools and data. It has three parts:

  • The host is the AI app you use: the Claude web and desktop apps, or Claude Code in a terminal.
  • The MCP server is a program that exposes tools (actions the AI can take), and optionally resources and reusable prompts, for one service, such as Google Ads.
  • The connection between them handles discovery ("here are the tools available") and authentication.

Because MCP is a standard, the same Google Ads MCP server can work with Claude, ChatGPT, Gemini, Cursor, and other MCP-compatible tools. It also means you're not locked into one vendor: you can switch servers or AI apps without rebuilding your process.

MCP servers come in two flavors:

  • Remote servers run on the internet and connect over HTTPS. In Claude, you add them as custom connectors, and they work across the Claude web app, desktop app, and Claude Code. Hosted services like Markifact are remote servers.
  • Local servers run on your own computer. You configure them in Claude Desktop or Claude Code, and they aren't available in the Claude web app.

What can Claude do in Google Ads through MCP?

What Claude can do depends on the tools your MCP server exposes and the permissions you grant. With a full-featured server, the work falls into four groups:

TaskRead or writeWhat Claude does
Account auditsReadReviews structure, bidding, budgets, conversion actions, match types, and ad coverage, then ranks issues by wasted spend or missed opportunity.
Search term miningRead, then writeGroups irrelevant search terms into themes, proposes negative keywords, and finds converting terms to add as keywords.
Performance diagnosisReadExplains why cost per conversion moved by comparing periods, segments, and the change history.
Budget and bid reviewRead, then writeChecks pacing, impression share lost to budget versus rank, and targets versus actual results.
Ad copy and assetsRead, then writeFlags weak responsive search ads and missing assets, then drafts new headlines and descriptions within character limits.
Conversion tracking checksReadLists conversion actions and flags miscounting, such as page views set as primary conversions.
ReportingReadWrites plain-English summaries for clients or owners, leading with leads and cost per lead.

A good rule: let Claude read everything, and let it propose changes freely, but make every write go through a human approval step. We cover how in the safety section.

What you need before you start

Here's the checklist:

  • A Claude account. Remote MCP servers can be added as custom connectors on every Claude plan, though Free accounts are limited to one. On Team and Enterprise plans, an owner adds the connector under Organization settings, then members connect to it. For local servers you'll need Claude Desktop or Claude Code.
  • Access to the Google Ads account(s). You'll sign in with a Google account that already has access. For agencies, that's usually through a manager (MCC) account.
  • An MCP server. Hosted (like Markifact), Google's open-source server, or your own. Details in the next section.
  • Clear goals for Claude. Your target cost per lead or ROAS, what counts as a real conversion, average deal value, service area, and anything off-limits (for example, "never touch the brand campaign").

If you run your own server (either Google's open-source one or a custom build), you need Google Ads API access. That process changed: Google sunset developer tokens on September 9, 2026. API access is now tied to the Google Cloud project behind your OAuth credentials, and you no longer need a developer token or a manager account to get started. Access comes in four levels:

Access levelProduction accountsDaily operationsHow you get it
TestNo (test accounts only)15,000 on test accountsAutomatic when you enable the Google Ads API in your Google Cloud project
ExplorerYes2,880 on production accountsApply on the Google Ads API Overview page in Google Cloud
BasicYes15,000Apply after completing brand verification for your project
StandardYesUnlimitedApply from Basic; includes a manual review (typically about 10 business days)

Explorer access is plenty for one person analyzing a handful of accounts. Agencies running daily reviews across many clients will want Basic or Standard. See Google's access levels documentation for the current details. Hosted services like Markifact handle API access for you, so none of this applies if you go that route.

Which Claude model should you use?

  • Claude Opus 5.5 for full account audits, diagnosing a performance drop across many campaigns, and anything where the quality of reasoning matters more than speed.
  • Claude Sonnet 5.5 for faster everyday work, such as weekly search term reviews, pacing checks, and report drafts.

Model names are current as of November 2026. Whichever you use, the quality of the answer depends far more on the context you give Claude (goals, margins, what a good lead looks like) than on the model.

Three ways to connect Claude to Google Ads

FactorHosted server (Markifact)Google's open-source serverBuild your own
Setup timeMinutesAn hour or twoWeeks
Skill neededNoneComfortable with a terminal and Google CloudA developer
Google Ads API accessHandled for youYour own Google Cloud projectYour own Google Cloud project
Changes to the accountYes, with an approval stepAssets and labels only, with propose-and-confirmWhatever you build
Works in the Claude web appYesOnly if you deploy it as a remote serverYes, if hosted remotely
Other ad platformsYes (Meta, Microsoft, TikTok, and more)NoWhatever you build
CostSubscription plus usage creditsFree software; you run itDevelopment time plus hosting
Best forMarketers and agencies who want it working todayTechnical teams who want control and mostly read accessAgencies with their own playbooks and guardrails

Option 1: A hosted MCP server like Markifact

Markifact is an AI marketing automation platform with a hosted MCP server. You connect your Google Ads account inside Markifact, then add one URL to Claude. It handles authentication and API access, supports Google Ads alongside more than a dozen other ad platforms (including Meta, Microsoft, TikTok, and Snapchat) plus GA4 and Search Console, and lets you restrict access by account and set any connection to read-only. Write actions go through a separate tool that asks for confirmation.

At the time of writing, Markifact offers a free trial with credits and paid plans starting at $79 per month, with reports and most data operations costing one credit each. Check their site for current pricing.

Option 2: Google's open-source Google Ads MCP server

Google's Ads developer organization publishes an open-source Google Ads MCP server on GitHub. It's a Python server you run yourself, with tools for running GAQL queries, listing accessible accounts, and managing assets and labels. It's free, transparent, and a good choice if you want full control and mostly need read access, but you'll set up your own Google Cloud project and credentials.

Option 3: Build your own

Building your own server gives you complete control over which actions exist, what guardrails apply, and how your team's playbook is encoded. It's the route Sloss Digital took. It's also the most work, so it makes sense mainly for agencies and in-house teams managing significant spend across many accounts.

How to connect Markifact to Claude, step by step

  1. Create a Markifact account and connect your Google Ads account (and any other platforms) inside Markifact.
  2. Create an MCP server in Markifact. Open MCP in the sidebar and click New MCP Server. Give it a name, such as "Claude - Client A."
  3. Set permissions before connecting anything. Under access, start with Read only. Under connections, choose Custom, tick your Google Ads connection, and set Accounts to Selected so Claude can only see the accounts you choose. Each level can only narrow access, never widen it.
  4. Add the connector in Claude. In Claude, go to Customize → Connectors, click + Add, then Add custom connector. Name it and paste the server URL: https://api.markifact.com/mcp. (On Team and Enterprise plans, an owner adds it under Organization settings first.)
  5. Sign in. Claude redirects you to Markifact to authorize access over OAuth. No tokens to copy.
  6. Set tool permissions in Claude. Allow the read-only tools, and set the write tool, run_write_operation, to Needs approval. Never set it to always allow.
  7. Test with a read-only question: "List the Google Ads accounts you can access and their spend over the last 7 days."

Prefer Claude Code? Add the same server from your terminal, then run /mcp inside Claude Code to sign in:

claude mcp add --transport http markifact https://api.markifact.com/mcp

A useful Markifact pattern for agencies is one MCP server per client: a Custom connection with only that client's accounts selected. Claude then physically can't see or change another client's account, no matter what you ask it.

How to run the open-source Google Ads MCP server from Google

This option suits technical marketers and developers. The project README is the source of truth; here's the shape of it.

What it exposes. Tools named search (run GAQL), get_resource_metadata, list_accessible_customers, create_asset, upload_video_asset, manage_labels, and manage_label_associations, plus reference resources for metrics, segments, and release notes. The write tools use a propose-and-confirm flow: calls default to validation only, and committing a change requires the same parameters plus a checksum from the proposal. To run it strictly read-only, turn off the label and asset tools in its tools_config.yaml.

Setup in brief:

  1. Enable the Google Ads API in a Google Cloud project (this grants Test access), then apply for Explorer access or higher for production accounts.
  2. Choose a credential method: Application Default Credentials with the Google Ads (adwords) scope, an OAuth client, or an existing google-ads.yaml from Google's Python client library.
  3. If you access client accounts through a manager account, set the login customer ID (GOOGLE_ADS_LOGIN_CUSTOMER_ID) or pass it per call.
  4. Run the server with pipx, and add it to Claude.

Adding it to Claude Desktop. Local servers go in Claude Desktop's claude_desktop_config.json:

{
  "mcpServers": {
    "google-ads": {
      "command": "pipx",
      "args": [
        "run",
        "--spec",
        "git+https://github.com/googleads/google-ads-mcp.git",
        "google-ads-mcp"
      ],
      "env": {
        "GOOGLE_APPLICATION_CREDENTIALS": "/path/to/credentials.json",
        "GOOGLE_PROJECT_ID": "your-google-cloud-project-id"
      }
    }
  }
}

Adding it to Claude Code:

claude mcp add --env GOOGLE_APPLICATION_CREDENTIALS=/path/to/credentials.json \
  --env GOOGLE_PROJECT_ID=your-google-cloud-project-id \
  --transport stdio google-ads \
  -- pipx run --spec git+https://github.com/googleads/google-ads-mcp.git google-ads-mcp

Because this runs locally, it won't appear in the Claude web app. To use it there, deploy it as a remote server (the repo documents a Cloud Run deployment with OAuth) and add its URL as a custom connector.

Building your own Google Ads MCP server

Sloss Digital built its own MCP server that connects Claude to the Google Ads API for the accounts we manage. If you're weighing the same route, these are the design decisions that matter most:

  • Start read-only. Ship reporting tools first and use them for weeks before adding any write tool.
  • Build purpose-made tools, plus one raw query tool. Tools like "get search terms with spend above X" return clean, compact data and are hard to misuse. A general GAQL tool covers everything else.
  • Make every write two-step. Have write tools return a proposal first, using the Google Ads API's validate_only option to confirm the change is valid without applying it, then commit only after a human approves.
  • Put guardrails in code, not prompts. Allowlist the customer IDs each user can touch, cap budget and bid changes (for example, no more than a 20% change per action), and prefer pausing over removing. A prompt can be talked around; code can't.
  • Return small, clean results. Convert cost from micros to currency, cap row counts, and pre-aggregate where you can. Huge raw tables waste Claude's context and invite mistakes.
  • Log every change with who approved it, when, and the before-and-after values.
  • Encode your playbook. MCP servers can expose reusable prompts, so your weekly audit or search term review can be a one-click starting point for the whole team.
  • Host it remotely with OAuth if your team should use it from the Claude web app, or run it locally if it's just for you.

The official MCP SDKs (TypeScript and Python are the most widely used) and Google's Google Ads API client libraries (Python, Java, .NET, PHP, Ruby, and Perl) do most of the heavy lifting.

How to keep Claude safe in a live ad account

AI mistakes in an ad account cost real money, so treat this section as required reading.

  1. Read-only first. Run read-only for your first few weeks. You'll learn where Claude is strong and where it needs more context before it can change anything.
  2. Approve every write. Set write tools to require approval in Claude, and read the proposed change before approving. Never use "Allow always" on a tool that can change your account.
  3. Scope access to specific accounts. One connector per client (or per account) is the simplest way to prevent cross-account mistakes.
  4. Watch for prompt injection. Account data can contain text written by strangers. Search terms, for example, are typed by the public, and a search term could contain instructions aimed at an AI. Anthropic warns that connected tools can expose you to prompt injection. Claude is trained to resist it, but your approval step is the real safeguard.
  5. Disable write tools before using Research mode. Anthropic notes that Claude's Research mode can call connector tools without asking for approval each time.
  6. Mind the data. You're sending account data to an AI provider. Check your Claude plan's privacy settings, and if you're an agency, make sure your client agreements allow it. Never upload customer lists or other personal data into a chat.
  7. Ask for receipts. Have Claude show the query and the numbers behind every recommendation, and spot-check important ones in the Google Ads interface.
  8. Keep a change log. Record what changed, why, and when to re-check it, so you can tell whether each change helped.

A repeatable Claude workflow for Google Ads optimization

The difference between a gimmick and a real process is consistency. This is the loop we recommend:

  1. Load the context once. Create a Claude Project per account (or per client) with instructions covering goals, target CPA or ROAS, average lead value, service area, conversion definitions, brand rules, and anything off-limits. Every chat in the Project starts with that context.
  2. Pull the data. Ask Claude for the specific report you need, with a clear date range. Narrow is better than broad.
  3. Diagnose before prescribing. Ask "why" before "what should I change," and have Claude check sample sizes and conversion lag.
  4. Get a proposal. Ask for specific changes with expected impact, in a table you can review.
  5. Approve and apply. Approve changes one at a time (or in small, reviewed batches) through the write tool, or make them yourself in Google Ads.
  6. Log and re-measure. Ask Claude to produce the change log, then come back after enough data has accumulated (usually one to two weeks, longer for low-volume accounts) to judge the results.

35 Claude prompts for Google Ads optimization

Copy these into Claude with your MCP connector enabled. Replace anything in brackets. They work best inside a Project that already knows your goals.

Account audits

  1. "List every Google Ads account I can access through this connector, with each account's name, customer ID, currency, and spend over the last 30 days."
  2. "Run a full audit of [account] over the last 90 days: structure, bid strategies, budgets, conversion actions, match types, ad coverage, and assets. Rank the top 10 issues by estimated wasted spend or missed opportunity, and show the data behind each one."
  3. "Which campaigns spent money without producing conversions in the last 30 days? Include spend, clicks, and conversions, and tell me whether conversion lag could explain it."
  4. "Compare the last 30 days to the previous 30 for every campaign. Flag anything where cost per conversion moved more than 20%, and check the change history for edits that line up with the shift."

Search terms and negative keywords

  1. "Pull the search terms report for the last 30 days, sorted by cost. Group irrelevant terms into themes (jobs, free, DIY, competitors, wrong location, wrong service) and propose negative keywords for each theme, with match type and whether each belongs at the campaign or account level."
  2. "Find search terms with at least [X] clicks and zero conversions in the last 60 days. Separate the clearly irrelevant ones from the relevant-but-underperforming ones."
  3. "Find converting search terms that aren't keywords yet, and recommend which to add, with match type and the ad group each belongs in."
  4. "Check my negative keywords for conflicts. Are any of them blocking keywords I bid on, or search terms that have converted before?"
  5. "Draft a shared negative keyword list that blocks job seekers, students, and DIY searchers for a [business type], and tell me which campaigns should use it."

Keywords and match types

  1. "Show keywords that spent more than $[X] in the last 30 days with no conversions, with their match type, Quality Score components, and the search terms they triggered."
  2. "Which broad match keywords are matching irrelevant searches? For each, show the worst search terms and recommend negatives or a tighter match type."
  3. "Find duplicate or overlapping keywords across ad groups and campaigns that could be competing with each other."

Quality Score

  1. "List keywords with a Quality Score of 5 or lower and meaningful spend. For each, tell me which component is weakest (expected CTR, ad relevance, or landing page experience) and what to fix first."
  2. "For my lowest-relevance ad groups, compare the keywords to the ad copy and landing page, and suggest a tighter ad group structure."

Ads and assets

  1. "Review the responsive search ads in [campaign]. Flag ads with weak Ad Strength, repetitive headlines, or missing assets, then draft 10 new headlines and 4 descriptions within the 30- and 90-character limits."
  2. "Which ad groups have only one active ad, or no responsive search ad at all?"
  3. "Show which sitelinks, callouts, structured snippets, and call assets are attached at the account, campaign, and ad group levels, and point out the gaps."
  4. "Write ad copy variations for [service] that address these customer objections: [list]. Keep headlines to 30 characters and descriptions to 90."

Bidding and budgets

  1. "For each campaign, show budget, spend, Search impression share, and impression share lost to budget and to rank over the last 14 days. Tell me where more budget would likely help and where the real problem is rank."
  2. "Compare each campaign's target CPA or target ROAS with its actual results over the last 30 days. Where is the target unrealistic, and where is there room to scale?"
  3. "Project month-end spend for every campaign at the current pace, and flag anything likely to miss its monthly budget by more than 10% either way."
  4. "Which campaigns are limited by budget but have a cost per conversion below my target? Recommend how to move budget from weaker campaigns."
  5. "Look at performance by hour of day and day of week over the last 90 days. Is any pattern strong enough to act on, and would my current bid strategy even use schedule bid adjustments?"

Performance Max

  1. "Summarize performance by asset group across my Performance Max campaigns, and identify asset groups with weak conversion volume or low-rated assets."
  2. "Pull the search term insights for [Performance Max campaign] and flag categories that look irrelevant or overlap with my Search campaigns."
  3. "Check whether brand searches are likely flowing into Performance Max instead of my brand Search campaign, and recommend how to verify and fix it."

Conversion tracking

  1. "List all conversion actions with their category, primary or secondary status, counting method, and conversions over the last 30 days. Flag anything that looks miscounted, like page-view conversions set as primary or duplicate actions."
  2. "Check whether enhanced conversions for leads is turned on for this account, and explain what I'd gain from it."
  3. "Did conversion volume drop suddenly on any day in the last 60 days? Show which conversion action was responsible and whether any campaigns changed around that date."

Segments, locations, and landing pages

  1. "Break down performance by device and by location over the last 90 days. Where is cost per conversion meaningfully worse, and is the sample big enough to act on?"
  2. "Show landing pages by spend, conversion rate, and cost per conversion, and identify pages worth testing or replacing."
  3. "Are any campaigns getting clicks from outside my service area? Check whether location targeting is set to presence or to presence or interest."

Reporting

  1. "Write a one-page weekly summary for [client] in plain English: what changed, why, and what we're doing next. Lead with leads and cost per lead, not clicks."
  2. "Turn this month's results into five bullet points for a business owner who doesn't know Google Ads terminology."
  3. "Build a table of every change you recommended today, with the expected impact and a date to re-check it, so I can track results next week."

GAQL queries worth knowing

Claude writes Google Ads Query Language (GAQL) for you, but knowing a few core queries helps you check its work and ask for exactly what you want. One detail trips everyone up: the API reports cost in micros, so divide metrics.cost_micros by 1,000,000 to get dollars. Ask Claude to do the conversion in every answer.

Search terms by cost (the starting point for negative keywords):

SELECT
  search_term_view.search_term,
  search_term_view.status,
  campaign.name,
  ad_group.name,
  metrics.impressions,
  metrics.clicks,
  metrics.cost_micros,
  metrics.conversions
FROM search_term_view
WHERE segments.date DURING LAST_30_DAYS
  AND metrics.cost_micros > 0
ORDER BY metrics.cost_micros DESC
LIMIT 200

Keywords with Quality Score components:

SELECT
  campaign.name,
  ad_group.name,
  ad_group_criterion.keyword.text,
  ad_group_criterion.keyword.match_type,
  ad_group_criterion.quality_info.quality_score,
  ad_group_criterion.quality_info.search_predicted_ctr,
  ad_group_criterion.quality_info.creative_quality_score,
  ad_group_criterion.quality_info.post_click_quality_score,
  metrics.cost_micros,
  metrics.conversions
FROM keyword_view
WHERE segments.date DURING LAST_30_DAYS
  AND ad_group_criterion.status = 'ENABLED'
ORDER BY metrics.cost_micros DESC

Budget versus rank (where more budget helps, and where it won't):

SELECT
  campaign.name,
  campaign_budget.amount_micros,
  metrics.cost_micros,
  metrics.conversions,
  metrics.search_impression_share,
  metrics.search_budget_lost_impression_share,
  metrics.search_rank_lost_impression_share
FROM campaign
WHERE segments.date DURING LAST_14_DAYS
  AND campaign.status = 'ENABLED'
ORDER BY metrics.cost_micros DESC

If a campaign loses most of its impression share to budget, more budget may buy more conversions. If it loses most to rank, more budget won't help much; the fix is usually ad relevance, landing pages, or bids.

Example: a weekly wasted-spend review with Claude

Here's how a typical 15-minute Monday review goes for a lead-generation account, using a Project that already knows the target cost per lead and service area.

1. Ask for the problem, not a report. "Show me where we wasted money last week: search terms, keywords, and locations that spent without converting. Use the last 7 days, but check 60 days of history before calling anything a loser."

2. Claude pulls three reports (search terms, keywords, and geographic performance), converts cost to dollars, and separates terms with real spend and no conversions from terms with too little data to judge.

3. Claude groups the waste into themes, such as job seekers, DIY searches, and a neighboring city outside the service area, and proposes negative keywords for each, with match type and level (campaign or account).

4. You push back where you know better. "Keep the 'emergency' terms. Those convert by phone, and calls aren't fully tracked yet." Claude updates the list.

5. Claude proposes the changes through the write tool. You review the exact keywords and the campaigns they'll be added to, then approve.

6. Claude writes the change log, and you schedule a re-check for next Monday.

The value isn't just speed. It's that the review actually happens every week, with the history check and the change log, instead of being skipped when things get busy.

Automating recurring Google Ads analysis

Once your prompts are proven, you can run them on a schedule:

  • Markifact Workflows run reporting, alerts, and audits on a schedule or trigger, with ready-made templates.
  • Claude Code can run non-interactively, for example claude -p "Run the weekly search term review for [account]", from a scheduled task on a computer or server that has the MCP server configured.
  • The Claude API can connect to remote MCP servers directly through its MCP connector (currently in beta), which is how development teams build scheduled reporting pipelines.

One rule for automation: unattended runs should be read-only. Let the automated job find problems and draft recommendations, then have a human review and approve changes.

Limitations and mistakes to avoid

  • The API doesn't expose everything. A few reports in the Google Ads interface aren't available through the API. Auction Insights is the one most people miss.
  • Conversion lag. Conversions are credited to the date of the click, so the last few days always look worse than they'll end up. Don't judge yesterday's performance today.
  • Small samples. A keyword with 8 clicks and no conversions isn't proven to be bad. Ask Claude to flag sample sizes and use longer windows for low-volume accounts.
  • Smart Bidding learning periods. Frequent bid strategy, target, and budget changes reset learning. Make fewer, deliberate changes and give them time.
  • Confident errors. Claude can misread a metric or pick the wrong date range. Ask for the query and the numbers, and spot-check anything important.
  • Missing business context. Claude doesn't know your margins, close rates, or which leads became customers unless you tell it. Importing offline conversions makes every recommendation smarter.
  • Too much data at once. Large accounts can overwhelm a single request. Ask for narrower date ranges, top results by cost, or one campaign at a time.
  • API operation limits. Explorer access allows 2,880 operations per day on production accounts, which a busy multi-account workflow can hit.

Do it yourself or work with an agency?

Connecting Claude to Google Ads is now easy. Using it well still takes judgment: knowing which recommendations to ignore, when a drop is noise, and how changes interact with Smart Bidding and your sales process.

If you'd rather have experienced people running that loop, Sloss Digital pairs our own Google Ads MCP server with senior campaign management. See our Google Ads management and AI-driven marketing services, or request a free consultation.

Related glossary terms: Search terms report · Negative keywords · Quality Score · Keyword match types · Performance Max · Smart Bidding · Target CPA · Impression share · Enhanced conversions · Offline conversion import

Frequently asked questions

Can Claude connect directly to Google Ads?
Yes. Claude connects to Google Ads through an MCP (Model Context Protocol) server, which gives Claude tools to query your account through the Google Ads API. You can use a hosted service like Markifact, run Google's open-source Google Ads MCP server, or build your own. Claude then pulls live data, analyzes it, and, with your approval, makes changes.
What is a Google Ads MCP server?
A Google Ads MCP server is a small program that exposes Google Ads API actions, such as running reports, listing accounts, or adding negative keywords, as tools an AI assistant can call. Claude decides which tool to use based on your request, and the server handles authentication and the API call.
Is it safe to give Claude access to my Google Ads account?
It can be, with the right setup. Start read-only, limit the connector to specific accounts, and require approval for every change. Never set write tools to always allow. Be aware that account data such as search terms can contain text that tries to steer the AI, so review every proposed change before approving it.
Can Claude make changes to Google Ads campaigns?
Yes, if the MCP server you use offers write tools and you grant write access. Hosted options like Markifact route changes through a separate write tool that asks for your confirmation. Google's open-source server limits writes to assets and labels and proposes changes before committing them. Custom servers can support whatever you build.
Do I need a Google Ads developer token for an MCP server?
Not anymore. Google sunset developer tokens on September 9, 2026. Google Ads API access is now tied to the Google Cloud project behind your OAuth credentials, starting at Test access and upgrading to Explorer, Basic, or Standard. Hosted services like Markifact handle API access for you, so you only sign in with Google.
Which Claude plan do I need to use an MCP connector?
Remote MCP servers can be added as custom connectors on every Claude plan, though Free accounts are limited to one. On Team and Enterprise plans, an owner adds the connector for the organization first. Local MCP servers run through Claude Desktop or Claude Code on your own computer.
What is Markifact?
Markifact is an AI marketing automation platform with a hosted MCP server that connects Claude, ChatGPT, Gemini, and other AI tools to Google Ads and more than a dozen other ad platforms. It handles authentication, offers read-only and per-account permissions, and requires approval for write actions, which makes it one of the fastest ways to start.
Is there an official Google Ads MCP server from Google?
Google's Ads developer organization publishes an open-source Google Ads MCP server on GitHub (googleads/google-ads-mcp). It's written in Python, runs with pipx, and exposes tools for GAQL queries, account listing, assets, and labels, with a propose-and-confirm flow for changes. You set up your own Google Cloud project and credentials to use it.
Which Claude model is best for Google Ads analysis?
Use Claude Opus 5.5 for deep audits, multi-campaign diagnosis, and anything where reasoning quality matters most. Claude Sonnet 5.5 is faster and well suited to routine checks like weekly search term reviews and pacing reports. Whichever you choose, give Claude your goals, margins, and conversion definitions so its recommendations reflect your business.
Will Claude replace a Google Ads manager?
No. Claude makes analysis dramatically faster, but it doesn't know your margins, sales cycle, or which leads actually closed unless you tell it, and it can be confidently wrong. The best results come from an experienced manager who uses Claude to do the digging, then applies judgment and approves every change.

Free consultation

Ready to turn ad spend into qualified leads?

Get a free, no-obligation consultation. We'll review your goals and show you exactly how a performance-driven media plan can grow your pipeline.

No obligation · Alabama grit. Nationwide impact.