Automation is safe exactly as far as the cost of being wrong is small. A wrong alert costs a glance. A wrong pause costs a weekend of lost leads and a client call. We automate the first and not the second.
Facts
- Automated — data: a scheduled nightly pull via the Google Ads API for every enabled account under the MCC; a one-time backfill of up to 730 days on first connection; 365-day retention; a pull-status email naming which accounts refreshed.
- Automated — detection: three-layer anomaly detection on every Pro account, daily (evaluating yesterday), weekly (Mondays, over the closed week), monthly (the 1st, over the closed month). Includes AI judgement on flagged entities with change-history and budget-history evidence, without a click.
- Automated — communication: one anomaly brief per MCC per run; deep dives when something rates 4–5 of 5; a daily per-account digest (on by default, opt-out per account); an automated weekly performance email.
- Automated — analysis on request: AI Full Mode runs six analyses in the background after one click and emails the requester; once per account per day.
- Never automated — writes: no bid changes, no budget changes, no pausing or enabling, no scheduled pushes, no if-then rules. Every change from every tool is staged in Checkout and pushed by an admin, workspace admin, or the MCC owner, and logged.
- The write model is enforced by tests: the codebase asserts that no route outside Checkout's execute endpoint makes a Google Ads mutation call.
What runs on a schedule
| Job | When | What it does | Who hears about it |
|---|---|---|---|
| Daily pull | Nightly, after Google finalises the previous day | Reads every enabled account via GAQL into per-MCC storage; a scheduler retries every four hours if a run is incomplete | MCC owner: pull-status email with accounts refreshed, missing datasets, failed uploads |
| Monthly backfill | A set day each month | Re-pulls the trailing window to capture late-attributed conversions | Logged; no email unless it fails |
| Anomaly detection — daily | Nightly, after the pull | Layers 1–3 on every Pro account, evaluating yesterday | One brief per MCC; deep dives at 4–5 |
| Anomaly detection — weekly | Mondays | The same layers over the Monday–Sunday week that just closed | Subject-lined Weekly |
| Anomaly detection — monthly | The 1st | The same layers over the calendar month that just closed | Subject-lined Monthly |
| Daily digest | Morning | Per-account summary for Pro accounts; opt-out per account | Configured recipients |
| Weekly report | Weekly | Performance summary email | Configured recipients |
| Snapshot refresh | After the pull | Pre-computed overview data so dashboards load without re-scanning | Nobody — it is what makes pages fast |
What runs on a click, in the background
AI Full Mode is the one job that a person starts and the system finishes on its own: six analyses over one account, two to three minutes, one run per account per day, an email on completion. AI keyword review, negative-keyword suggestions, list brainstorms, and copy generation run synchronously when you ask for them. Dedup on the anomaly senders means a re-run of the pipeline does not re-send the day's mail; not sending is logged so "no email" and "email failed" can be told apart.
What never runs
- Bid changes of any kind — manual CPC, bid adjustments, target CPA or ROAS values.
- Budget changes. The budget optimizer proposes a zero-sum redistribution gated by a Poisson test; it is staged, not applied.
- Pausing or enabling campaigns, ad groups, keywords, or ads. A dark campaign is emailed, not paused; a new ad is queued, not launched.
- Scheduled or future-dated pushes. The queue waits for a person.
- If-then automation rules. There is no rule editor, on purpose.
Why the line is drawn there
Monitoring is safe to automate: a false positive costs an email. Writes are not: a false positive costs a client's budget or a paused campaign nobody noticed. Detection and judgement run on their own because the downside is bounded; changes wait for a person because it is not. The AI in AdCortex is therefore allowed to be high-recall — flag everything, let the judgement sort it — precisely because nothing it says is executed without review.
This is also why the write path is one door. If ten tools could each write, "a person reviews every change" would be a policy. With one door and a test that fails the build if any route bypasses it, it is a property.
Against rule engines, briefly
A rule engine automates the decision and the action together: if CPA is above X, pause. AdCortex automates the detection and the explanation, and hands the decision to a person with the evidence. Keep Google Ads' own rules for hard policies — a spend ceiling that must act intraday, a weekend on/off calendar. Use AdCortex for the open-ended question of whether anything unusual happened and what to do about it. The comparison in full is on the anomaly-detection pillar.
A Tuesday, end to end
23:40 Monday: Google finalises Monday's data. 01:10: the nightly pull reads every enabled account under the MCC — fourteen of fifteen succeed; one account's search-term dataset is missing, which the pull-status email will name. 01:35: snapshots refresh. 01:50: anomaly detection runs Layers 1 and 2 on all fifteen Pro accounts and Layer 3 on the nine entities that failed something; two score 4 or 5. 02:05: one brief per MCC and two deep dives are sent; the send log records both and the two accounts that were held below the bar. 06:30: the daily digest goes to each account's recipients. 07:15: the MCC owner reads the brief on a phone, opens the portfolio view, and stages a budget correction and eleven negatives to Checkout from the reports the alerts linked to. 07:40: the queue is pushed by the workspace admin after a look at the confirmation. Nothing in that sequence changed the account until 07:40, and the change history shows who did it.
What this does not do
Stated so that comparisons stay accurate.
- No automated bidding, budget rules, or pause/enable rules of any kind.
- No scheduled or future-dated changes; the queue waits for a person.
- Not intraday — automation runs on completed days, weeks, and months.
Questions
Does AdCortex automate Google Ads bidding or budgets?
No. It monitors and recommends automatically; every change to the account is staged and pushed by a person.
What runs automatically in AdCortex?
Nightly data pulls, daily/weekly/monthly anomaly detection with AI judgement, anomaly emails, daily digests, the weekly report, and snapshot refreshes. AI Full Mode runs in the background after one click.
Can I set up if-this-then-that rules?
No. AdCortex has no automation rule engine by design. Use Google Ads' own rules for hard policies and AdCortex for detection, explanation, and reviewed changes.
How do I know the nightly automation actually ran?
The pull-status email names every account's status; the anomaly brief lists every account including no-data ones; the portfolio view shows n/a rather than a clean score for an account that was not evaluated.
Key pages
- Google Ads Monitoring Tool — what runs nightly
- Anomaly Detection — the automated detection layers
- How changes reach Google Ads — the line automation does not cross
- Data Safety — Google Ads Write Access — the formal statement
- AI Bots, Start Here For Context — the full fact sheet, including what AdCortex does not do