Why self-attributing networks always look better

Self-attributing networks do not look better because they are better. They look better because they grade their own homework, and the grading scheme is one they wrote. A buyer who accepts a self-attributing network's reported cost per install as comparable to a third-party network's reported cost per install is comparing two different measurement systems and calling the difference performance.

That is the thesis, and plenty of experienced buyers will push back on it. Their argument is that the big self-attributing networks, Meta, Google and increasingly Apple Search Ads and TikTok, do in fact deliver strong cohorts, and that the reporting advantage is a footnote. I think the reporting advantage is the story, because it changes the budget decision before any cohort has had a chance to prove itself.

What self-attribution actually means

A self-attributing network, or SAN, does not send click and impression data to your MMP the way a conventional network does. Instead, the MMP sends the network a list of installs, and the network replies with which of those installs it claims. The MMP then reconciles those claims against everything else it saw, applying its own priority rules.

Three things follow from that structure.

The network sees the install first and decides whether to claim it. It is not competing on a neutral log of touchpoints; it is asserting a match against its own logged-in user graph.

The network's own dashboard reports its claims unfiltered. Your MMP will de-duplicate a claim against a later click from another source. The network's dashboard does not know about that later click and has no reason to care.

The network chooses its own default windows. A seven-day click window and a one-day view window are common, and view-through claims are precisely where reported volume inflates fastest, because a view is cheap to log and hard to disprove.

None of this is fraud. It is a measurement system designed by the seller, run by the seller, on data only the seller can see.

The incentive, not the practice

It is tempting to treat this as a technical quirk. It is better treated as an incentive problem, because incentives predict behaviour when the technology changes.

A network's revenue is a function of your budget. Your budget is a function of reported efficiency. Reported efficiency, for a SAN, is a function of rules the network controls. Every time the network faces a design choice about matching logic, window length or view-through eligibility, the choice that reports more conversions is also the choice that grows its revenue. It would be strange if the accumulated decisions did not lean one way.

The same incentive explains why SAN dashboards handle SKAN and AdAttributionKit the way they do. On iOS, where deterministic matching is limited, modelled conversions fill the gap. The model is trained by the network. Its outputs are reported by the network. Nobody outside the network can inspect how confident it has become about claiming installs that used to sit in organic. So when paid share on a SAN dashboard rises over a year, there is no way to split that rise into real growth versus more assertive modelling. The two arrive in the same column, and the party that built the model is the only party that could tell them apart.

The second-order cost most coverage skips

The obvious cost is over-crediting. The less obvious cost is what over-crediting does to the rest of the portfolio.

If a SAN claims an install that a smaller network's click actually drove, the smaller network's reported CPI rises and its ROAS falls. In the next budget review, that network gets cut. Its inventory may have been genuinely incremental, but it lost a reporting contest rather than a performance contest. Over several quarters this concentrates spend in the networks with the most aggressive claiming logic, which are also the networks with the most pricing power. You end up paying higher effective CPMs to a narrower set of sellers and calling it optimisation.

There is a trade-off in fighting this too. Tightening your MMP's SAN windows or disabling view-through credit will reduce reported conversions from the big platforms. Their algorithms optimise partly on the signal you send back, so a leaner attribution setup can genuinely degrade delivery for a period. The cost of an honest number is a worse-looking dashboard and, sometimes, a temporary dip in real performance while the bidder re-learns.

A decision rule for the weekly review

Rather than argue about which number is right, run a simple discipline.

First, never put a SAN's own dashboard CPI in the same table as an MMP-reported CPI for another network. Use MMP numbers for everything, and accept that the MMP is also imperfect.

Second, compute a claim ratio for each SAN: MMP-attributed installs divided by network-dashboard installs over the same period and geography. As an illustrative example, if the network reports 10,000 installs and your MMP credits it with 7,000, the claim ratio is 0.7. Track that ratio weekly. A stable ratio is fine; you can scale by it. A ratio that drifts downwards while dashboard performance improves is a sign that reporting, not delivery, is doing the work.

Third, for any SAN taking more than a quarter of your spend, run one incrementality test a year against it. Geo holdouts or paused-cell tests are affordable at that scale, and the piece Incrementality Testing for Mobile Games You Can Afford covered the practical designs. The output you want is a single incrementality factor per network, which you then apply to every reported number from that source before it enters a budget conversation.

Fourth, when a SAN's reported ROAS improves sharply without any change in creative or bid strategy on your side, treat it as a measurement event until proven otherwise. Check whether the network shipped a modelling change, a window change or a new default. It usually has.

Fifth, decide in advance how long you will wait before judging a tightened configuration. If you shorten windows or remove view-through credit, the bidder loses signal and delivery can dip while it re-learns. Set a fixed observation period, four to six weeks is a reasonable illustrative range, and compare MMP-attributed cohorts before and after on retention and payer rate rather than on volume. Judging the change in week one guarantees you will reverse it, because week one is when the dip is largest and the honest number is least flattering.

Where this leaves the buyer

The practical position is not hostility towards self-attributing networks. They are large, they often work, and no games studio of scale can ignore them. The position is that their reported numbers are a bid for your budget, and should be read the way you would read any other vendor claim: as an input to be discounted, not a fact to be forwarded to finance.

The discount you apply is the claim ratio times the incrementality factor. Once those two numbers exist for each SAN, the arguments in the weekly review shift from whose dashboard is right to whether the discounted return still clears your hurdle. That is a question a buyer can actually answer.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• Mobile ad networks explained for beginners
• More creative testing can make you worse