Across 640 incrementality experiments run since the start of 2024, Haus found that for every 100 dollars of platform-attributed direct-to-consumer revenue on a 7 day click window, Meta actually generated 115 dollars of incremental revenue.

Under-reporting. Fine, you think, the dashboard is conservative, so apply a correction factor and move on.

In the same dataset, Advantage+ campaigns over-reported by an average of 12 percentage points against manual campaigns.

Read those two together and the comfortable response disappears. The error has a different sign depending on the campaign type, inside the same account. No global adjustment fixes that, and comparing two campaigns in the same dashboard is already invalid before anyone starts arguing about channels.

Last-click was never a measurement. It was a filing rule

Last-click does one thing well: it assigns every conversion to exactly one touchpoint, quickly and without argument. It answers “where do I file this install”, not “what caused this install”.

That was tolerable while the signal was rich enough to make the two look similar. In mobile they have now visibly separated, for reasons that are structural rather than temporary.

Start with the number everyone quotes wrong

The most cited statistic in mobile measurement is the ATT opt-in rate, and it is not one number. Adjust reported an industry-wide opt-in of 35 percent in Q2 2025, measured on users who were shown the prompt. Singular, measuring at first open across the whole install base, reported 9.1 percent for Q3 2025. AppsFlyer has documented the same underlying event producing 40, 36 or 30 percent depending purely on which denominator you choose.

None of those sources is wrong. They are answering different questions, and most teams comparing themselves to a benchmark have never written down which question their own number answers.

Then read what Apple actually documents

Apple’s own StoreKit documentation for SKAdNetwork 4 sets conversion windows at days 0 to 2, days 3 to 7 and days 8 to 35, with a random delay of 24 to 48 hours on the first postback and 24 to 144 hours on the second and third. AdAttributionKit gives a person 24 hours to install after a view-through ad and 30 days after a click, records a maximum of 15 view-through impressions per publisher app, and sends one non-winning postback to up to 5 ad networks.

The part that quietly reshapes reporting is crowd anonymity. Apple’s documentation is explicit that in tier 0 the system sends only the first postback, carrying a two digit source identifier and no conversion value, no source app ID and no country code. Tier 1 returns only a coarse value.

So on low volume campaigns the granular data does not arrive late. It does not exist. Any bid decision made on long-tail geo or placement detail in those campaigns is being made on something reconstructed rather than observed.

On Android there is no equivalent replacement coming. Google announced on 17 October 2025 that it is retiring the Privacy Sandbox technologies, including Attribution Reporting and Topics on both Chrome and Android, and its own status page, updated 14 August 2026, lists every Android ad API as scheduled for phaseout with no published end date.

What replaces it is a stack, not a model

1. Incrementality testing is the closest thing to ground truth. Show ads to one group, withhold from a comparable one, measure the lift. The entry cost has collapsed: Google cut the minimum budget for its incrementality tests from around 100,000 dollars to 5,000, reported in November 2025.
2. Media mix modelling estimates each channel’s contribution from aggregate data, needs no user-level identifier, and no longer needs a six figure vendor. Google Meridian has been open to everyone since January 2025. Meta’s Robyn is open source, and its own repository describes the Python version as experimental, which is worth knowing before it goes into a board deck.
3. Multi-touch attribution still helps where the underlying signal survives, which in mobile is a shrinking share of traffic.
4. The MMP becomes the reconciliation layer rather than the source of truth, collecting SKAN and AdAttributionKit postbacks alongside platform-side reporting and presenting one view.

Three things to do this quarter

5. Publish your opt-in rate three ways, on prompted users, on total installs, and as actual identifier collection, before you compare yourself to any benchmark. One afternoon of analytics ends a recurring argument.
6. Pull your postback tier distribution from your MMP and treat tier 0 as a budget line, not a technical footnote. Collapse source identifier hierarchies on low volume campaigns and stop optimising on detail the system is not sending you.
7. Run a geo holdout before the next budget increase, not after. A workable rule: any channel above 15 percent of UA spend needs a valid holdout each quarter. If the test disagrees with the dashboard, the test wins and the target CPA for that channel is recalibrated on the measured factor.

The limitation

Incrementality and MMM are not the source of truth last-click pretended to be. They are slower, coarser and openly contested.

The IAB’s State of Data 2026, surveying more than 400 senior brand and agency decision makers, found that 60 to 75 percent say advanced measurement falls short on rigour, timeliness, trust and efficiency, and that none of them believe all paid channels are well represented in today’s marketing mix models. EMARKETER and TransUnion found only 27.6 percent of US marketers rate MMM as the most reliable methodology, with multi-touch at 19.4 percent.

For a performance team the practical consequence is a hierarchy rather than a replacement. A holdout gives you one number per channel every two or three weeks. It cannot run your daily optimisation. Tests should govern budget allocation between channels, dashboards should govern adjustment within a channel, and anyone who throws one away for the other will be blind daily and still unresolved quarterly.

Why the commercial model is part of the measurement problem

Every model above is an attempt to answer one question: what did this spend actually cause. Most media contracts do not require anyone to answer it, because the invoice is settled on impressions delivered.

For example, at SpinX the commercial model is CPA, and the conversion is verified by the client’s MMP rather than by us. That does not solve attribution, and it would be dishonest to claim it does. What it changes is where the burden sits. If the action is not recorded by a third party the advertiser controls, it is not billable, so the incentive to inflate attributed volume does not exist on our side of the table.

If you want to see what your current spend looks like when only third-party verified actions are billable, that is the conversation to have at spinx.io.