Practical guide
Claude and ChatGPT for Google Ads and Meta Ads: a setup guide
Claude and ChatGPT can read your Google Ads and Meta Ads data once you give them a connection. There are three realistic ways to do it, and they differ in setup effort, in whether the assistant can change live campaigns, and in how much arithmetic it has to get right on its own. This guide compares them, walks through the setup and lists the prompts and checks that keep the answers trustworthy.
By Adlane · Updated
Three ways to give an assistant your ad data
Most setups use the Model Context Protocol (MCP), an open standard that lets an assistant call tools on a server you connect. The server holds the connection to Google or Meta; the assistant asks it for accounts, campaigns and reports. The alternative is to export a report and upload the file.
Pick the option by what you want the assistant to do. Analysis needs read access only. Editing campaigns from chat needs write access, and with it a clear idea of who reviews a change before it reaches a live account.
| Option | What you need | Can it change live ads? | Best for |
|---|---|---|---|
| Provider’s official connector (Meta’s ads AI connector, Google’s open-source Google Ads MCP server) | A supported assistant; for Google, a Google Cloud project with Google Ads API access and the skill to run a Python server | Meta’s connector is described as supporting campaign management; Google’s current release is read-only | Teams that want the provider’s own tooling and can manage the setup |
| Hosted MCP server from a third party (Adlane is one) | The server URL and a sign-in with the Google or Meta login that already sees the account | Depends on the service; Adlane’s endpoints are read-only | Analysis in a couple of minutes, with nothing to install |
| Exported report (CSV or spreadsheet) uploaded to the chat | Access to Ads Manager or Google Ads reports | No | A one-off question about a fixed date range |
Connect Claude to Meta Ads
In Claude on the web or desktop, open Settings, choose Connectors and add a custom connector. Paste the remote MCP server URL, save, then click Connect and finish the Meta sign-in in the browser. Custom connectors depend on your Claude plan; an organisation owner may need to add the connector for the team.
For Meta’s official connector, use the URL Meta gives you during its setup. For Adlane’s read-only Meta Ads MCP, the URL is https://mcp.adlane.app/metaads. During consent, choose only the business portfolios and ad accounts you intend to analyse. A Facebook Page role alone does not grant access to advertising reports; the login needs a role on the ad account itself.
In Claude Code, add the same URL from the terminal with claude mcp add --transport http adlane-meta-ads https://mcp.adlane.app/metaads, then run /mcp inside Claude Code to sign in.
Connect Claude to Google Ads
Google’s Ads API team publishes an official Google Ads MCP server on GitHub. It runs as a local Python process or on your own Cloud Run service and needs a Google Cloud project with Google Ads API access plus an OAuth client or service account. Its current release reads data only: it lists accessible customers, runs GAQL searches and describes resource fields. It suits developers who want to run and audit the code themselves.
A hosted server skips that setup. With Adlane, add https://mcp.adlane.app/gads as a custom connector and sign in with the Google login that already sees the account. If that login reaches clients through a manager account, ask the assistant to list customer clients first; campaign data lives in the client account, not in the manager.
Use ChatGPT, Cursor or VS Code instead
The same server URL works in any client that supports remote MCP servers with OAuth. In ChatGPT, custom connectors are only available on plans and workspace settings that allow them, and an administrator may need to enable them. In Cursor, add the server to .cursor/mcp.json; in VS Code, add it to .vscode/mcp.json and start it from the MCP view. Each client opens a browser for the provider sign-in the first time it uses the server.
Whichever client you use, the report data you request is returned to that client and may be processed by its AI provider. Check the client’s data and training settings before asking about sensitive accounts.
Prompts that produce answers you can check
Vague questions get confident, unverifiable answers. Give the assistant a scope, a period and a rule about changes, and ask it to show which report each number came from.
| Goal | Prompt to adapt | Why it works |
|---|---|---|
| Pick the right account | List the ad accounts I can access, with currency and timezone. Ask me which one to use. | Stops the assistant guessing between similar accounts and fixes the units |
| Explain a change | Compare the last 28 complete days with the previous 28 days for campaign spend and conversions. Cite the report for each figure. | Complete days and a cited source make the comparison reproducible |
| Find wasted search spend | List search terms with cost but no conversions in the last 60 days. Group them by intent and suggest which need a closer look. Do not change anything. | Separates review candidates from actions |
| Check Meta frequency | Show reach and frequency for this ad set over the whole period, not summed by day. | Reach is unique per period; daily rows cannot be added |
| Reconcile platforms | Report Google Ads and Meta attributed conversions side by side with their attribution settings. Do not add them together. | Each platform attributes the same order by its own rules |
Five mistakes assistants make with ad data
Even a capable model can misread ad reports when the tool hands it raw API output. Check for these before you act on an answer.
| Mistake | What to check |
|---|---|
| Cost figures a million times too large | The Google Ads API reports money in micros; divide by 1,000,000 for account-currency units |
| Reach added across days or ad sets | Ask for reach over the whole period and scope; daily or per-ad-set reach overlaps |
| Google and Meta conversions summed as unique sales | Keep them separate and compare with your own order or CRM data |
| Recent days judged as poor performance | Conversions can arrive days after the click; treat the latest days as incomplete |
| Mixed currencies or timezones | Read both from the account before comparing accounts or periods |
Read-only first, then decide how changes happen
Connecting an assistant with write access means a prompt can change a budget, pause a campaign or publish an ad. That can be useful, but a misunderstood instruction then reaches a live account directly. A safer pattern is to keep analysis read-only and make each change a separate, explicit step that a person approves after seeing the current setting and the proposed one.
Adlane follows that pattern: its MCP endpoints only read data, and inside the Adlane workspace a change is drafted as a proposal you review and approve. New campaigns are created paused. If you prefer to manage campaigns from chat, Meta’s official connector is the option built for that; set up permissions deliberately and limit it to the business portfolio you intend.
Questions, answered
Can Claude manage my Meta Ads?
Claude can read and, with a connector that has write access, change Meta campaigns. Meta’s own ads AI connector is described as supporting campaign management. Read-only connectors such as Adlane’s Meta Ads MCP let Claude analyse accounts without any risk of editing a live ad.
Is there an official Google Ads MCP server?
Yes. Google’s Ads API team publishes an open-source Google Ads MCP server. You run it yourself, locally or on Cloud Run, with a Google Cloud project that has Google Ads API access. The current release is read-only.
Do I need a developer token or API keys?
Not for a hosted server: you sign in with the Google or Meta login that already sees the account, and the service holds the API application credentials. Running Google’s official server yourself needs a Google Cloud project with API access and an OAuth client or service account.
Does ChatGPT work with Google Ads and Meta Ads?
Yes, through custom MCP connectors, where your ChatGPT plan and workspace settings allow them. You add the server URL and complete the provider sign-in in the browser.
Is it safe to connect my ad account to an AI assistant?
Grant the narrowest access that answers your question: read-only for analysis, and only the ad accounts you need. Remember that report data you request is sent to the assistant’s provider. Keep changes to live campaigns as a separate step that a person approves.
Sources
Continue your review
- Google Ads audit: turn campaign evidence into a review plan
- Facebook Ads reporting: connect the numbers to a decision
- Google Ads cost_micros: convert micros to currency without errors
- Meta reach is not additive: why you cannot sum reach across rows
- Meta Ads account access permissions: fix an empty connection
- Cross-channel reporting: shared evidence without false totals