Change history is Google Ads' built-in log of everything that changed in an account: the entity touched, the old and new values, the timestamp, the user who made the change, and the surface it came through. It answers the only question that matters after an unexplained performance shift — what changed, when, by whom, from where — and it holds two years of answers.
It is also the report that makes AI operations in an ad account inspectable rather than mysterious, because it records machine-made changes with exactly the same fidelity as human ones.
How change history works
Every change lands as a row with source attribution. The source values are worth learning to read, because they carve the account's activity into different kinds of actor: the web interface (a person clicking), the API (software acting under a user's authorisation — including AI connectors), Google Ads Editor and bulk uploads, automated rules and scripts you scheduled, and auto-applied recommendations — changes Google itself made under a setting you enabled. A surprising number of "who changed this?" mysteries resolve to that last category.
The log is filterable by date range, campaign, change type and user, which is what makes it usable as a diagnostic instrument: narrow to the window where performance bent, and read what happened just before.
Many rows carry a control that reverses them — budget changes, status changes and bid changes among them — but reversal has limits worth stating precisely. Not every change type can be reversed from the log, and reversing a change does not erase history: it applies the old value back as a new change, itself recorded. The log is append-only in spirit — a black-box recorder, not an eraser.
When the AI both reads and writes the log
Two things change, one in each direction.
First, the log becomes readable at conversational speed. Correlating changes with performance is the classic manual chore — export the history, line it up against the metrics, squint. An agent with account access does the join directly: performance bent on the 14th; here are the changes in the preceding days, and the two that plausibly explain it. Auto-applied recommendations, the source people forget to check, stop escaping attention because the agent reads all sources equally. The habit of verifying what changed after any shift becomes cheap enough to be routine.
Second, the log now contains the AI's own work — and the attribution chain matters. A connector that runs under each member's own Google sign-in produces API changes attributed to that person, which means "the AI changed it" is never the end of the trail: it is this person's assistant, this change, this timestamp. That is the property AdCopilot's per-member design is built around, and it is what keeps a team's accountability model intact when assistants join it.
One honest limit: change history records writes only. What was read — which reports an AI pulled, what it looked at and was refused — never appears in Google's log. That other half of the story is the connector's audit trail to tell, and the two ledgers together are the full account of AI activity. The trust architecture around both lives on the security page.