Daily state snapshots, a full event timeline and AI causal reasoning that attribute growth to what actually moved it — not last-click guesses.
The AI State-Transition Attribution & Growth Engine treats marketing as a sequence of states, not a pile of isolated charts. Every entity — account, campaign, page — gets a daily snapshot of metrics, every optimization becomes a timestamped event, and AI links the two with causal reasoning and evidence. When a metric jumps or falls, the engine can say which transition caused it, with confidence, and proposes the next experiment.
Accounts and campaigns are snapshotted every day into a time-series state model — metrics, waste cost and a health score, with delta computation.
Every optimization — budget, bid, keyword, ad, negative — becomes a timestamped growth event with old/new values.
Anomaly detection triggers LLM reasoning that links cause and effect with evidence_links — and human corrections feed back into the model.
Statistical thresholds catch meaningful metric shifts before they compound, deduplicated per campaign+metric per day.
For every detected anomaly, AI estimates the expected direction and magnitude of impact before you decide what to do.
State a goal and the engine decomposes it into tasks, budgets and a timeline — then tracks execution as events.
Any fix can become an experiment with baseline snapshot, automatic evaluation and a SUCCESS/FAILURE verdict.
State sequences render as trend lines for impressions, clicks, cost, waste and impression share — a dashboard that reflects time, not totals.
Every diagnosis carries confidence, reasoning and evidence links, tracked through adopt/reject/implement.
Human verdicts and experiment outcomes feed back to sharpen future reasoning — the engine gets better the longer you run it.
Attribution is a time-series model: daily state snapshots plus an event timeline, so credit lands on the transition that moved the metric.
State transitions, evidence chains and experiments explain what changed and why — not just that a number moved.
Every conclusion is a transparent chain: anomaly → evidence links → reasoning → confidence → recommendation.
The engine compounds: experiments and human feedback make each round of diagnosis sharper than the last.
A single spend number can't tell you if an account is improving. How to build a daily state snapshot time-series model with idempotent snapshots, delta computation and 14-day trends.
Read article →Budget, bid, keyword and negative changes were scattered and unverifiable. How we auto-log every optimization as a timestamped event — entity, field, old and new value — and what it took to make the record reliable.
Read article →LLM JSON parsing kept failing with empty strings and prose-wrapped arrays. Twelve logged failure patterns traced to one root cause: thinking mode. Turn it off, batch, and parse in layers.
Read article →