Google Ads experiments review
Reads each A/B test with its type, dates, status and the traffic split between control and treatment, compares both arms over the test period and says whether the result is readable.
| Item | Value |
|---|---|
| Skill name | google-ads-experiments-review |
| Required MCP server | Pilot MCP, https://app.pilot.braike.com/mcp |
| Writes | None: the skill only reads |
| Download | google-ads-experiments-review.zip |
MCP tools used
| Tool | Platform |
|---|---|
google_list_accounts | Google Ads |
google_list_experiments | Google Ads |
google_list_experiment_arms | Google Ads |
google_list_campaigns | Google Ads |
google_get_campaign_performance | Google Ads |
Optional MCP servers
A file storage server to keep the result, a team messaging server to share it, a note-taking server to archive it. None is needed: the skill works with Pilot MCP alone.
Example requests
- "How is our Google Ads experiment going?"
- "Should we promote the bidding test?"
What the skill returns
A table of experiments, each arm's results and the gap, a verdict (readable winner, needs time, stop) and ideas for the next tests.
Install
Download google-ads-experiments-review.zip, then follow the steps for your client: install a skill. With the Pilot CLI, one command is enough:
npx @braike/pilot skills install google-ads-experiments-reviewThe SKILL.md file
See the file as your assistant reads it
---
name: google-ads-experiments-review
description: "Reviews Google Ads experiments through the Pilot MCP server: every A/B test with its type, dates and status, its control and treatment arms and their traffic split, and the results of each arm's campaign over the test period, then says whether a test has a readable winner, needs more time or should be stopped, for a person to decide. Use when the user asks how an experiment is going, whether to promote a test, or what tests ran. Triggers: Google Ads experiments, A/B test results, campaign experiment, should we promote the test. Déclencheurs : expériences Google Ads, résultats du test A/B, test de campagne, faut-il appliquer le test."
compatibility: "Requires the Pilot MCP server (https://app.pilot.braike.com/mcp) connected in the client and authorized on a workspace with a Google Ads account attached."
metadata:
pilot-mcp-server: "https://app.pilot.braike.com/mcp"
pilot-mcp-tools: "google_list_accounts google_list_experiments google_list_experiment_arms google_list_campaigns google_get_campaign_performance"
pilot-mcp-read-only: "true"
---
# Google Ads experiments review
Tell the user what each experiment tests, how its arms compare, and whether the result can be trusted yet. Answer in the user's language.
## Requirements
- **Required MCP server:** Pilot MCP (`https://app.pilot.braike.com/mcp`), authorized on the workspace that holds the account.
- **Optional MCP servers:** a document server to keep a test log. None is needed.
- **Read-only:** never promote, end or change an experiment. The decision is for a person.
## Tools
- `google_list_accounts`: the Google Ads accounts attached to the workspace.
- `google_list_experiments`: experiments with name, type, status and dates.
- `google_list_experiment_arms`: the control and treatment arms of one experiment, with their campaigns and traffic split.
- `google_list_campaigns`: campaign names and settings, to see what differs between arms.
- `google_get_campaign_performance`: daily cost, clicks, conversions and value of each arm's campaign.
## Steps
1. Call `google_list_accounts` and pick the account. Call `google_list_experiments`.
2. For each running or recently ended experiment, call `google_list_experiment_arms` with its `experiment_resource_name`: note the control campaign, the treatment campaign and the split.
3. Call `google_list_campaigns` to see what the treatment changes (bidding, budget, settings).
4. Call `google_get_campaign_performance` with `campaign_id` for each arm, from the experiment start date to its end or to yesterday.
5. Compare arms on the test's goal: conversions, CPA, ROAS or conversion value. Scale each arm by its traffic share before comparing volumes.
6. Judge readability: with fewer than about 30 conversions per arm, or less than two weeks of data, the test is not readable yet. A gap smaller than the day-to-day swing of the control is not a result.
## Output
1. **Experiments**: a table with name, what it tests, status, dates and split.
2. **Results**: per experiment, each arm's figures and the gap.
3. **Verdict**: winner readable, needs more time (with an estimate), or stop, and why.
4. **Next tests**: ideas the results suggest, for a person to decide.
## Rules
- Never call a winner on too few conversions; say how many are missing.
- The first days after a bidding change are learning: leave them out or say they are included.
- Name the dates and the currency.
## Examples
- "How is our Google Ads experiment going?"
- "Le test d'enchères a-t-il un gagnant ? Faut-il l'appliquer ?"
- "Quels tests ont tourné ce trimestre, et avec quel résultat ?"Google Ads audience review
Reads remarketing and customer lists, their size and state, the audiences targeted, observed or excluded in each ad group, and the results of each audience.
Keyword research
Draws ideas from Keyword Planner from seed words or a page, with volumes, competition and bids, removes what the account already has, adds the search terms that convert, and groups everything into themes.