Custom Rules Engine — Complete Guide

Where Custom Rules Fits Within Autopilot

First, get the framing right: Custom Rules is a capability inside Autopilot — not a separate product, and not equal to Autopilot itself.

Autopilot's default mode is AI Inspection — it proactively diagnoses your account, gives evidence-backed optimization suggestions, and executes after you confirm. Custom Rules adds a layer of deterministic execution on top: take the repetitive operations you've already thought through (stop-loss, scaling, alerts), write them as fixed rules, and let them run on schedule — no need to wait for an AI inspection each time.

Think of it as a transitional capability on Autopilot's path toward "more automatic" — AI Inspection handles "surfacing what you didn't think of," Custom Rules handles "auto-executing what you've already decided." They work together; it's not either-or, and neither replaces the other:

Can both run simultaneously? Yes — they are two parts of the same Autopilot, covering the same ads without interfering with each other.

What happens when the same ad is covered by both? Custom Rules run on schedule and auto-execute at each slot. AI Inspection only evaluates when you manually trigger it. If Custom Rules already paused an Ad Set, AI Inspection will read its current status as paused and won't suggest pausing it again. There's no priority conflict — whichever runs first takes effect, and the later one works with the latest state.

---

Safety Mechanisms

Pausing / Disabling / Deleting a Rule

↑ The rule list on the Autopilot detail page — each rule shows a status badge (Active / Paused / Error) with pause / resume / delete on the right

Does it Respect "Learning Phase Protection" and "Budget Caps"?

• Learning phase protection: Custom Rules V1 does not have built-in platform learning phase detection. If you don't want rules to act on ads still in learning, add a creation time condition (e.g., since_creation window with a minimum days threshold), or rely on AI Inspection's learning phase protection.
• Budget caps: The adjust_budget action supports a cap parameter. Once set, no matter how many times the rule fires, the budget will never exceed your cap. The system will suggest adding a cap during creation but won't force it.

Can I Roll Back a Misfire?

One-click rollback is not supported yet. If a rule paused an Ad, you'll need to manually resume it (or create a reverse rule). Best practice: start new rules with notify only for a few days, confirm the hit logic is correct, then switch to auto-execution.

"Notify First, Auto-Execute Later" — The Transition Approach

Fully supported. Set the action to notify when creating. Matched results are recorded in the action logs, viewable in your inspection report. After observing for a period and confirming the logic is sound, edit the rule to change the action to pause or adjust_budget — no need to delete and recreate.

---

Operations × Platforms × Levels Matrix

Full Support Table

Where Does adjust_budget Land Under CBO / ABO?

• Meta CBO enabled: Budget lives at the Campaign level. Adjusting Ad Set budget has no effect (auto-skipped at execution).
• Meta ABO (non-CBO): Budget lives at the Ad Set level. Adjusting Campaign budget has no effect.
• Best practice: If your account mixes CBO and ABO, set both Campaign and Ad Set items in the same rule. The system auto-determines which level is effective and skips the inapplicable one. You don't need to check whether each Campaign is CBO or ABO first.

What Exactly Does Google Ads Support?

Google Ads fully supports pause / resume / adjust_budget (Campaign level) / adjust_bid_target (Campaign and Ad Group levels) / notify. Ad-level pause/resume is not supported yet. Google has no ad group budget concept, so adjust_budget only applies at the Campaign level.

---

Condition Capabilities — Full Reference

These metrics, time windows, and operators ultimately combine into a rule's condition in the UI, like this:

↑ What a rule looks like in the UI — metrics (spend / conversions_registration) + operators + time window (yesterday) + action (notify) combine into a condition

Available Metrics

The following are commonly used metrics (not a closed whitelist — the system continues to expand):

For conversion-related metrics, we recommend specifying the exact event: cost_per_install, cost_per_purchase, cost_per_add_to_cart, etc. When you say "CPA," the system will ask: "Which event's cost? Install / purchase / add-to-cart?" to help you target precisely.

Time Windows

Operators

Condition Logic (AND / OR)

• AND (all must be met): Wrap multiple conditions in all logic.
• OR (any one triggers): Wrap multiple conditions in any logic.
• Nested combinations supported: e.g., "(spend > $50 AND CPA > $15) OR (spend > $100 AND conversions = 0)".

Consecutive Days / Consecutive Hours

• consecutive_days: N (N ≤ 30): Each of the past N days must individually satisfy the condition.
• consecutive_hours: N (N ≤ 24): Each of the past N hours must individually satisfy the condition.

Difference from last_Nd: last_7d checks whether the 7-day aggregate exceeds the threshold; consecutive_days: 7 checks whether each individual day exceeds it — the latter is stricter, suitable for "sustained poor performance before acting" scenarios.

---

Scheduling

Supported Frequencies

Minimum granularity is 1 hour, on the hour only. If you ask for "every 30 minutes," the system will inform you the minimum is 1 hour and adjust accordingly.

One-time schedule (run once only): Not natively supported yet. Workaround: create the rule, let it run once, then manually pause it.

How Is Timezone Determined?

The schedule's timezone is specified when you create the rule. The system defaults to your account timezone, which you can override. For example, if your account timezone is America/Los_Angeles, setting "daily 09:00" means 9 AM Pacific Time.

How Are yesterday / today Calculated?

All time windows are based on account timezone. yesterday = previous day 00:00–23:59 in account timezone. today = current day 00:00 to query time in account timezone. The system handles timezone conversion automatically.

---

Scope

Rules can precisely define the target scope:

You don't need to target "everything" — you can scope down to "Account A, objective contains sales, name contains 'US-', Ad Set level."

---

Multi-Rule Conflict Handling

At Creation: Conflict Detection

When you create a new rule, the system automatically compares it against existing rules on the same channel:

• Conflicting operations (one wants to pause, another wants to increase budget, with overlapping scope) → System prompts you to confirm: intentional separation of duties, or should you add mutual exclusion conditions?
• Duplicate / mergeable rules (highly similar conditions, same-direction operations) → Suggests merging into one to avoid stacking effects (e.g., two rules each reducing by 20% → effective reduction ≈ −36%).

These are advisory prompts only — they do not block activation. You can dismiss them and proceed.

At Runtime: Same-Cycle Conflict

If the same ad is matched by two rules in the same execution cycle with mutually exclusive operations (one stops, one scales):

• Conflicting operations are both skipped — neither executes.
• Conflict details are recorded in the inspection report's anomaly section.
• You need to adjust rule logic to eliminate the conflict.

There is no priority or weight mechanism currently — conflicts default to conservative inaction. Recommendation: use scope_filter to isolate ranges between different rules (e.g., one manages "ROAS > 2 winners," another manages "CPA > $30 losers" — naturally mutually exclusive).

---

Creation Flow: Complete Walkthrough

Here's a real example demonstrating the full process.

Your request: "In my Meta accounts, for Ad Sets that spent over $50 but CPA is above $15, automatically reduce budget by 20% daily, with a daily budget floor of $30. Run every day at 10 AM."

Step 1: Natural Language Description

Just tell GrowthGPT the above in plain language. Chinese or English both work. No code required, no need to know field names.

Step 2: Intent Parsing

The system automatically extracts 7 dimensions:

If there's ambiguity, the system asks once. For example, if you say "CPA," it will ask: "Which event's cost? Install / purchase / add-to-cart? Or just the optimization goal's cost per result?"

Step 3: Feasibility Check

The system auto-validates:

• Does Meta Ad Set level support adjust_budget? → ✅
• Is cost_per_result available on Meta? → ✅
• Does −20% adjustment exceed the ±50% advisory limit? → No ✅
• Pure ratio rule missing a base threshold? → Already has spend > $50, ✅ no additional prompt needed

Step 4: Rule Card Draft

The system outputs a text-based rule card:

Rule name: High CPA Budget Reduction Channel: Meta Schedule: Daily 10:00 Scope: All accounts · Ad Set · status = ACTIVE Conditions: spend > $50 yesterday AND cost_per_result > $15 yesterday Actions: Reduce daily budget by 20%, floor $30

⚠️ Note: With continuous budget reduction, once the budget hits the $30 floor, no further adjustments will be made.

At this point the system pauses and waits for you to confirm the draft — verify conditions, actions, and scope are correct before proceeding.

↑ The Rule Card GrowthGPT outputs in chat: rule name, channel, schedule, conditions, actions, plus the protection note at the bottom

Step 5: Dry Run

After you confirm the draft, the system runs a simulation (no platform changes) — applies your real account data against the conditions to see how many Ad Sets would be matched and what percentage of total spend they represent.

Possible outcomes:

Normal: "Matched 3 Ad Sets, representing 18% of yesterday's spend, all meeting spend > $50 and cost_per_result > $15." → Proceeds to confirmation card.

Dry Run incomplete: Some account data didn't return (common reasons: account just authorized, no spend yesterday, platform API delay). The system marks it as Dry Run incomplete and doesn't show match counts. You can choose:

• Activate anyway — accounts with missing data will auto-skip during actual execution (circuit breaker), no misfire risk.
• Wait a day — retry Dry Run after data normalizes.

↑ Dry Run feedback: matched Ad Set count and share of yesterday's spend; incomplete data is flagged

Step 6: Confirm & Activate

The system displays a confirmation card with the final rule summary (rule name + schedule + scope + conditions → actions).

• Click Active → Rule is saved and goes live, awaiting the next scheduled trigger.
• Click Cancel → Not created, return to chat to continue editing.

↑ Confirmation card: rule name + channel badge on top, conditions → actions in the middle; only clicking Active activates it

---

After Activation: Where to See Results

You can also ask directly in chat: "Which ads did the 'High CPA Budget Reduction' rule match today?" GrowthGPT will pull up the execution logs for you.

---

adjust_budget / adjust_bid_target — Adjustment Rules

adjust_budget

adjust_bid_target

---

FAQ

Q: How is Custom Rules different from Meta / TikTok / Google's built-in automated rules?

In short: if you only run one platform with simple rules — built-in rules are fine. If you run multiple platforms, want natural-language rule creation, want Dry Run and conflict protection — Custom Rules is the better fit.

Q: How does it compare to third-party SaaS (Revealbot, Madgicx, Smartly)?

Q: What are the capability boundaries — what can't Custom Rules do?

Q: How many rules can I create per channel?

Maximum 10 rules per channel. If you hit the limit, consider merging similar rules — one rule can contain multiple "condition → action" mappings.

Q: How many actions can a single rule have?

3 execution actions maximum. The system suggests splitting into multiple rules if exceeded.

Q: What happens with consecutive failures?

If the same rule errors 3 consecutive times → auto-paused, status changes to Error, system notifies you. Investigate the issue and manually resume when ready.

Q: Will it misfire if data is unavailable (no spend yesterday, newly authorized account)?

No. Circuit breaker mechanism — when retrieved data is empty (especially spend), the rule skips entirely for that cycle with no actions taken, recorded as "data missing, skipped" in the inspection report.

Q: How are ratio metrics (e.g., CPA) calculated when there are zero conversions?

• Spent money but conversions = 0 → CPA = +∞ (infinity), will match "CPA > XX" conditions. This is intentional: spending with zero results → should stop loss.
• No spend and no conversions (0/0) → Entity is skipped, no match, no action.
• Recommendation: add a base threshold to pure-ratio rules (e.g., spend > $30) to avoid premature kills on minimal spend. The system proactively reminds you during creation.