Reserve a learning line in your 2027 budget for inventory you cannot isolate
Ad slots, AI answers and map placements are arriving inside existing campaigns or under new controls. A budget structure that funds learning without putting core performance at risk.
The 2027 plan most teams are about to write will have the same three rows as last year: Apple, Meta, Google. This year those rows hide more change than usual. Apple added ad positions in App Store search that you cannot target separately. Google is testing ad formats inside AI answers that run through existing campaigns. Apple has announced ads in Maps, and a Supreme Court case could decide the cost of sending users to a web checkout. When you cannot isolate new inventory, you cannot judge it on the usual metrics. Budget for that explicitly instead of letting it leak into core numbers.
The problem with blended inventory
If a new placement serves inside an existing campaign, its results arrive averaged into the numbers you already track. A modest change in the mix shifts your blended CPA, and you have no label telling you why. That is the situation with Apple's additional search positions: Apple said existing search results campaigns are eligible automatically and advertisers cannot choose or bid on a position. It is also the likely situation with Google's AI Mode formats, which Google described in May as being tested, with preparation through AI Max and Performance Max campaigns rather than a separate buy.
The trouble is not only that you cannot see it. It is that two bad habits follow. One is to treat any CPA drift as a performance problem in the campaign, and respond by cutting bids. The other is to treat any improvement as proof that the new inventory works. Neither follows from the data.
Fund a learning line, and keep it small and finite
A learning line is a fixed share of the budget with a purpose: find out whether something new earns a permanent place. A workable structure for a mid-sized budget looks like this, using round numbers as an example only.
- Core, around 85 percent. Channels and campaigns with a track record, managed to payback targets.
- Controlled tests, around 10 percent. Experiments with a design: a holdout, a geo split or a clear before and after, a pre-registered metric and an end date.
- Reserve, around 5 percent. Unallocated money for things that appear mid-year, such as a new placement that becomes available to you, so that you do not have to raid core budgets to try it.
The proportions depend on your risk tolerance. The principle is that the learning line is decided in advance, so a test is a planned expense and not a reaction to a headline.
Write the test before you spend the money
For each item in the learning line, write down four things on one page:
- The question. For example: do Maps placements generate orders or visits in your stores, at a cost we can accept? Or: does a creative-asset header lift trial starts per tap?
- The metric. One outcome you pay for, measured outside the platform where possible.
- The design. How you will separate the effect from everything else: matched markets, a holdout, or alternating weeks.
- The decision rule. What result would make you scale, keep, or stop. Write the threshold before you see data.
This sounds bureaucratic until the first time a stakeholder asks whether a new surface is working and you can answer with a decision rather than an opinion.
What should be in the 2027 learning line
Your list will differ, but these are the candidates we would consider, with the facts as announced so far.
- Apple search positions. Not separately controllable, so the test is on brand defense and bid levels, using the first-position and all-positions impression share metrics Apple provides.
- Apple creative assets. In an October 5 developer post Apple described new product page header and search result assets and a preview tool in App Store Connect. Apple's post does not say how they interact with ads, so test them against your existing product page variants and read the result on a downstream metric.
- Google AI formats. Tests in AI Mode were announced in May. If they reach your account, an AI Max experiment against a control is the cleanest way to read them.
- Apple Maps. Only relevant if you have physical locations, and Apple has announced it for the US and Canada. A matched-market holdout is the right design.
- Web-to-app checkout. The Supreme Court's review of the link-out case means the commission rule may change, so test with scenario margins and keep the in-app path healthy.
Protect the core from the experiments
Keep test campaigns in separate campaigns or accounts where the platform allows, with their own budgets, so that a failed test cannot drag a core campaign's learning phase down. Where a placement runs inside a core campaign and cannot be split, change nothing else in that campaign during the test window.
Agree in advance how long a test runs and who decides. Tests with no end date turn into permanent line items.
Plan to be wrong about some of it
The best use of this structure is the discipline it forces. Some of these placements will turn out to matter and some will be background noise. A 2027 plan that says which is which, with evidence, will be much easier to defend than one that adds a new line for every announcement.
The takeaway: set aside a fixed, small learning budget, write each test's question, metric, design and decision rule in advance, and keep the core budget managed to payback.