"Testing" is not changing things now and then. It is a cycle with rules —the same one any experimental discipline uses— repeated with no end. This lesson formalises it for Search ads and gives you the calendar to apply it.
The cycle (it never ends — that is the point)
1 · HYPOTHESIS a new angle, grounded in what is already known about the group
2 · EXPOSURE the challenger competes in real auctions, with real data
3 · MEASUREMENT piece by piece: impressions, label, combinations
4 · VERDICT the good stays; the bad is pulled — with human judgement
5 · MEMORY the result is archived: the account learns forever
… and round again, each time with more memory
The difference between "changing ads now and then" and continuous optimisation lies in steps 3 and 5: measuring piece by piece and archiving the result. With no memory, the cycle accumulates nothing: it just goes round.
The three rules that make it work
- Every hypothesis is an angle, not a paraphrase. If the champion sells trust, the challenger attacks on price, on urgency or on novelty. Only then does the test teach you something new about the ad group's audience.
- The verdict comes from the data per piece, not from a feeling. A headline with thousands of impressions and a «Low» label is a verdict; "I prefer this one" is not.
- Nothing is forgotten. Winners and losers are archived with their result. Round 20 of the test builds on what the previous 19 taught: that is the compounding advantage the eternal ad never accumulates.
How it works in an RSA, step by step
| Step | In practice |
|---|---|
| Hypothesis | Pick the piece to replace (the «Low» one with the most impressions) and write one candidate with an angle the group has not tried yet (archive sheet, previous lesson) |
| Exposure | Replace the piece in the existing RSA (do not create a new ad: you keep the learning on everything else). Wait until it accumulates enough impressions: thousands, usually 3-4 weeks |
| Measurement | Asset report: impressions and label for the candidate; combinations it appears in; and the whole chain for the group (CTR, conversion rate, CPA) before and after |
| Verdict | «Best/Good» → it stays as part of the champion. «Low» with volume → it is pulled. No volume → wait one more round |
| Memory | A row in the archive sheet: piece, angle, dates, impressions, label, verdict |
One or two pieces per ad group per round; never the whole ad (Module 4 of the Basic level). To compare two complete approaches (two RSAs with opposing angles), a Google Ads custom experiment is the clean tool.
The calendar for a 20-group account
- Every two weeks, 90 minutes: capture the asset report of all 20 ad groups into the sheet (wall 1), verdicts on the candidates that have volume, archive.
- Same day: hypotheses and new candidates in the groups that need them (pieces labelled «Low» with volume, groups with fewer than 2 RSAs, «Poor» ad strength). Maximum 5-6 groups per round, so it fits the time.
- Monthly, 30 minutes: a cross-account read of the sheet — which angles win across the whole account? Which angles never work in this sector? That feeds next month's hypotheses.
- Quarterly: review the ad → QS chain: is expected CTR rising in the groups you have worked on? Is their CPC falling?
Three hours a month for 20 ad groups. It is manageable… the first month. Wall 5 shows up in the third.
Compound interest
Continuous optimisation pays like compound interest: every round starts from the level of the previous one. A modest, sustained improvement in CTR —month after month, round after round— ends up moving the whole ad group's Quality Score, and with it the price of every click. The eternal ad, meanwhile, accumulates the opposite compound interest: decay (the five forces from lesson 2). Two identical accounts in January are two very different accounts in December, and the difference is not visible in any single month: it is visible in the total.
The mistakes that break the cycle
- Skipping measurement: changing pieces "because it was time" without reading the asset report.
- Skipping memory: with no archive, round 20 is the same as round 1.
- Two hypotheses at once in the same ad group: if it improves, which one did it?
- Judging on CTR alone: the price angle raises clicks and may lower sales (lesson 1). The whole chain.
- Stopping "because it's fine now": that is the eternal ad coming back through the back door.
💡 Ninja trick: the RSA Optimizer runs all five steps every night and leaves the verdict in your hands: AI hypotheses grounded in the group's archive (lesson 6), exposure when you tick the box to replace, measurement per piece and per combination, permanent memory. And it respects the rule that costs most by hand: anti-repetition windows, so a freshly refreshed group gets no new proposals until it has accumulated data.
What you should remember
- Five steps: hypothesis → exposure → measurement → verdict → memory, in a loop.
- Three rules: angle, not paraphrase · data per piece, not a feeling · nothing is forgotten.
- One or two pieces per group per round; 3-4 weeks of exposure; never the whole ad.
- It pays like compound interest; the eternal ad accumulates the opposite.