A reader selling outdoor gear sent over an account last month asking why it got worse the more they optimized. For six months the team adjusted bids daily, added negative keywords weekly, and installed two scripts, while CPA climbed. Two problems were doing the damage. Conversion actions were double counting, booking one order twice, so reported CPA had been too low all along. And three main ads had been disapproved because their landing pages would not open, cutting traffic off weeks earlier. The team had been optimizing a broken dataset.

This kind of misdiagnosis happens daily. The order of diagnosis has to be fixed. Confirm tracking is trustworthy, confirm ads and products are eligible, and only then discuss search terms, budgets, bids, creative, and landing pages. Run it backwards and effort compounds the error. This piece strings together a daily, weekly, and monthly cadence, an experiment framework, and automation governance. For the overall structure of the optimization layer, see the optimization guide.

Two Checks Before Any Optimization

Tracking is the foundation. Whether conversions are double counted or missing decides whether every ratio is real. Check three places, the conversion action setup (counting method, attribution window, cross-device inclusion), whether backend orders reconcile with conversions reported in the ads account, and whether consent settings are blocking a share of conversions. When the two sides diverge badly, do not touch bids. Fix the data first. The full treatment of tracking and attribution lives in this piece.

Eligibility is the second gate. Ad review status, policy disapprovals, rejected feed items, landing pages that fail to load, expired payment methods, budgets that run out early. Any one of them shows up as a sudden performance drop. These are delivery interruptions and do not belong on an optimization list. The daily check starts with status precisely to keep this class of problem away from optimization work.

Daily Looks for Breaks, Weekly for Drift, Monthly for Direction

The three cadences divide labor clearly. Daily asks whether anything broke, weekly asks whether anything drifted, monthly asks whether the direction is right. Each level needs an explicit answer for what to watch, what it triggers, and what it must not trigger.

The daily check covers four things. Spend and burn rate, whether budget is gone by midday. Impressions and clicks, whether either fell off a cliff. Conversions, whether they dropped to zero. Review and policy alerts plus script run status. Triggering these means checking tracking, eligibility, and delivery interruptions. What the daily check must not trigger matters just as much. One bad day is not a reason to change bids, and the daily check does not make structural decisions.

The weekly check covers five areas. Search terms that newly appear with high spend and low conversion, negative them out. CPA and conversion trends, read against last week and a four-week average. Whether budget is chronically limited, with the share of impression share lost to budget as a usable signal. Whether the bidding strategy is still learning. Which is fading first, creative or landing page. Audience and frequency also belong here, while list and exclusion design lives in that piece, so this level only checks for anomalies. The weekly check triggers negatives, small budget adjustments, and creative rotation. It does not trigger strategy switches or large rebuilds.

The monthly check reads structure. Whether campaigns and ad groups are too fragmented, whether each group can accumulate conversions. Whether bidding strategy and targets still match the business stage. Whether last month's experiment results should ship. Whether the cost structure moved, including the split between people and tools, which the cost and budget piece covers. Merges, splits, and strategy switches belong here.

On search terms and negatives, the search ads guide already gave the execution order, check search terms and negatives first, then landing pages, then bids, changing one variable class at a time on a weekly cycle. This piece expands that sentence into three cadences. The two do not conflict. One is the order of a single pass, the other is the calendar for running it continuously.

When Bids Deserve a Change

High CPA is not a reason to change bids. It is a symptom. Rule out three things first, whether tracking is trustworthy, whether eligibility is intact, and whether budget is chronically limited. With all three cleared, check a fourth question, whether the bidding strategy is still in its learning period. Changing targets or budgets during learning restarts the process.

Even a justified change should move in small steps. Change a target, change a budget, switch a strategy, one at a time, with an observation window after each. How bidding mechanics actually work is covered in the bidding and budget piece. This piece only answers when to touch them.

Experiment Design Starts with a Written Hypothesis

Write the hypothesis before opening an experiment. What changes, which metric should move, how much movement counts as success, and what would count as failure. If those four sentences cannot be written, the thinking is not done, and the experiment should wait.

Google Ads experiments run a parallel version on a portion of the original campaign's traffic and budget for a specified length of time, then compare performance. The platform judges results with confidence intervals and significance, using bucketing and a 95 percent confidence interval, and the monitoring page states whether results are statistically significant or whether there is no clear winner. The second outcome is normal. It means the change produced no discernible difference, not that the experiment failed. How Performance Max experiments are built and where their boundaries sit is covered in this piece.

Four design decisions matter.

One variable. Change one thing at a time. If the creative changes, the bid does not. If the landing page changes, the audience does not. Mixed variables win without telling you why.

Sample by conversions, not impressions. Impressions can pile up while conversions stay too few for a conclusion. Estimate weekly conversions before opening the experiment. If the number is too small, consolidate structure first or run the test in a larger campaign. Search and Performance Max experiments expose an official Experiment Power score that estimates the likelihood of reaching statistical significance. Check it before launching rather than after.

Duration covers a full business cycle. Cross-border categories often take days from click to order, so run at least a week, ideally across two full weekly patterns, before concluding. Short experiments produce noise.

Early stopping is reserved for technical failure. Broken tracking, disapproved ads, exhausted budgets, those justify stopping. Ugly numbers do not. Stopping after two days turns the experiment into a mood swing.

Read intervals and direction. A good point estimate with an interval crossing zero means the result is not there yet. A narrowing interval with a stable direction is what earns a rollout. For how a PMax versus Shopping comparison sets its sample definition, this case has the full example, and the numbers are not repeated here.

Automation Belongs Under Governance

Scripts and automated rules save a lot of repetitive work. They should not make judgments for the team. A sane division leaves scripts responsible for repeated checks, data pulls, reports, and alerts, while automatic actions are graded by risk. Low-risk actions can run on their own. High-risk actions go through human review.

Six governance items.

Least privilege. A script that only reads never gets write access. A script that runs in one account never gets manager account access. Bigger permissions mean a bigger blast radius when something goes wrong.

Preview first. Google Ads scripts offer a preview mode that simulates changes to campaign data without modifying it, affecting only calls made through AdsApp while other services run normally. Run every new script in preview, read the change list, and only then go live.

Leave a trail. Change history records two years of account changes, including interface edits, automated rules, API calls, and scripts, with the user column showing the operator or system name. Per-entry detail is limited, and the official guidance suggests keeping each change small, ideally under two thousand entities. Scripts should also keep their own logs, recording what ran, what changed, and what came back.

Make it reversible. Change history supports undoing most recent changes, and automated rules can be undone from version history with a preview first. Before any automation goes live, confirm it can be reverted, and how long that takes.

Human review gates. Budget changes, bidding strategy switches, bulk pauses, and bulk negative keyword additions should all pass review. The reviewer does not need to read code, only the change list and expected impact.

Alert tiers. P1 handled immediately, such as conversions at zero, account disapproval, or payment failure. P2 same day, such as spend anomalies or script errors. P3 into the weekly report, such as creative fatigue trends. Write the tiers into alert rules so the team knows when a midnight call is warranted.

Two official details belong in the runbook. Automated rules have a standard turnaround of within two hours after conditions trigger, not on the dot. And when multiple rules target the same entities, the system does not order them, which can produce conflicts, so the official advice is to stagger their run times. On the scripts side, when an execution times out, changes made before the timeout still apply, so bulk operations should be batched and idempotent rather than assumed to finish in one pass.

One Operating Table

Putting the cadence on a calendar beats documenting it. The daily check takes the campaign owner ten to fifteen minutes. The weekly check brings media and content together for half an hour. The monthly review pulls in the business owner for an hour. Every check leaves one line, what was reviewed, what was found, what was done, who reviewed it. Run that table for three months and account stability looks different.

Common Misconceptions

  • Adjusting bids on broken tracking, drifting further with every change
  • Reading a disapproval as a performance dip instead of a delivery stop
  • Making structural calls in the daily check and switching bidding strategies in the weekly one
  • Stopping an experiment after two days of ugly numbers
  • Granting scripts full permissions with no logs and no rollback plan