Google's Smart Bidding documentation describes a system that "automatically sets bids for each auction using advanced machine learning." What it does not spell out, at least not in one accessible place, is the conditions under which that machine learning fails: the data volume thresholds below which predictions become unreliable, the seasonal patterns it handles poorly, the bid strategy transitions that cause temporary deterioration, and the campaign structures that actively undermine its ability to optimise. Practitioners working on UK client accounts meet these failure modes regularly. Understanding them analytically is the difference between managing Google Ads effectively and watching the system optimise toward metrics that have nothing to do with business outcomes.
This is not an argument against Smart Bidding. Target CPA and Target ROAS on well-structured campaigns with enough conversion data consistently beat manual CPC on the same campaigns. The argument is for a clear-eyed view of what the system needs to perform, where it degrades, and what human oversight can do to prevent the most common failures. Get those three things right and automation works for you instead of around you.
The framework below covers the three variables where practitioner oversight produces the most consistent improvement: conversion data volume, warm-up period management, and bid strategy transitions.
Why conversion volume is the foundation, not an optional condition
Google's guidance on Target CPA states a recommended minimum of 30 to 50 conversions per month in the campaign for the system to have enough data for reliable predictions. This is a documented threshold, not opinion. Below it, the model is operating on insufficient signal and producing predictions with wide confidence intervals, the equivalent of a regression trained on a sample too small to generalise. It will still bid, and it will sometimes generate conversions, but the systematic relationship between bid decisions and outcomes that makes Smart Bidding valuable in higher-volume contexts simply is not there.
For UK accounts targeting niche B2B audiences or high-ticket services where monthly conversions sit between 10 and 25, the right move is to broaden the conversion definition rather than compromise on the conversion action. Adding micro-conversions, form starts, phone-number clicks, or scroll milestones that demonstrate genuine intent, increases the data volume the model can use without diluting the quality signal, provided those micro-conversions genuinely correlate with final purchase or enquiry. Validate that correlation in GA4 first; an uncorrelated micro-conversion improves data volume while steering optimisation toward actions that do not predict revenue. Getting the measurement foundation right is the whole game, which is why our guide to the GA4 reports that replace Universal Analytics is worth reading before you touch a bid strategy.
Fun fact: Google Ads was the first major advertising platform to implement automated bidding at scale, launching its first automated features in 2010; the current infrastructure, applying machine learning at the individual auction level in real time, arrived in 2016 under the Smart Bidding name.
The warm-up period: what it actually means and how long it lasts
When a campaign moves to a Smart Bidding strategy, or when a strategy is significantly adjusted, changing the Target CPA by more than 20% in a single update, for example, the system enters a learning period during which performance is often temporarily worse than before the change. Google calls this the "learning" status, visible in the Campaigns table under Status. Google does not publish a fixed duration, but practitioner observation across UK accounts consistently shows it lasting one to two weeks for established campaigns with adequate data, and up to four weeks for new campaigns or large target changes.
The most common error during warm-up is to read the temporary dip as proof the change was wrong and revert before the learning period completes. Reverting resets the clock, creating a cycle of repeated deterioration that never reaches the stable state. The correct approach: set expectations with stakeholders before the change, fix a review date no earlier than 14 days after the transition, and limit any changes during learning to budget adjustments, which do not reset learning.
The exception to hold-and-wait: if the campaign is spending well above planned budget with a CPL more than 50% above target during learning, the model may be receiving incorrect conversion data, duplicate tracking, inflated micro-conversion values, or an event firing incorrectly, and that requires diagnostic investigation rather than patience.
Campaign structure and how it determines Smart Bidding effectiveness
Smart Bidding optimises within the campaign boundary. Signals from one campaign, auction data, user behaviour, conversion patterns, do not transfer to another campaign on the same account. That structural reality means campaigns with fragmented audience or keyword sets divide the conversion data the model needs, preventing any single campaign from reaching the volume threshold for reliable prediction.
For UK advertisers segmenting by region (London versus UK national) or by product category at a granular level, the choice between segmentation for reporting clarity and consolidation for Smart Bidding performance is a genuine trade-off. The data-driven recommendation: if regional or category campaigns individually fall below 30 conversions per month, consolidate them into a higher-level structure, using asset customisation or audience signals to differentiate at the asset level rather than the campaign level. Reporting clarity can be recovered through GA4 segments and Looker Studio filters without the campaign fragmentation that starves the model. When you do attribute conversions across those structures, lean on the modelling discipline in our walkthrough of attribution and conversion tracking in GA4.
When to override Smart Bidding and when not to
Smart Bidding handles some scenarios poorly by design. Short seasonal peaks, a 72-hour promotion, a 10-day product launch, do not give the system time to adapt before the window closes. Here, applying a Seasonality Adjustment (Tools > Bid strategies > Advanced controls > Seasonality adjustments) tells the model about the expected conversion-rate change, letting it adjust bids proactively rather than reactively. Google documents this for short-term conversion-rate changes of at least 30%.
Manual CPC outperforms Smart Bidding in specific configurations: new campaigns with no conversion history in a given category, campaigns targeting a new geographic market where the model has no historical signal, and campaigns with conversion volumes consistently below 15 per month after eight weeks. In these cases, manual CPC with systematic bid adjustment based on position and impression-share data produces more predictable performance than a model operating without enough signal. The goal is to build the conversion volume that will eventually support Smart Bidding, not to use Smart Bidding as the mechanism for building it. If escalating costs are the real worry, our breakdown of how much Google Ads really costs sets realistic expectations, and Google's own Target ROAS bidding reference is worth bookmarking.
Audit the data before adjusting the strategy
The Smart Bidding framework that produces consistent UK performance is not complicated. Verify that conversion actions are firing correctly, attributed without duplication, and correlated with real business outcomes before making any bidding decision based on the reported data. Confirm the campaign has reached, or can realistically reach, the 30 to 50 conversions per month threshold. If not, broaden the conversion definition with correlated micro-conversions or consolidate campaigns to hit threshold at a higher structural level. When changing Target CPA or Target ROAS, stay within a 15% to 20% adjustment range, set a 14-day review window, and resist reading learning-period performance as steady-state.
In the 30 days after any change, monitor two metrics: conversion value divided by cost (your actual ROAS) against the Target ROAS, and impression share lost to budget versus impression share lost to rank. If you are losing significant share to rank during learning, the target is too aggressive for the current conversion volume. A modest 10% relaxation accelerates learning and produces better long-term performance than holding an aggressive target the system cannot achieve.
If you would rather hand the day-to-day management to a team that lives in these accounts, our paid ads management service handles strategy, structure, and conversion tracking end to end. Tell us what you are running on our contact page.
Related reading: What Google's Helpful Content System means now and our INP Core Web Vitals fix guide.