The Measurement Plan to Agree Before You Spend

Why this conversation is easier before spend starts than after

Before any money has been spent, everyone in the room is reasoning abstractly — nobody has a stake in defending a specific number yet, because no number exists. That's precisely the window in which a genuinely honest measurement plan gets agreed. Once spend starts and early results come in, every subsequent conversation about methodology carries the shadow of whoever's position it currently favours. The plan agreed in week one, before results exist, is a fundamentally more trustworthy document than one negotiated in week eight, after a disappointing number has already made someone defensive.

It also keeps the plan short. A good one fits on a single page and is signed by both the brand and whoever runs the media, so that nobody has to remember in month three what was meant in week one.

What the plan actually needs to specify

Name the default attribution model for routine reporting, and the reporting window for each ad product involved — Sponsored Products, Sponsored Brands, Sponsored Display and DSP each run different windows, and DSP itself splits by whether inventory is in-store or offsite. Amazon's own guide to the basics of Amazon Attribution is a good place to confirm what each product reports before you write the windows down.

Name the incrementality test design specifically: holdout or geo-lift, on which channel or audience, over what minimum duration — with an explicit note that the learning window shouldn't be cut short before audience models settle. Pulling a test early is one of the most common ways to end up with a read nobody trusts.

Name a minimum spend threshold below which a formal incrementality test isn't statistically reliable, so nobody demands a definitive causal answer from a budget too small to produce one. Where DSP and sponsored ads both run, also say how the two will be reconciled so the same purchase isn't counted twice; Amazon Marketing Cloud is the usual place that happens.

The clause that matters most: what happens if it's bad news

This is the part most measurement plans skip, and it's the one that actually earns the document its value. State explicitly, in advance: if the incrementality test shows minimal or no lift, what happens next — does the channel get cut, does budget move, does a second test run at a different scale before a final call. Agreeing this before the result exists means a disappointing test doesn't turn into a negotiation about whether the test itself was valid; the response was already decided when nobody had a reason to fight about it.

A worked version of what this looks like in practice

For example, say a brand is launching a new DSP prospecting line at $15,000 a month. The agreed plan: last-touch attribution for weekly reporting, a geo-lift incrementality test starting in week 5 (after the initial learning period) running 6 weeks, a $10,000 monthly minimum before treating any single-market read as reliable, and a bad-news clause stating that if the geo-lift shows under 5% lift after the full 6 weeks, budget reallocates to retargeting for the following quarter with a re-test scheduled in 6 months. When the test comes back at 3% lift in month three, there's no debate — the plan already said what happens, and the conversation moves straight to execution rather than relitigating whether 3% counts as "good enough."

The arithmetic is worth doing on your own numbers too. In this example, a $15,000 monthly budget over a 6-week test is about $22,500 of spend (15,000 × 1.5 months), so the plan should say up front that this is the budget the test is allowed to consume, and that it sits above the $10,000 monthly floor.

A one-page checklist to copy

Before the first dollar goes out, each line below should have a written answer:

• Attribution model: which model governs weekly reporting, and which window applies to each ad product.
• Primary metric: the one number the decision will be made on (for example return on ad spend, cost per acquisition or new-to-brand share), so a different metric can't be swapped in later.
• Test design: holdout or geo-lift, which audience or market, start date and minimum length.
• Who designs and computes it: and whether the design is written into the statement of work.
• Spend floor: the monthly spend below which results are read as directional.
• Bad-news clause: the agreed action if lift is minimal or absent.
• Review dates: when the group meets to read results, and who has the final say.

If any line can't be filled in, that is the conversation to have before launch, not after.

Why a headline return is not a measurement plan

Across 30 advertisers in July 2026, the reMKTR portfolio delivered a 6.04x return on ad spend, with 20.1% of attributed purchases coming from shoppers new to the brand. Those are figures from our own book in one month, not a promise about any other account, and they show why the plan matters: a return figure only means something once you know the attribution model and window behind it. The same spend can read very differently under a different window, and none of those readings says whether the ads caused the sale. That is a question for the test design, which is why it belongs in the plan.

If you want to see how we think about the wider picture, our guide to advertising on Amazon covers sponsored ads versus DSP, and reMKTR runs Amazon DSP as a service.

The common mistake, including ours

The mistake is treating the measurement plan as a formality to draft after the media plan is already locked, rather than as part of the media plan itself. We've launched DSP spend for a client before the incrementality test design was fully agreed, planning to finalise it "once things were running" — and by the time we circled back, three weeks of spend had already happened without a clean test window, meaning the earliest reliable read was pushed back by the exact amount of time we'd spent not finishing the plan. We now treat the measurement plan as a launch blocker, not a follow-up task.

What to do if a plan wasn't agreed and spend is already running

If spend is already underway without a written plan, the honest fix isn't to pretend one existed retroactively — it's to build the plan now, dating it clearly as starting from today, and accepting that any test window measured against the period before the plan existed is weaker evidence than one measured cleanly after it. A late plan is still meaningfully better than none, and the sooner it's written, the sooner a genuinely clean test window can begin.

Who should design the test, and why that's part of the plan too

"Is this incremental, or would they have bought anyway" is the single most common objection we hear in DSP conversations, and it's fundamentally a trust question, not a technical one. The honest answer is naming, in the plan itself, who designs and computes the test — a holdout or geo-lift run entirely by the party being paid on the result, with no agreed design written down in advance, isn't a test a skeptical stakeholder should fully trust, including from us. Writing the test design, the holdout definition and the reporting cadence into the plan before spend starts is what makes that objection answerable rather than just reassuring.