Re: [attribution] Methodological concerns regarding Attribution Level 1 and causal measurement

Hi Charlie,

Thank you for the response. I appreciate the clarification, and I largely
agree with your characterization of what could theoretically be done using
the API within an explicit experimental framework.

My concern is less that the API could never be used as part of an
incrementality study, and more that the specification itself overwhelmingly
presents attribution as though it is already sufficient for measuring
advertising effectiveness in and of itself.

For example, the document repeatedly states that attribution helps
advertisers understand “what ads perform best,” allows them to spend more
on “effective advertising,” and helps determine “which creative works
best.” It also defines attribution itself as measuring “correlation between
one or more ad placements … and the outcomes that an advertiser desires.”

At the same time, the document only briefly references controlled
experiments almost in passing, for example noting that “not displaying an
advertisement” could be used for “controlled experiments that seek to
confirm whether an advertising campaign is effective.”

That distinction is important. Correlation telemetry and causal measurement
are not the same thing, and the advertising industry has spent well over a
decade routinely conflating attribution reporting with incrementality
measurement. My concern is that a browser-level standards document risks
institutionalizing that confusion if it does not much more clearly
distinguish between:

   1.

   privacy-preserving event telemetry infrastructure, and
   2.

   scientifically valid causal measurement methodology.

I agree that an experiment of the type you describe could potentially be
implemented on a single publisher or platform. But in practice, I think
there are substantial operational and methodological limitations.

The phrase “divert users using a 1P identifier” is doing a great deal of
work in that workflow. Many advertisers do not possess stable first-party
identity for the populations they most care about influencing, particularly
in acquisition campaigns aimed at prospective customers rather than
existing CRM populations.

Likewise, many publishers on the open web do not operate large
authenticated identity systems at all. A publisher such as the NY Post, for
example, does not have persistent logged-in identity for most readers. This
creates a structural asymmetry in favor of large walled gardens and retail
media platforms with massive authenticated user bases and deterministic
identity infrastructure, while many independent and mid-sized publishers
are already under severe economic pressure.

In practice, many implementations of user-level incrementality measurement
in these ecosystems appear likely to depend on some combination of:

   -

   hashed email onboarding,
   -

   phone-number matching,
   -

   CRM reconciliation,
   -

   identity graphs,
   -

   clean rooms,
   -

   LiveRamp-style interoperability layers,
   -

   or platform-controlled measurement environments.

At that point, the Attribution API itself no longer appears to be the
primary source of causal validity. The causal validity derives from the
randomized experimental design and the surrounding identity-reconciliation
infrastructure.

The API seems primarily to contribute privacy-constrained exposure and
conversion telemetry.

More importantly, many of the hardest scientific problems remain unresolved:

   -

   highly lossy match rates,
   -

   identity fragmentation,
   -

   sample bias between matched and unmatched populations,
   -

   cross-platform inconsistency,
   -

   optimization feedback loops,
   -

   and external-validity limitations.

The match-rate issue is particularly important. In many real-world systems,
exposure matching against advertiser sales or CRM systems is highly
incomplete, often near or below 50%. The matchable population is
systematically different from the unmatchable population, making it
difficult to generalize measured lift to the total exposed audience.

I also believe user-level platform experiments are intrinsically vulnerable
to targeting and auction-system dynamics. Modern large platforms
increasingly optimize delivery toward users already exhibiting behavioral
indicators associated with imminent purchase intent. In many cases, the
experiment therefore measures the effect of advertising on a highly
pre-filtered audience already selected by self-selection, behavioral
targeting, and AI optimization systems.

That is not necessarily the same question many advertisers, CFOs, or
economists care most about, which is broader marketing allocation across
the full media mix.

More broadly, I think there is an important privacy paradox embedded here.

The specification is motivated by legitimate privacy goals, which I
support. But many practical implementations of identity-based
incrementality measurement appear likely to increase incentives for:

   -

   more first-party identity harvesting,
   -

   more authenticated-user environments,
   -

   more CRM onboarding,
   -

   and more identity reconciliation infrastructure.

In other words, the ecosystem incentives may move further toward persistent
identity systems rather than away from them.

This dynamic also risks further concentrating measurement and optimization
advantages inside the largest platforms with the richest identity and
commerce graphs.

For these reasons, I continue to believe the specification would benefit
from much clearer language distinguishing:

   1.

   attribution telemetry infrastructure, and
   2.

   causal advertising effectiveness measurement.

I do appreciate the discussion very much, and I appreciate your engagement
on the topic.

Best,
Rick

>