Turning AI Search Agents into Your Attribution Analysis Sidekick

Attribution analysis is the classic problem of determining why a metric rises or falls. The standard workflow consists of three stages: anomaly confirmation (comparing against a baseline such as yesterday, last week, or a historical average), investigative analysis (decomposing factors like revenue = impressions × CPM, examining dimensions such as merchant, channel, or placement, and correlating with internal or external events), and producing a readable, verifiable attribution report.

To automate this workflow, the team designed an Agent that mirrors the analyst’s reasoning as a Plan (choosing analysis path) and Action (executing specific operations). The Agent’s action space defines which tools it can call, what data it can query, and which calculations it can perform. A large language model supplies basic analytical knowledge, while the systematic design translates human experience into machine‑executable steps.

The action‑space design starts with the Plan stage, where the Agent decides the time window (e.g., 7 days vs 30 days), factor‑prioritisation order, and which dimension to explore first. The SubAgent then carries out the Action stage using a toolbox that includes data‑query APIs, contribution‑calculation utilities, code‑execution engines, event‑lookup services, and schema‑query interfaces that resolve naming ambiguities (e.g., the exact definition of “revenue” in different business contexts).

In a concrete scenario, a user reports a sudden rise in a revenue metric over the past two days. The main Agent generates a Plan to compare recent data with the same period last year, confirms the anomaly, and spawns two SubAgents in parallel: one evaluates the factor decomposition (revenue = impressions × CPM) and finds that CPM contributes 79 % of the uplift; the other checks merchant‑level revenue changes. Depending on the results, the Agent either drills deeper along the high‑contribution path or switches to another dimension such as region.

Pain point 1 – exponential search space: Factor branching can produce dozens of paths and dimension combinations, leading to linear‑time growth if explored serially. The solution is a parallel‑plus‑pruning architecture that generates multiple attribution hypotheses simultaneously, builds an "attribution tree", and prunes branches whose validation results are insignificant (e.g., uniformly distributed factor contributions).

Pain point 2 – optimal solution may lack business value: The Agent might discover a mathematically optimal answer that is irrelevant to the business, such as a long‑tail merchant contributing 30 % of the change but having a negligible revenue base. To steer the Agent, a "script" mechanism acts as a lightweight SOP template. Users can supply full scripts or script fragments (e.g., "When category A is abnormal, next examine sub‑category B") which the Agent follows, ensuring analysis stays within business‑approved constraints.

Pain point 3 – constraint drift: In long chains, downstream queries must inherit filters from upstream steps (e.g., "merchant = A"). The Agent may lose or misuse these constraints, breaking the attribution chain. The team introduced a dynamic structured memory that records each node’s purpose, parameters, schema information, and success/failure status. During back‑validation, the system reconstructs the full chain, detects drift, and automatically corrects mismatched constraints.

Interactive capability: Because attribution is time‑consuming, the system presents the current analysis plan to the user before each verification round, allowing manual edits. Users can also issue a "no need to confirm" command, prompting the Agent to switch to a fully automatic mode and output the final report, thereby reducing trial‑and‑error cost.

Future directions: The roadmap includes expanding the toolbox with more scientific methods and visualisation, evolving from a single‑agent to a multi‑agent organisational structure where different agents play distinct business roles, building a closed‑loop "discover‑investigate‑adjust" capability, enabling composite attribution across business lines, exploring "unknown attribution" to surface blind‑spot factors, and upgrading experience reuse by archiving every attribution attempt as structured knowledge with timestamp‑based priority and conflict‑resolution rules.