Managing Google Ads With Gemini

Published August 4, 2026 · 6 min read

Gemini is the odd case: Google already built it into Google Ads, so the question is not "can it access my account" but "whose interests is it optimizing for?" This covers what the native assistant does well, what Gemini CLI adds through MCP, and why keeping one independent advisor is worth the effort.

Three different things called "Gemini"

People conflate the three and end up disappointed. The chat app cannot see your campaigns. The in-product assistant can, but only within Google's own framing.

What the built-in assistant is good at

• Asset generation. Headlines, descriptions and images generated in context, already fitted to the format.
• Campaign construction. Conversational setup that pulls keyword and asset suggestions from your landing page.
• In-product answers. "Why is this ad disapproved" answered where the ad lives.

It removes friction from work you were already doing inside the interface. That is genuinely useful.

The structural conflict

Google's revenue is your ad spend. The recommendations surfaced in-product skew consistently in one direction: broader match types, more automation, Performance Max, higher budgets. Not because anyone is being dishonest, but because "increase budget on a converting campaign" is both a reasonable suggestion and a revenue event.

Optimization score makes this concrete. Accepting every recommendation gets you to 100% and frequently costs more per conversion than a deliberately imperfect account. The questions the in-product assistant will never volunteer:

• Should this Performance Max campaign exist at all, or is it harvesting brand traffic you would get for free?
• Is broad match with Smart Bidding actually cheaper here than phrase match was?
• Should 30% of this budget move to Meta this month?

None of those are answerable by a tool inside the platform being questioned.

Gemini CLI plus an MCP server

Gemini CLI supports MCP servers, which means you can point Gemini at an independent tool layer instead of Google's own framing. The setup mirrors any other client — an entry in the CLI's MCP configuration pointing at the remote server:

https://klienta.co/mcp

Configuration formats differ between clients and change between versions; the setup page carries current instructions per client. Once connected, the model queries your live account through the same tool surface Claude and Cursor use — 83 Google Ads tools, 19 Meta Ads tools, guardrails and undo on every write.

Questions worth asking an independent tool

• "Is my Performance Max campaign cannibalizing brand search? Show impression share and search terms overlap."
• "Which recommendations did Google apply automatically in the last 90 days, and what happened to CPA after each?"
• "If I cut the worst-performing 20% of spend, what would I lose in conversions?"
• "Compare Google and Meta for last month — spend, conversions, CPA — and tell me where the next $1,000 goes."

The last one is out of scope for any in-platform assistant by construction. Cross-channel comparison requires a tool that is not owned by one of the channels.

If you prefer another client

The same server works with Claude Desktop, Claude Code and Cursor — see managing Google Ads with Claude, or the ChatGPT version for what changes when your client has no connector support.