Key takeaways
- An experiment needs evidence, a hypothesis, a change, a measure, a window, and a decision. Missing the window is a wish.
- Prioritise with explicit uncertainty — not a fake ICE score that pretends you know the outcome.
- If measurement cannot name the event, fix the architecture before you backlog another test.
The backlog has forty-three ideas. Twelve are live. Nobody wrote down what success means. Three “wins” were seasonal noise.
A growth experiment backlog is not a wish list in a whiteboard app. It is a ledger where every row has evidence, a hypothesis, a change, a measure, a window, and a decision. Prioritise without fake certainty. If the window is missing, it is not an experiment. It is a mood.
How to sell more from your website named the leak when the offer stalls. What to do in the 90 days after you launch named the post-launch rhythm: measure, care, one improvement cycle. This page is how you choose that improvement without fooling yourself.
Ideas are cheap; decisions are the product
Teams confuse activity with learning. They change button colour, headline, and form on the same week, during a holiday, while sales ran a promotion. Then they pick the chart that went up and call it strategy.
An experiment exists to change a decision at the end of a defined window. If the row cannot end with ship, revert, or iterate with written reason, it should not enter the backlog.
| Field | Question it answers |
|---|---|
| Evidence | What do we observe today, with a number or trace? |
| Hypothesis | If we change X, Y will move because Z |
| Change | One scoped change only |
| Measure | Primary metric and guardrail metric |
| Window | Start, end, and minimum sample rule |
| Decision | Ship, revert, or iterate — recorded |
The Scientific Method is not branding. It is how you stop debating taste.
Log every row in the same shape
Use one template. No free-text cards that omit the window.
Evidence: “Form completion on `/quote` is 2.1% on mobile; call clicks are high; inbox shows wrong-service inquiries.”
Hypothesis: “If we replace the generic headline with the ad’s scoped offer, completion will rise because comprehension will rise.”
Change: headline and proof block only — no form change in the same window.
Measure: primary `form_complete` rate on mobile; guardrail qualified-lead rate in CRM.
Window: fourteen days or 300 sessions on the URL, whichever comes last — no peeking daily except for breakage.
Decision: ship if primary up and guardrail flat; revert if guardrail down; iterate if inconclusive with noted confounders.
Monday: open your experiment tool or spreadsheet. Delete any row without a window and a measure. What remains is the backlog.
Prioritise without inventing precision
Prioritisation needs ranking, not fake math. A common failure is an ICE or RICE score where impact is guessed to two decimal places.
Use explicit bands instead:
- Evidence strength: weak / medium / strong
- Effort: hours / days / weeks
- Risk: reversible / hard to revert
- Leverage: which funnel factor moves — arrivals, comprehension, belief, completion, fit
Run high-evidence, reversible, short-window tests first. A headline test beats a checkout rebuild when comprehension is the suspected leak. A tracking fix beats any UI test when nobody trusts the numbers.
Two experiments may not change the same surface in the same window. Queue them. Parallel tests on one URL are how teams learn nothing.
Guard against bad measurement
An experiment backlog that cannot hide bad measurement refuses rows when:
- the primary event is undefined or fires twice;
- baselines come from a week that included an outage or promo;
- the window is shorter than the sales cycle for the offer;
- nobody owns reading the result.
If those conditions fail, the work is conversion tracking architecture — not another backlog item. Fix the event ontology before you optimise headlines.
Document confounders when you close a row: seasonality, budget change, sales follow-up change, site deploy on day three. Honest inconclusive is better than a false win.
Share the decision in one paragraph — what moved, what did not, what you will try next. A backlog that ends in “we think it worked” without numbers is a diary. A backlog that ends with ship, revert, or iterate with evidence is a growth programme.
Review the backlog monthly
Once a month, close stale rows. Anything older than two windows without a decision is either blocked on measurement or was never scoped. Blocked rows go to tracking or analytics work first. Unscoped rows leave the backlog.
Reorder what remains by evidence strength, not enthusiasm. The loudest idea in the room is rarely the best next test when comprehension is still broken.
Pair the ledger with platform guidance where it helps — Google Ads — About experiments for ad tests, web.dev — A/B testing for on-site changes — but keep the window and decision columns yours.
Connect backlog to post-launch cadence
After launch, the backlog should stay small and sharp.
- Weeks 1–2: measurement and breakage only — forms, analytics, speed, obvious 404s.
- Weeks 3–6: one comprehension or completion experiment on the money URL.
- Weeks 7–12: one belief or fit experiment — proof, qualification, routing — or a second comprehension test if the first was inconclusive.
What to do in the 90 days after you launch is the calendar. This ledger is the unit of work inside it.
Do not redesign in week two because the inbox is quiet. Run a row with evidence first. Quiet inboxes are often offer or reply problems, not layout problems.
When the backlog says build, not tweak
Some rows fail the “one scoped change” rule because the URL is wrong:
- campaign traffic lands on a homepage;
- proof lives on another domain;
- form routing sends every lead to a black hole.
Those are conversion repair projects, not button tests. Likewise, when spend scales without offline feedback, the row is paid ads analytics before creative.
Digital marketing and web development split implementation, but the backlog owner is one role — usually whoever signs the decision column.
The estimator at get started turns the first inconclusive rows into scoped build work — evidence attached, window closed, decision pending on facts not vibes.
