How Attribution Changes When Agents Make Thousands of Small Decisions

The attribution model for AI agents making micro-decisions across channels is a policy-level one. Credit moves off individual touches and onto the versioned set of rules the agent runs under, and three instruments carry the measurement: a holdout the agent cannot reach, a log of every decision it makes, and an incrementality test that compares one policy with another. Touch-based reports stay on as steering signals. Readers new to the term can start with our explainer on what agentic AI is.

What happens to attribution when agents make thousands of micro-decisions across channels?

An attribution model built around customer touches loses its footing once agents are making thousands of micro-decisions across channels, because the events that move results stop being touches. A bid target nudged overnight, budget moved from paid social to search, an audience excluded, a subject line swapped: none of these appears in a customer path, yet each one changes which paths exist.

Ad platforms were operating at this grain before agents arrived. Google Ads Help says Smart Bidding sets bids for each individual auction instead of a few times a day, reading signals such as the time of day, the ad creative being shown, and the user's device and location. An agent sitting above several platforms adds a second layer of choices about where money and messages go.

That selection is what undermines observational credit. In a paper posted to arXiv in January 2022 and revised that October, Gordon, Moakler and Zettelmeyer analyzed 663 large-scale experiments at Facebook, run in the United States and started between November 2019 and March 2020, to test whether non-experimental methods could "undo" the way the platform selects ads for users. Median lift in the randomized trials was 29%, 18% and 5% for upper, middle and lower funnel outcomes. Double/debiased machine learning, with access to more than 5,000 user-level features, returned 83%, 58% and 24%. Their conclusion on those methods: "we are unable to reliably estimate an ad campaign's causal effect." Touch-based attribution is non-experimental too, and under an agent it has a second selector to see past.

Which attribution models lose meaning, and which keep a job?

The attribution model changes in one direction when AI agents take over micro-decisions across marketing channels: models that weight touches lose authority, and methods that randomize exposure gain it.

According to Google Analytics Help, three models remain in its Attribution reports: data-driven, paid and organic last click, and Google paid channels last click. The page says data-driven attribution "uses your account's data to calculate the actual contribution of each click interaction", which makes it a reasonable signal for an agent to steer by. It remains credit per click inside one vendor's view, so the agent's cross-channel choices go unmeasured.

Why the policy becomes the unit of measurement

A policy, in this sense, is everything that determines how the agent decides: its objective, the actions it may take, its spend caps and approval gates, and the model and prompt version behind it. Individual decisions are draws from that policy, too many and too entangled to credit one at a time. What a finance team can evaluate is whether the business earns more under version 2 than under version 1, and an experiment can answer that cleanly.

Governance already works at this level. The rules for which budget moves an agent makes without asking, and the governance operating model that sets its approval gates, are the policy. Version them and each version becomes something you can test.

How do multi-touch attribution and incrementality divide the work?

With agentic AI, marketing attribution for micro-decisions splits into two jobs: multi-touch attribution stays as the fast signal the agent steers by, and incrementality testing delivers the periodic verdict on the policy. The six-part test design below is our proposal.

1. Freeze the policy. Any change to prompts, models, caps or allowed actions is a new version that needs its own test.
2. Pick a unit the agent cannot leak across: geographic markets for consumer brands, account lists for B2B.
3. Randomize a holdout that keeps the previous policy and the same budget per unit, and exclude it from the agent's credentials, audiences and sends.
4. Log every decision with a timestamp, policy version, channel, object changed, old value, new value, stated reason and any human approver.
5. Measure outcomes outside the ad platforms: orders, qualified pipeline or retained revenue from the system of record.
6. Have a named person approve the holdout, the stop conditions and the threshold for keeping the policy before launch, and every version change after it.

The log needs its own storage. The Google Ads API documentation shows what a platform-side record offers: the change event resource returns the change time, user email, client type, changed fields, and old and new resource values. It also requires a date range within the past 30 days, limits a query to 10,000 rows, and may not include every row from Change History in the web client. An agent making thousands of changes a day needs its own ledger, written at the moment of action, and in a stack with several agents that ledger belongs in the shared governance layer.

Platform lift tools complement the design. Google Ads Help describes Conversion Lift as an incrementality tool with user-based and geography-based methods, and states: "Conversion Lift isn't available for all Google Ads accounts. To use Conversion Lift, contact your Google account representative." Meta Open Source presents GeoLift as a research tool for geo-based experiments where people-based measurement is not feasible, and says Meta recommends people-based experimentation where possible for its higher statistical power. Conversion Lift is a Google Ads tool and GeoLift is an R package for geo tests you plan and run yourself, so a cross-channel agent needs its holdout defined above the platforms.

A worked hypothetical with 40 markets and two policies

Every figure in this example is invented to show the arithmetic; none is a client result or a benchmark. A subscription brand splits 40 matched markets at random. Twenty stay on policy v1: platform auto-bidding with a weekly human budget review. Twenty move to policy v2, where an agent adjusts bid targets, shifts budget, edits audiences and swaps messages across search, paid social and email. Each arm spends $15,000 per market over six weeks.

The agent logged 126,000 decisions in 42 days, or 3,000 a day: 88,200 bid target changes, 18,900 budget shifts, 12,600 audience edits and 6,300 message swaps. A touch-based report has no row for any of them. The policy comparison covers all of them at once: v2 produced 540 more new customers on equal spend, a 9.0% lift on a base of 6,000.

At an assumed $120 of contribution per new customer the lift is worth $64,800, and after an assumed $9,000 to run the agent (tokens, tooling and review hours) the six-week net is $55,800. The platforms told a rosier story, with summed conversions up 17.4%, close to double the measured lift. Their count ran 15% over the order system in holdout markets and 24% over it in agent markets, so in this example an agent steering by platform conversions produced more of them than it produced customers.

Suppose the analysis puts a 90% interval of 2% to 16% around the lift. The rule agreed before launch was to keep v2 if the lower bound sat above zero and the net was positive, so v2 stays. The log then sets up the next test: v2 with message swaps switched off, to learn whether those 6,300 decisions were doing any work. Attribution under agents turns into this, crediting classes of decision through successive policy tests. To run the cost side with your own figures, the AI ROI calculator is a further resource.

Where this design fails

Leakage is the first thing to check. National campaigns, shared audiences and email lists cross market lines, so the holdout has to be enforced in the agent's permissions as well as in the test plan. Policy drift comes second: a model or prompt update mid-test means two policies shared one treatment arm, which is why the version belongs on every log row. Thin data is the third: with few units per arm the interval can be wide, as in the example, and a result that straddles zero should be reported as unresolved.

Where is marketing attribution for AI agents heading after multi-touch?

The future of marketing attribution with AI agents keeps multi-touch data as an input and hands the verdict to experiments. What vendors document today is per-auction bidding, three remaining models in Google Analytics, a lift tool with restricted availability, and an API change log limited to the past 30 days. A policy-level holdout above the platforms, paired with a decision ledger, is our proposed assembly of those parts. In the pages we read, we found no ready-made report that attributes revenue to an agent's policy version across channels, so our agentic AI automation practice treats the holdout and the ledger as part of the build.

The practical change shows up in the monthly review. "Which channel gets the credit?" gives way to a three-part question: which policy version was live, what was held out, and what did the difference earn?

Sources

• Google Ads Help, Google Ads automated bidding
• Gordon, Moakler and Zettelmeyer, Close Enough? A Large-Scale Exploration of Non-Experimental Approaches to Advertising Measurement (arXiv, 2022)
• Google Analytics Help, Get started with attribution
• Google Ads API documentation, Change Event
• Google Ads Help, About Conversion Lift
• Meta Open Source, Welcome to GeoLift