The industry that invented positional credit rules retired them

Multi-touch attribution (MTA) is ad-tech's name for the credit-assignment problem: a customer saw eight ads before buying, so which ads get credit for the sale? For years the standard answers were rule-based models with names that will sound uncomfortably familiar to anyone building agent evals: first-click (all credit to the first touch), last-click, linear (spread it evenly), time-decay (recent touches matter more), and position-based (bonus credit to first and last).

Google no longer supports most of them. Its Ads documentation states plainly that "the first click, linear, time decay, and position-based attribution models are no longer supported by Google," and that conversion actions using the deprecated four "have been upgraded to use data-driven attribution," now the default model for most conversion actions.1 Last-click is the one rule-based holdout still available. Google's current documentation gives no date for the deprecation, only the fact of it, so treat this as the industry's standing position rather than a datable event.

What replaced the rules is worth reading closely. Data-driven attribution (DDA) is a counterfactual method: "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," then gives more credit to the interactions that actually moved outcomes.2 Not "the last touch matters most because we said so." Instead: compare journeys that ended in a conversion against journeys that didn't, and let the difference assign the credit.

One caveat before leaning on this story. Google is not a neutral witness to 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 real alternative explanation, distinct from "the empirics settled it." That doesn't make DDA's counterfactual logic wrong. It does mean "the market leader retired the rules" is weaker evidence than a controlled comparison, and this post leans on the former.

The deprecation is only half the lesson. The other half is what Google's own documentation admits the replacement costs.

First, data volume. Google recommends "at least 200 conversions and 2,000 ad interactions in supported networks within a 30-day period" for DDA to assign credit precisely, while noting the model "will still function with less data."2 That is a precision recommendation, not an eligibility gate, and the distinction matters: counterfactual credit assignment degrades gracefully but needs volume to be trustworthy. There is a floor below which you are back to guessing, just with more machinery.

Second, auditability. On the Ads help pages, the mechanism behind DDA is described only as credit "as determined by Google AI" — no named algorithm, no inspectable math.1 To be fair to Google, this opacity is specific to the Ads help surface; its analytics documentation elsewhere discloses more about the counterfactual framing. But the practitioner-facing product documentation asks advertisers to accept the credit split on trust. Google judged that trade worth making. Whether you can make it, in a regulated eval pipeline where someone may ask you to justify why step 7 was blamed, is a different question, and an honest reading has to leave it open.

Step-level credit assignment for agents is structurally the same problem MTA was solving. That includes the current wave of process reward models (PRMs), which score each intermediate step of a reasoning or tool-use trace rather than only the final outcome; "Let's Verify Step by Step"3 and Math-Shepherd4 are the field's reference points. We cite them as landmarks, not as evidence for the analogy: PRM-literature overlap with the attribution mapping was outside our research pass's verification budget. The structural rhyme is what matters. A trajectory of touches ends in an outcome; you want per-touch credit; outcomes are sparse and delayed; and rule-based positional heuristics are the cheap, tempting first answer.

The record above is the strongest argument we know of for not adopting those heuristics permanently. Not because a linear or last-step weighting won't ship; it will, and that is its appeal. Google Ads is a large-scale production example where several positional attribution rules were retired in favor of a counterfactual alternative. (We have no citable measure of Google's data volume against any given eval team's, so read "they had more data than you ever will" as illustrative, not measured.) Google's public documentation doesn't narrate a decade of practitioner disagreement, either; it records only that the rules are deprecated, with no date given. But retiring every rule-based model except last-click in favor of a method built to resolve exactly this kind of arbitrary-weighting dispute is a strong signal that unprincipled adjudication was a real, felt cost. When your first-step-weighted eval metric and your colleague's last-step-weighted one disagree, there is still no principled way to pick a winner.

Here is where this post refuses to become a feature pitch, because the underlying research is blunt about it. The internal research spike behind this piece rendered its verdict on the MTA-to-eval transfer as "inverted": the transfer's entire value is a "do not build this" constraint, not a feature spec.

The reason is data plumbing. The system studied in that research (an internal multi-model agent-evaluation harness KellerAI builds and operates itself) captures exactly one trace event per subject invocation, treating the agent as a single opaque call: no internal tool-call sequence, no list-of-steps field anywhere in the result schema. We haven't audited other harnesses to confirm how far that generalizes. Treating the agent as one opaque call is a common default in the eval tooling we've seen, but check your own stack before assuming either way. DDA-style counterfactual credit needs step-level, outcome-labeled path data for successful and failed runs alike, and you cannot run a converting-versus-non-converting path comparison on traces that don't record paths.

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, or a genuine cold start before any labeled data exists. The ad-tech warning is about quietly adopting the heuristic as the permanent system, the way an "interim" attribution rule survives a decade. It is not an argument against using one that is labeled temporary while you build the real instrumentation.

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

• Don't ship positional heuristics as the permanent credit system. The industry that invented them retired nearly all of them. If you need a stopgap, label it as one, with a review date attached.

• Instrument first. Step-level, outcome-labeled trace capture is the prerequisite for any principled credit method, and it earns its keep in debugging regardless.

• Budget for both published costs before building the counterfactual version: a real data-volume floor before the credit split is trustworthy, and an auditability story, because "the model decided" is a harder sentence to say to a compliance reviewer than to an advertiser.

This post proposes no new method, and that is deliberate. It is the measurement industry's own deprecation notice, read across to a field about to re-run the experiment. Ad-tech already paid for this lesson once; the cheapest option is to read the receipt.