AI in PPC: Smart Bidding, Performance Max and How to Set Guardrails

AI in PPC means the automated systems inside ad platforms that set bids, choose where ads appear, combine your creative and, increasingly, find new search queries for you. They can improve results when they are fed accurate conversion data and sensible goals. They waste money when tracking is wrong, targets are unrealistic or nobody checks what they are doing. Your job shifts from adjusting bids to supplying good inputs and setting guardrails.

Google, Meta and other platforms have moved most campaign mechanics into machine learning. Smart Bidding, Performance Max and AI Max for Search on Google, and Advantage+ campaigns on Meta, make decisions at a speed and scale humans cannot match. But they are not magic, and they are far from transparent. This guide covers the PPC fundamentals separately in our PPC advertising guide. Here we focus only on the AI layer: how automated bidding works, what Performance Max and AI Max do, which inputs matter, how to set guardrails, how to test and what the limits are. We rely on Google's and Meta's own documentation, which changes often, so check the current help pages. For Meta specifically, see our Meta ads strategy guide.

Smart Bidding: automated bidding in plain words

Google's help documentation describes Smart Bidding as a set of automated bid strategies that uses Google AI to optimise for conversions or conversion value in each auction, a feature it calls auction-time bidding. It takes account of a wide range of signals at the moment of the auction, including device, location, time of day, language and operating system. Google lists five Smart Bidding strategies, including Target CPA, which aims for conversions at a target cost per action, and Target ROAS, which aims for conversion value at a target return on ad spend. Google notes that for conversion-focused bidding you must make sure conversion tracking is set up correctly and the conversion actions you want to optimise for are set as Primary. Source: Google Ads Help, Determine a bid strategy based on your goals.

What you control	What the system controls
The conversion goals and their values	The bid in each individual auction
The target cost per action or return on ad spend	How it weighs device, location, time and other signals
Budget	How it spends within that budget during the day
Which keywords, audiences and ads are available	Which combinations it shows in each auction

The system can only optimise for what you tell it counts. If the conversion is "any form click" and many clicks are junk, it will find more junk cheaply.

Performance Max and AI Max

Performance Max

Google describes Performance Max as a goal-based campaign type that lets you access all Google Ads inventory, including YouTube, Display, Search, Discover, Gmail and Maps, from a single campaign. It combines the assets you provide, such as images, text, video and product feeds, with Google AI to serve across that inventory and optimise toward your goals, using audience signals to guide learning. It is meant to complement keyword-based Search campaigns, and includes generative AI tools to help create assets. Source: Google Ads Help, About Performance Max campaigns.

AI Max for Search

Google announced AI Max for Search campaigns at Google Marketing Live 2025 as an AI-powered option within Search campaigns, and in 2026 described AI Max as its primary AI-powered campaign solution. In general terms it extends Search campaigns beyond your keyword list by matching ads to relevant searches and tailoring text and landing pages. Because features and names change often, read the current Google Ads documentation before relying on any detail here.

Meta Advantage+

Meta offers Advantage+ campaign types that automate targeting, placements and creative combinations, particularly for sales campaigns. Meta's learning system needs data to settle, which we explain in our Meta ads guide.

The trade-off

Automation can find customers you would not target by hand and saves management time. In return you get less visibility into where ads ran, which queries triggered them and why. The more automated the campaign type, the more important your inputs and your checks become.

The inputs that matter most

Input	Why it matters	What to do
Conversion tracking	It is the learning signal. Wrong data trains wrong behaviour	Track real outcomes once each. Mark the right actions as primary. Test after site changes
Conversion values	Value-based bidding needs real values to prioritise high-value customers	Pass revenue or profit. Import offline results such as qualified leads and closed deals
Enough data	Systems learn from volume	Avoid splitting budget across many tiny campaigns. Consolidate where sensible
Realistic targets	An unrealistic target starves delivery or wastes spend	Start near recent actual performance and tighten gradually
Creative assets	The system can only combine what you give it	Provide varied, accurate headlines, descriptions, images and video. Review combinations
Audience signals	They guide early learning	Use customer lists, past converters and site visitors, with consent
Product or business data	Feeds and landing pages affect where and how you appear	Keep Merchant Center feeds and landing pages accurate and fast
Brand and exclusion settings	They protect against poor placements and branded search cannibalisation	Set brand exclusions, negative keywords where available and placement controls

Offline conversions: teach the system what a good lead is

For lead generation, an online form fill is not the goal. A closed sale is. If you only tell the platform about form fills, it optimises for people who fill in forms, including those who never buy. A better approach:

1. Record the source of each lead, using a click identifier and UTM parameters.
2. Track the lead through your CRM to qualified, quote and won stages.
3. Send back the stages that matter, such as "qualified lead" and "sale", with values, using the platform's offline conversion import.
4. Use the more meaningful action as the optimisation goal once there is enough volume.

This requires consent and privacy care, since you are sharing customer data. See our first-party data guide.

Guardrails for automated campaigns

• Budgets and caps. Set daily budgets you can afford and review after the first days. Avoid sudden large increases.
• Targets with room. Do not set the tightest target on day one.
• Brand safety. Use brand exclusions, negative keywords and placement exclusions where the campaign type supports them.
• Review the reports. Check search terms and insights, asset performance and where spend goes.
• Change slowly. Frequent edits restart learning. Make fewer, deliberate changes and note the date of each.
• Human approval for creative. Review automatically generated assets and text before they run. Ads must be true and compliant, whoever or whatever wrote them.
• Experiments. Use platform experiments or holdouts to compare automated campaigns with your existing setup.
• Alerts. Set rules or reports that flag spend jumps, conversion drops and tracking failures.

A testing plan

1. Baseline. Record current results: spend, conversions, cost per acquisition, return on ad spend and lead quality.
2. Change one thing. For example, move a campaign to Target CPA, or add a Performance Max campaign alongside Search.
3. Use an experiment so traffic is split fairly, instead of comparing before and after.
4. Run it long enough. Automated systems need weeks, and enough conversions, to settle. Meta states that learning stabilises with a minimum of about 50 optimisation events over a seven-day period, as a reference for how much data algorithms want.
5. Judge on business outcomes. Profit, qualified leads and revenue, not platform conversions alone.
6. Keep or revert. Decide, then document what you learned.

Where AI PPC struggles

• Low volume. With few conversions, systems have little to learn from, and results are erratic.
• Long sales cycles. If a sale takes months, early signals mislead unless you pass back later outcomes.
• Brand cannibalisation. Automated campaigns can take credit for conversions from people already searching your brand name, flattering the numbers.
• Opaque placements. Some campaign types give limited detail on where ads ran, so quality can be hard to audit.
• Misleading goals. Optimising for cheap leads can reduce quality.
• Creative sameness. If you provide weak or similar assets, you get weak combinations.
• Policy risk. Auto-generated text can include claims that breach policy or law.

The role of the PPC specialist now

Moving away from	Moving toward
Manual bid adjustments	Conversion tracking design and quality
Keyword list micromanagement	Account structure, signals and negative controls
Writing every ad variation	Creative strategy, asset quality and review
Reporting numbers	Interpreting results, testing and challenging platform claims
Single-platform focus	Cross-channel measurement and blended results

See our guides to attribution models and GA4 for measurement that does not rely on a single platform's report.

AI-generated ads and compliance

Platforms now generate headlines, descriptions, images and video from your inputs. You remain responsible for what runs. Check every claim is true and can be substantiated, that restricted categories are handled properly and that the content is not misleading. For endorsements and influencer-style content, label advertising clearly as UK, US and Indian regulators expect. Do not use AI to invent testimonials, reviews or customers. See our guide to authenticity in marketing.

Ads inside AI answers

PPC is also moving into AI answers. Google has announced ads in AI Overviews and ads formats in AI Mode, and OpenAI has begun testing ads in ChatGPT. We cover this in ads inside AI answers.

A 30-day checklist before you automate more

Week	Check
1	Audit conversion tracking and primary conversions. Fix double counting and junk conversions
2	Define break-even cost per acquisition or return on ad spend. Add conversion values. Set up offline imports if relevant
3	Prepare assets, feeds and audience signals. Set exclusions and caps
4	Launch an experiment against your current setup, and schedule a review for four to six weeks later

Common mistakes

• Turning on automated bidding with broken or junk conversion tracking.
• Setting an unrealistic target on day one.
• Splitting a small budget across many campaigns.
• Editing constantly, so the system never settles.
• Trusting platform-reported conversions over real sales.
• Letting automated campaigns absorb brand search credit.
• Running auto-generated creative without review.
• Leaving campaigns unattended.

A worked example: lead generation with offline conversions

This is an invented example. A Manchester solar installer spends £3,000 a month on Google Search ads. The campaign uses Target CPA bidding and reports 60 leads at £50 each, which looks good. But sales report that only about one in eight leads becomes a quote.

1. Diagnose. The conversion being optimised is a form submission. The system finds cheap submissions, including many from people outside the service area or with no intention to buy.
2. Record outcomes. The company adds a field to record the click identifier and source on each lead, and tracks leads in its CRM through to "site survey booked" and "contract signed".
3. Import offline conversions. It sends "survey booked" events back to the platform with a value, and uses them as the main conversion once there are enough per month.
4. Tighten the inputs. It adds postcode exclusions, negative keywords such as "diy" and "grants only" and a form question about property ownership.
5. Test. It runs the new setup as a campaign experiment against the old for eight weeks.
6. Result. Cost per form lead rises, but cost per survey booked falls, and the sales team reports better quality. The company keeps the new setup and reviews monthly.

The platform did not get smarter. The company gave it a better definition of success.

A worked example: ecommerce and Performance Max

This is also invented. An online shop sells home fitness equipment and wants to try Performance Max alongside its standard Shopping campaign.

1. Set break-even. Average margin is 35 percent, so break-even return on ad spend is about 2.9.
2. Prepare the feed. It improves titles and categories, and adds high-margin and low-margin labels.
3. Split by margin. It puts high-margin products in the Performance Max test and leaves low-margin items in standard Shopping.
4. Add exclusions. It excludes brand terms from the test, so existing demand is not credited to the new campaign.
5. Run a six-week experiment with a capped budget.
6. Judge. It compares blended revenue, margin and new customer share, not just platform-reported return, and decides to expand the test to a second product category.

Signals that automation is going wrong

Signal	Possible cause	What to check
Spend rises, sales do not	Targets too loose or broken tracking	Conversion tracking, target level, search terms
Conversions rise, revenue does not	Counting low-value or duplicate actions	Primary conversion settings, values, duplicates
Delivery drops suddenly	Target too tight or learning reset by changes	Recent edits, targets, budget
Lots of brand search conversions	Automation absorbing existing demand	Brand exclusions, brand versus non-brand reports
Irrelevant queries or placements	Weak exclusions and inputs	Search terms, placement reports, negatives
Strong platform ROAS, flat blended return	Overlap and attribution inflation	Blended revenue against total spend, holdout tests

A monthly review routine for automated accounts

1. Tracking. Confirm conversions still record correctly, once each, with values.
2. Money. Compare spend, revenue and margin with plan. Check blended return.
3. Queries and placements. Review search terms, insights and placements where available. Add exclusions.
4. Assets. Review which creative performs, replace weak assets and check claims.
5. Experiments. Check test results, decide to keep or revert, and record the decision.
6. Notes. Log changes and dates, so future results can be understood.

Questions advertisers ask about automation

Question	Answer
Should I let the system find my keywords?	It can find useful searches, but control matters. Use negative keywords and review search terms regularly
How many conversions do I need?	More is better. Systems learn from volume. Meta's guidance references about 50 optimisation events a week per ad set. Google's needs vary, so check current advice
Is it wise to set a very low target CPA?	Usually not. A tight target can starve delivery. Start near current performance and tighten gradually
What if automation takes credit for brand searches?	Use brand exclusions and compare brand and non-brand performance separately
Can I pause and restart campaigns freely?	Frequent changes can restart learning. Make fewer, deliberate changes
Does automation reduce costs?	It can improve efficiency in some accounts. It can also overspend if inputs are wrong. Measure against break-even

Creative asset quality for automated campaigns

• Headlines and descriptions. Provide many distinct, accurate options that read well in any combination.
• Images. Use original, high-quality images in the required sizes, with a clear subject.
• Video. Provide short, well-lit videos with captions, in vertical and horizontal formats where needed.
• Product feeds. Accurate titles, categories, prices and stock, with high-quality images.
• Brand guidelines. Provide logos and colours and any restrictions the system should respect.
• Review generated assets. Remove anything that is inaccurate, off-brand or non-compliant.

A governance checklist for automated ads

• A named owner reviews performance weekly and the full account monthly.
• Budgets and targets are documented with the date and reason for each change.
• Access to ad accounts is limited and uses two-factor authentication.
• Claims in ads are checked by someone with subject knowledge.
• The business owns the ad accounts, conversion data and audiences, even when an agency manages them.
• Experiments are logged with start and end dates, hypotheses and results.

A story of an automation project that nearly went wrong

This is an invented composite to illustrate common problems. A UK online retailer of garden furniture moved most of its Google Ads budget into Performance Max after a platform representative suggested it. The first fortnight looked excellent: platform-reported return on ad spend was far above the previous Shopping campaigns, and the team celebrated.

Then the finance director compared the platform's figures with actual orders and margin. Revenue from paid sources had risen only slightly, but the platform was claiming much more. Digging in, the team found three reasons. First, a large share of the new campaign's conversions came from people searching for the retailer's name, who would probably have bought anyway through the old brand campaign, which had been paused to "simplify the account". Second, the campaign was winning cheap conversions on low-margin items such as cushions, while bestselling and higher-margin sofas received less budget. Third, some conversions were being counted twice, because both a purchase event and a thank-you page view were marked as primary conversions.

The team fixed these in order. It restored a separate brand campaign and added brand exclusions to Performance Max so that existing demand was not being counted as new. It corrected the conversion setup so that only purchases with values were primary. It split products by margin so that high-margin lines had their own asset group and target. And it ran a campaign experiment comparing the new structure with the old for six weeks, rather than judging by before-and-after numbers.

The result was less dramatic than the first fortnight suggested, but more real. The retailer's blended return on marketing spend improved modestly, new customer numbers rose, and, importantly, the team now trusted its numbers. The founder summarised the lesson in a note to the team: automation is a powerful tool that does what you tell it to, not what you meant. Tell it the wrong thing, or let it take credit for demand you already had, and it will happily produce impressive-looking reports.

Three habits came out of the episode. The team now reconciles platform reports with real orders every month. It tests every major automation change with an experiment. And it keeps a log of changes, so that when results move, it can see what it did and when. Those habits cost little and prevent the most expensive mistakes.

A short note on platform recommendations

Ad platforms offer optimisation scores, recommendations and automated suggestions, such as "raise your budget", "add broad match keywords" or "switch to a different bid strategy". Some are helpful. Others primarily serve the platform's interest in higher spend or wider adoption of its products. Treat recommendations as prompts for investigation, not instructions. For each one, ask what it would change, what evidence there is for the effect on your results, what the risk is and how you would test it. Accept the ones that match your goals, test those that might, and ignore those that do not. Keep a record of what you accepted and why, so you can learn which kinds of suggestion have worked for your account.

Where we can help

We are a digital marketing agency in Manchester, UK and Mumbai, India. Our AI PPC services and Google Ads management start with tracking and break-even numbers, then add automation with human oversight. Contact us for a review of your account.