Natural Language Google Ads: How It Works

Natural language Google Ads management means typing a sentence — "pause every keyword in the brand campaign with zero conversions in the last 30 days" — and having that sentence translated into the right Google Ads API calls and executed against your live account. It's not a wrapper around the Ads UI. It's an AI model with structured tool access reading your data and writing back through the API.

This works because of a specific architecture: an AI model (Claude, GPT, etc.) plus an MCP server that exposes Google Ads API operations as named tools. NotFair is the production MCP server most people use. The AI maps your sentence to one or more tool calls, the tools hit the API, results come back, and the AI summarizes them in plain English.

The prompt-to-API pipeline, end to end

• Step 1: parse intent — The model classifies the sentence as a read or write, identifies the entities (campaigns, keywords, search terms), and pulls out the filter expressions ("zero conversions", "last 30 days").
• Step 2: pick the tool — Reads almost always go through runScript with a GAQL query. Writes go through specific tools like pauseKeyword, bulkPauseKeywords, addNegativeKeyword, or updateCampaignBudget.
• Step 3: build the structured call — The model fills in the tool arguments: customer ID, resource names, filter clauses. For runScript, it writes a JS sandbox snippet that calls ads.gaql or ads.gaqlParallel.
• Step 4: server validates — NotFair checks scopes, validates the GAQL syntax, and for writes, gates the call behind an approval prompt that you confirm.
• Step 5: API execution — The Google Ads API receives the call under NotFair's developer token, authenticated as your Google identity via OAuth.
• Step 6: summarize — The model takes the raw API response (JSON, often verbose) and renders it as a readable answer with the numbers that matter.

A real prompt mapped to real tool calls

Why this is more reliable than "AI clicking buttons"

Browser-automation agents that drive the Google Ads UI exist, but they're slow and brittle. They depend on DOM selectors that Google changes, struggle with auth, and can't handle bulk operations cleanly. The MCP path skips the UI entirely. The AI talks to the same REST API that an enterprise PPC platform would talk to. No clicking, no waiting on page loads, no breakage when Google ships a redesign.

How ambiguity is handled (this is the important part)

Natural language is ambiguous. "Pause my underperforming campaigns" is not a precise instruction. The AI's job, when faced with ambiguity, is to ask a clarification question rather than guess and execute. NotFair's tools are structured so that destructive writes require explicit parameters — there is no API call for "pause underperforming things," so the model has to resolve the ambiguity into specific resource names before any write can happen.

Failure modes you should know about

• Hallucinated metrics — Models occasionally invent numbers when context is thin. The fix: every claim should be backed by a tool call. If Claude says "CPA is $40" without a runScript call before it, treat that number as suspect.
• Wrong scope — "Pause this keyword" can refer to one of many keywords with similar text. Tools require resource IDs; ambiguity should produce a clarification, not a guess.
• Date interpretation — "Last week" means different things (last calendar week vs last 7 days). The AI should state which interpretation it picked.
• Stale changes — If you make a change in the Google Ads UI mid-conversation, the AI is working from cached numbers. Re-running a query is cheap; do it after manual UI edits.

What a good natural-language ads system looks like

• Grounded — Every numerical claim ties to a specific tool call. No vague "your CTR is low."
• Reversible — Writes flow through an approval gate. Many operations also expose an undoChange tool for emergency rollback.
• Composable — One sentence can fan out into 20 parallel queries (gaqlParallel) when correlation across surfaces is needed for an audit.
• Auditable — The conversation log is the audit log. You can scroll back to see exactly what was asked, what tools fired, and what changed.

Setting up a natural language interface to your account

The shortest path: visit notfair.co/connect, complete OAuth with the Google identity that has Ads access, pick the customer IDs you want exposed, and install the connector in Claude Desktop, Claude Web, or ChatGPT. Cursor and Windsurf both support MCP via stdio config and work the same way. From OAuth to first natural-language audit is about five minutes.

FAQ