5 articles
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 →A 2σ scan catches campaign drops like a 67% impression fall. Statistics detect, an LLM explains, evidence persists — snapshot first, dedup, review.
Read article →An LLM blamed 43 correct negative keywords for a 67% impression drop with 70% confidence. Five limitations, four lessons on AI causal reasoning.
Read article →