Tutorial · A custom experiment from start to finish

In short: a custom experiment is a copy of your campaign with a single change, split 50/50, for 4-6 weeks. A search-based split works for bids and match types; a cookie-based split for landing pages. The result is read with three things at once: significance, effect size and the whole chain — and negative results get documented too.

Several lessons have told you to "run an experiment". This tutorial actually runs one, screen by screen.

Step 1 · Where do you create experiments?

Campaigns → Experiments (side menu) → + → Custom experiment → campaign type (Search, Display, Shopping, Demand Gen; PMax has its own experiments) → Continue.

Step 2 · Which campaign do you use as the base?

Name ("tCPA test 40 vs 35 · Implants") and description (the hypothesis: "cutting the target by 12% will hold volume"). Base campaign: the one you want to test. Google creates a copy (the "trial" version).

Step 3 · How many changes do you test at once?

The trial campaign opens: edit only what you are testing:

Everything else, untouched. Budget: the split sets it (do not touch it).

Step 4 · How do you split the experiment's traffic?

• Traffic split: 50% (recommended: balanced data).
• Split type: search-based (each search is assigned at random to one version; more data, faster comparison) or cookie-based (each user always sees the same version; better for landing page or experience changes). For bidding and match types, search-based; for landing pages, cookie-based.
• Dates: start tomorrow; end in 4-6 weeks (you can extend it).
• Option to sync changes from the base campaign to the trial (if you touch the base during the experiment, it is replicated): switch it on so the only difference stays the one you introduced.

Create. The trial version starts on the date you set.

Step 5 · What should you do while the experiment runs?

• Do not touch either the base or the trial (beyond what syncs).
• Experiment panel: metrics for both versions, difference in % and confidence interval; the star/indicator marks when the difference is statistically significant (95%).
• Check after a week that both versions are getting traffic (real split ≈ 50%).

Step 6 · How do you read the results?

1. Primary metric (the one you defined): is the difference significant?
2. Size: by how much? Does it justify the cost of applying it (learning period reset)?
3. The whole chain: business conversions (qualified leads, value), not just clicks or tag conversions.
4. Side effects: IS, CPC, spend.

No significance = no result (extend it, or accept that the effect is smaller than the noise).

Step 7 · How do you apply a winning experiment?

• Apply: two options — update the original campaign with the changes (it keeps its history; the trial closes) or convert the trial into a new campaign (and pause the original). Almost always the first.
• End without applying: everything reverts to the base; document the result (a negative one counts too).
• Apply outside peak periods; watch it for two weeks.

Which mistakes invalidate an experiment?

• Touching the base without sync switched on.
• A 90/10 split.
• Two changes in the trial.
• Judging it at 7 days.
• An experiment on a campaign with < 30 conversions/month: it never reaches significance.
• Primary metric = clicks when the objective is conversions.

💡 Ninja trick: leave the SBNS judge in TEST mode over both campaigns in the experiment: on top of Google's official verdict, you will have the daily series of how each version is meeting the target — and you will know the day one started to pull away from the other.

What to remember

• Experiments → custom → copy of the campaign → one single change.
• 50/50; search-based split (bidding, match types) or cookie-based (landing pages); 4-6 weeks; sync switched on.
• Result = significance + size + the whole chain.
• Apply by updating the original; document the negatives too.