Experiments for Google Ads
Give your AI tool the campaign report by day, and your leads or sales by region for a geo test. It says how many days a test needs, reads the result with a verdict on how far to trust it, and works out the profit at your margin.
npx skills add parallel-ai-inc/google-ads-skills -gInstalls all 13 skills in Claude Code, Codex, Cursor and other tools. Steps for each tool
Test readout
The test regions and dates were found from the data, not given by the person. Confirm them before acting on this.
The answer
Verdict: Strong. In the test group (Evanston; Skokie; Oak Park), with ads switched off, from 2026‑10‑22 to 2026‑11‑04: Leads came in at 310.0 against 438.1 predicted from the control group: a difference of −128.1 (−29.2%), range −189.4 to −66.7 at 95% confidence. That is distinguishable from zero.
Why that verdict: a gap this large appeared in 0% of 69 placebo windows in the pre‑period. Fit before the test: R squared 0.38.
- The ads caused about 128.1 leads over the 14 days (9.1 a day), range 66.7 to 189.4.
- Spend removed: 1,934.34 over the 14 days.
- Cost per incremental lead: 15.11 (range 10.21 to 28.98).
- Google Ads attributed 13.89 conversions a day to the test group before the test. The ads caused 9.1 leads a day: 66% of the attributed figure (range 34% to 97%). Attributed is not incremental.
- Incremental revenue per unit of spend: 15.83 (range 7.96 to 23.70).
- Profit: at this margin, each unit of spend returns 5.33 of profit after paying for itself (range 2.18 to 8.48); 10,315.90 in total. Break even is 2.50 of revenue per unit of spend. The ads make money.
What to give it
To design a test or work out return on investment, the campaign report by day for 8 to 13 weeks, from Campaigns with the Day segment or the Report editor for a long range.
To read an experiment, the two arms by day or Google’s experiment scorecard. To read a geo test, the business’s leads or sales by region and day from its booking system or CRM, and the Google Ads cost by region and day from the Report editor. If you do not remember the test regions or dates, those two files are enough to find them.
The skill also asks for the decision the test will drive, the lift that would change it, the gross margin, whether conversion value is revenue or profit, and what a lead is worth.
What you get back
test-plan.xlsx and test-plan.mdThe design. The days the test needs, the smallest lift it can detect, the dates, what changes and what stays fixed.readout.xlsx and readout.mdThe readout. The lift with its range, a verdict on how far to trust it, the checks on the result, and profit at your margin.roi.xlsx and roi.mdReturn on investment, and whether the next dollar pays.
In the chat, your AI tool answers what you asked first, then the verdict and what it rests on, and what to confirm before acting on it.
It says when a test cannot answer the question
Before a test runs, the design works out the smallest lift it could detect in the time you have. When the answer is too small to see, it says so rather than sending you into weeks of a test that cannot settle the decision.
The answer
Too little volume. A lift of 15% in conversions needs 359 measured days, longer than the 56 day limit. In 56 days the smallest lift this test could detect is 42%.
Design: a Google Ads custom experiment (a split of one campaign). Run 2026‑11‑16 to 2027‑01‑17, reading from 2026‑11‑23.
What lift it can detect
At 95% confidence and 80% power:
| Measured days | Smallest lift it can detect |
|---|---|
| 14 | 102.9% |
| 21 | 78.2% |
| 28 | 64.9% |
| 42 | 50.4% |
| 56 | 42.4% |
How it decides
Scripts compute every number, and your AI tool states no figure they did not print. Each rule it applies is labelled Official, from Google’s documentation, or Practitioner rule, from Parallel AI’s operating knowledge.
- A result is Strong only when its range excludes zero, a gap that large appeared in 5% or fewer of the placebo windows before the test, and the pre‑period is at least three times the test.Practitioner rule
- Keep at least three times the test’s length in data from before it.Official
- An experiment’s 50% split sets which auctions each arm can enter, not equal impressions or spend.Official
- Conversions in Google Ads are what it attributed. Only a test shows what the ads caused.Practitioner rule
What it does not cover
- Setting up the test. The plan says what to set up in Google Ads and what to keep fixed, and you set it up.
- Which campaign caused a geo result. A geo test reads every campaign in the test regions together.
- Anything that hit only the test regions during the test, such as a local event or a competitor. It shows up as an effect.
- Real downloads of every layout. The experiment and geographic report layouts have not yet been checked against real downloads.
When you want it on live accounts
The skill plans and reads tests from exports, and setting them up is up to you. Parallel AI plans and reads experiments on your live Google Ads accounts.
The page for this job in Parallel AI is How Parallel AI works for Google Ads.
Install
In Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, OpenCode and other tools that read Agent Skills, the open skills CLI installs every skill in the pack.
npx skills add parallel-ai-inc/google-ads-skills -gIn Claude Code, add the pack as a plugin from a session.
/plugin marketplace add parallel-ai-inc/google-ads-skills/plugin install parallel-ai-for-google-ads@parallel-ai
To install this skill alone, name it.
npx skills add parallel-ai-inc/google-ads-skills --skill experiments-for-google-ads -gAdd -a with a tool’s id, such as codex or cursor, to install for that tool alone. Without -g, the skills install into the current project.
Questions and answers
Does the skill cost anything?
No. The skills are free to use for any Google Ads work, for yourself or for clients. They may not be republished as your own skills. Parallel AI is a separate product, with a 5‑day free trial.
Does it change my Google Ads account?
No. It never connects to an account on anyone’s behalf and never changes one. It reads the reports you give it and writes files. Where a job ends in a change, it writes a Google Ads Editor file, and you check it there before posting.
Where does my data go?
The skill’s scripts make no network calls, and nothing goes to Parallel AI. Your AI tool sees the files you give it, as it would any file, under that tool’s own terms. If the export is a client’s, check your agreement with that client before you give it to an AI tool.
What if I do not remember when a test ran?
For a geo test, give it your leads by region and day and the Google Ads cost by region and day. It finds the regions and dates where ads were off, and asks you to confirm them before acting on the result.
Which AI tools can run it?
Tools that read Agent Skills and can run Python, such as Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf and OpenCode. The scripts use Python 3’s standard library only.
Is Northside Plumbing a real business?
No. Northside Plumbing is the pack’s made up sample account, with problems planted for each skill to find. Run the skill on the sample files and you get the same figures as this page.
Google Ads is a trademark of Google LLC. Parallel AI is not affiliated with or endorsed by Google.
Plan and read tests on every account
Parallel AI reads your live Google Ads accounts, plans experiments that can settle the decision and reads their results, across every client at once.
Starter $20, Pro $60 and Max $200 a month
You connect Google Ads first, then add a card to start the trial.