Automated bidding sets every bid, while you define what it should buy. Choose the conversion event, value, target, budget ceiling, and exclusion list. Those five settings steer your ad budget.
Set boundaries before the auction, then give the system enough time to learn. Many accounts reverse that order. They leave inputs on defaults and intervene every week.
What are AI bidding guardrails?
Humans set five constraints before automated bidding starts: conversion event, value, efficiency target, budget ceiling, and blocked inventory. The algorithm then optimizes within those limits.
Five inputs carry almost all of the weight:
- The conversion event. What the model treats as a win.
- The conversion value. How much each win is worth relative to the others.
- The efficiency target. Target CPA, target ROAS, cost cap, or bid cap.
- The budget ceiling. Daily and campaign-level spend limits.
- The exclusion set. Placements, audiences, brand terms, and geographies the campaign may not buy.
The platform owns auction-level bids. Human judgment still controls the five inputs above. Those inputs determine what the account buys.
Why does Google’s August 17 change make your target matter more?
Starting August 17, 2026, Google will pull budget-limited Target CPA and Target ROAS campaigns closer to their stated targets. A campaign set to a $10 target CPA while delivering at $5 may drift toward $10. A loosely set ceiling can become the performance you get.
Many accounts set a target CPA well above real performance and treat it as a safety ceiling. They expect the algorithm to keep finding cheaper conversions. On budget-constrained campaigns, that assumption changes this month.
Google has pushed back on the wider reading of the change. According to PPC Land’s reporting, Google’s Ginny Marvin said the update “will not lead to spend increases.”
Daily and monthly budget limits still apply. Efficiency may move even if total spend does not.
Google announced the change alongside a broader bidding package. Search Engine Land reported that Google expanded Smart Bidding Exploration to Performance Max and Shopping. Google also added promotion mode for demand spikes. Exploration pursues new queries within your ROAS tolerance, which becomes another guardrail.
Review every campaign flagged “Limited by budget” and compare its stated target with the last 30 days of delivery. Where the two are far apart, reset the target to what the campaign already produces.
Which bidding decisions belong to a human?
Split the account by who is better at each call. The algorithm handles high-speed auction decisions. Humans handle business knowledge that never enters the platform, like what a customer is worth after month three.
| Decision | Owner | Why |
|---|---|---|
| Bid per auction | Algorithm | Millions of decisions per day against signals you cannot see |
| Which conversion counts | Human | Only you know which event maps to revenue |
| What a conversion is worth | Human | Margin and retention data live outside the platform |
| Efficiency target | Human | Set from unit economics, not from last week's CPA |
| Audience and placement mix | Algorithm | Adjusts delivery continuously from live performance signals |
| Exclusions and brand safety | Human | Reputational limits are not a performance metric |
| Budget between campaigns | Human | Portfolio calls sit above what any one campaign sees |
| Budget inside a campaign | Algorithm | Reallocates continuously within the configured budget |
The human-owned decisions are set before the auction. None requires daily manual bidding.
Not sure which of those rows is set correctly in your account? We audit bid strategy configuration as part of every paid media engagement, before touching spend. Book a Free Strategy Call and we will walk through your account structure with you.
How do you set a target that works as a guardrail?
Set it from your economics, then sanity-check it against delivery. A target derived from margin tells the system what you can afford. A target copied from last month’s CPA only tells it to repeat last month.
On Google, choose a target CPA or ROAS you would accept at scale. Aspirational targets can throttle delivery. After August 17, loose targets may pull performance toward the looser number.
Meta gives you two shapes of the same guardrail, and they behave differently:
- Cost cap. Holds average cost near your number while volume stays open. Common practice, per TheOptimizer’s 2026 bidding breakdown, is to set it 10% to 20% above your true target CPA so delivery does not choke.
- Bid cap. A hard ceiling on any single auction. Tighter control, slower spend, and a real risk of underdelivery if the number is low.
Cost cap fits most accounts that prioritize average efficiency and delivery. Use bid cap when one expensive conversion would cause real damage. A high-ticket B2B account is one example, where a single bad lead can burn a week of sales capacity.
Inside Advantage+ Shopping, manual bid control is limited by design. You supply a ROAS goal as a guide and the rest of your guardrails move to the signal and exclusion layers. If you are still deciding how much automation to hand over, our comparison of Advantage+ and manual campaigns covers that trade.
What stops the algorithm from buying the wrong conversions?
Clean conversion data stops the algorithm from buying the wrong conversions. It sits upstream of every target you set. An automated strategy can hit a $40 target CPA while optimizing for the wrong event. The number looks efficient even when the outcome is useless.
Three checks hold this together:
- Event depth. Optimize toward the deepest event with enough weekly volume to learn from. A form-fill target produces form fillers. We covered the full version of this in why Google and Meta keep finding the wrong customers.
- Value accuracy. If a $900 order and a $40 order both register as one conversion, the model treats them as equal wins.
- Tracking uptime. A broken pixel may not show an obvious error. It can quietly teach the model that your best days produced nothing.
Google offers a specific repair for data corrupted by tracking problems. Data exclusions tell Smart Bidding to ignore a past data window. Use them after a tracking outage or checkout failure.
Google’s API documentation separates data exclusions from seasonality adjustments. Seasonality points forward and fits short events of roughly 1 to 7 days. Do not use it to patch bad historical data.
When should you step in, and when should you leave it alone?
Intervene when tracking breaks or economics change. Ignore three bad days unless a hard failure accompanies them. Automated bidding needs a stable read window, and frequent significant edits can restart the learning period.
| What you are seeing | Step in? | Action |
|---|---|---|
| Conversions stopped recording | Yes, same day | Fix tracking, then apply a data exclusion for the outage window |
| Campaign spending with zero conversions | Yes | Pause and diagnose signal before touching the target |
| CPA up 15% over three days | No | Often normal variance; wait for a longer read |
| CPA up 15% over three weeks | Yes | Review target and creative supply together |
| Margin or pricing changed | Yes | Recalculate the target from new unit economics |
| A placement showing up you dislike | Yes | Add to exclusions, do not change the bid strategy |
| Reported ROAS looks too good | Yes, but test it | Run a holdout before scaling |
| Competitor launched a campaign | No, not by itself | Let the auction adjust; act only if your economics change |
That “too good to be true” row deserves attention. Automated campaigns often find the cheapest available conversions, including buyers who were already likely to purchase.
The gap between reported and incremental ROAS can be widest in campaigns you want to scale. Include incrementality testing in the guardrail review instead of waiting for quarterly reporting.
What does a weekly guardrail review look like?
Use a fixed list in the same weekly slot. Consistency matters more than depth.
- Tracking check. Did every conversion action fire every day this week?
- Target versus delivery. Which campaigns are running far from their stated target, and in which direction?
- Budget-limited flags. Which campaigns are newly capped by budget, and how does that change target behavior?
- Exclusion drift. New placements, new search terms, new automatically-created assets you did not approve.
- One portfolio decision. At The Remarkable, we keep any one budget move under 20%. Our budget reallocation cadence explains why the step size matters more than the frequency.
In our experience, 15 minutes covers this review in most accounts. Daily target changes can keep campaigns in a learning state and make results harder to read.
Teams get more from automated bidding when they manage fewer settings. They keep conversion data clean and set credible values and margin-based targets. Then they protect exclusions, leave bidding alone, and focus on creative.
If your bid strategies use default settings and instinctive reviews, our paid media service rebuilds the signal and guardrail layer before scaling spend. Book a Free Strategy Call and we will review your current setup.
Like this? Get the next one.
Short emails. New posts as they ship.
