The deprecation notice, read in full

Multi-touch attribution (MTA) is ad-tech's name for the credit-assignment problem. A customer sees eight ads across three weeks and then buys; which ads get credit for the sale? The answer determines where the next dollar of budget goes, so the industry spent two decades on it, and for most of that time the standard answers were rules. First-click gave all credit to the first touch. Last-click gave it to the final one. Linear spread credit evenly across the path. Time-decay weighted recent touches more heavily. Position-based split a bonus between first and last. Every one of these is a positional heuristic: a fixed weighting over the trajectory, chosen by convention rather than derived from outcomes.

Google's current Ads documentation records where that era ended. "The first click, linear, time decay, and position-based attribution models are no longer supported by Google," it states, and conversion actions that used the deprecated models "have been upgraded to use data-driven attribution," now the default attribution model for most conversion actions.1 Two precise qualifications matter here, because a sloppy retelling of this story over-claims. Last-click was not deprecated; it remains the one rule-based model still available. And Google publishes no date for the deprecation. The live page records only the standing position, so this is best treated as the industry's current posture rather than a datable event in a timeline.

A third qualification is about the witness. Google is not a neutral narrator of its own deprecation. A single-vendor model is easier to sell than an auditable rule an advertiser could second-guess, and that commercial motive is a genuine alternative explanation for the retirement of the rules, distinct from "the empirics settled it." It does not make the replacement's logic wrong, and it does not undo the fact that the rules are gone. It does mean the deprecation is weaker evidence than a controlled comparison would be, and this paper's argument leans on the deprecation. We flag that openly rather than dressing a market actor's product decision up as a scientific verdict.

Data-driven attribution (DDA) is a counterfactual method, and Google's practitioner documentation describes its core move in one sentence: "by comparing the paths of customers who convert to the paths of customers who don't, the model identifies patterns among those ad interactions that lead to conversions."2 The model then assigns more credit to the interactions its path comparison associates with conversion. The contrast with the rules is structural, not incremental. A positional heuristic asserts a weighting; DDA estimates one, from the difference between converting and non-converting trajectories.

Notice what the counterfactual framing buys. When two engineers disagree about whether the first retrieval or the last tool call deserves the blame, a rule-based system has no way to adjudicate: both weightings are conventions, and the argument is a matter of taste. A path-comparison system turns the same question into an empirical one. Do trajectories that include this touch convert at a different rate than matched trajectories that don't? That is a question with an answer, and the answer does not depend on which engineer shouted last.

It is worth being precise about what the industry gave up to get that property, because the costs were published too, and they are the half of the lesson that transfers most cleanly.

Counterfactual credit assignment needs volume, and Google says so with numbers. Its recommendation for DDA is "at least 200 conversions and 2,000 ad interactions in supported networks within a 30-day period," alongside an important softener: the model "will still function with less data," but sufficient volume "allows the model to better identify patterns and assign credit more precisely."2 Read carefully, that is a precision recommendation, not an eligibility gate. The method does not refuse to run below the floor. It degrades, gracefully but genuinely, and below some volume you are back to guessing, with more machinery between you and the guess.

The transfer to evals is direct. Positive outcomes in an agent-eval setting are the conversions: sparse, delayed, and expensive to label. A counterfactual credit method for agent traces would need enough outcome-labeled trajectories, on both the success and failure sides, for path comparison to find real patterns instead of noise. Teams sizing a step-credit initiative should treat "how many outcome-labeled traces per month do we actually produce?" as the first feasibility question, before any modeling choice. Google's published floor is not your floor (different data, different signal-to-noise), but the existence of a floor is the durable lesson, and it is one the vendor selling you a step-scoring product is unlikely to volunteer.

The second published cost is the one regulated deployers should sit with. On the Ads help pages, the mechanism behind DDA is described only as credit "as determined by Google AI."1 No named algorithm, no inspectable math, no per-decision explanation. Scope that claim precisely: it describes the practitioner-facing Ads help surface, and our source research notes that Google's analytics documentation elsewhere discloses more about the counterfactual framing (a carried claim we did not re-fetch live, flagged in Section 7). But the product documentation an advertiser actually reads asks them to accept the credit split on trust.

Google judged that trade worth making, and for advertising it arguably is: the decision being explained is a budget split, and the counterparty asking "why" is a marketing team. An eval pipeline in a regulated environment faces a harder version of the same question. If a step-level credit system blames step 7 of a clinical-workflow agent, or assigns fault inside a lending decision trace, someone with authority may eventually ask for the reasoning, and "the model decided" is a materially harder answer to give a compliance reviewer than an advertiser. The lesson is not "never accept opacity." It is that opacity is a cost with a price that varies by context, the industry that pioneered counterfactual credit paid it knowingly and wrote it down, and a team adopting the same architecture should decide, in advance and in writing, whether their context can afford it.

Step-level credit assignment for agents is structurally the same problem MTA was solving. The current wave of process reward models (PRMs) makes the parallel concrete: a PRM scores each intermediate step of a reasoning or tool-use trace rather than only the final outcome. Lightman et al.'s "Let's Verify Step by Step" established the process-supervision reference point with step-labeled human feedback3, and Wang et al.'s Math-Shepherd extended it with automatically constructed step labels, no human annotation required.4 We cite these as the field's landmarks, not as evidence for the attribution analogy itself; the overlap between the PRM literature and the ad-tech mapping was outside our research pass's verification budget, and we say so rather than imply a literature review we did not do.

The structural rhyme is what carries the argument. A trajectory of touches ends in an outcome. You want per-touch credit. Outcomes are sparse and delayed relative to the touches. And a fixed positional weighting is the cheap first answer that many teams reach for, because it ships in an afternoon. First-step weighting is first-click. Last-step weighting is last-click. Uniform step credit is the linear model. The eval field is currently writing, with fresh conviction, the exact heuristics whose retirement the measurement industry has already documented.

Two honesty notes on the comparison's scale. Google Ads is a large-scale production example, and it is tempting to say the deprecation happened "with more data than any eval team will ever have"; but we have no citable measurement comparing Google's data volume to any given eval team's, so treat that framing as illustrative. Similarly, Google's documentation does not narrate a decade of practitioner disagreement over dueling rule models; it records only the deprecation. What licenses the inference is the shape of the replacement: retiring every rule-based model except one, in favor of a method whose design resolves arbitrary-weighting disputes empirically, is a strong signal that unprincipled adjudication was a real, felt cost.

Here is where this paper declines to become a feature pitch. The research spike behind this series evaluated the MTA-to-eval transfer against a real system and rendered its verdict as "inverted": the transfer's entire value, today, is a "do not build this" constraint rather than a feature specification. The reason is data plumbing, and it is worth stating mechanically.

The system studied (an internal multi-model agent-evaluation harness KellerAI builds and operates itself, not a third-party or client system) captures exactly one trace event per subject invocation. The agent is treated as a single opaque call: no internal tool-call sequence is instrumented, and the result schema is flat, with no list-of-steps field anywhere in it. DDA-style counterfactual credit requires step-level, outcome-labeled path data for successful and failed runs alike. You cannot run a converting-versus-non-converting path comparison on traces that do not record paths. Two structural gaps, event granularity and schema shape, sit between this harness and the method, and neither is a modeling problem. They are instrumentation work.

How far does that generalize? We have not audited other harnesses, so we will not claim an industry statistic. Treating the agent as one opaque call is a common default in the eval tooling we have seen, and the honest recommendation is to check your own stack before assuming either way: if your trace store cannot answer "what was the full step sequence of a failed run last Tuesday," you have the same gap.

None of this rules out a cheap heuristic as an explicitly provisional stopgap. Forensic debugging of a single trace, a low-stakes prototype eval with no audit requirement, a genuine cold start before any labeled data exists: all defensible uses of a positional rule, provided the rule is labeled temporary and carries a review date. The ad-tech warning is about the other pattern: the "interim" heuristic that quietly becomes the permanent system and survives a decade, which is roughly the history the deprecation notice closes out.

The sequencing for a team tempted by step-level credit today:

• Don't ship positional heuristics as the permanent credit system. If you need a stopgap, label it as one, with a review date attached and the retirement criteria written down.

• Instrument first. Step-level, outcome-labeled trace capture is the prerequisite for any principled credit method, it is useful for debugging regardless, and it is the long-lead-time item.

• Budget for both published costs before building the counterfactual version: a data-volume feasibility check against your actual monthly volume of outcome-labeled traces, and a written auditability decision for your regulatory context.

And the limits, so a skeptical reader knows exactly where this analysis is load-bearing and where it is not. The Google Ads claims were verified against the live help pages on 2026-07-06.1 The claim that Google's analytics documentation discloses more methodology elsewhere is carried from our source research and was not re-fetched live. The PRM citations are landmarks, not a surveyed literature.3 The "common default in eval tooling" observation is anecdotal. No deprecation date exists in Google's current documentation, and we assert none. And the entire argument leans on a market actor's own account of retiring its own products: strong enough to justify the negative constraint this paper actually draws, and not stronger.

This paper proposes no new method, and that is deliberate. It is the measurement industry's own deprecation notice, read closely and read across. Ad-tech already paid for this lesson once; the cheapest option available to the eval field is to read the receipt.