AI Paid Ads Optimisation

The short answer

In a modern Google or Meta account you rarely set individual bids, pick every placement or decide which headline shows next to which description. Smart Bidding, Performance Max, AI Max for Search and Meta's Advantage+ products do that. What you still control is what the machine learns from and what it is allowed to do. Optimisation is now the work of improving those inputs.

The order matters, because each layer depends on the one before it:

1. • The conversion signal. What you tell the platform counts as success, and how accurately.
2. • Structure. How campaigns and asset groups divide budget and learning.
3. • Creative supply. The range and quality of ads the system can choose from.
4. • Guardrails. Negatives, exclusions, targets and budget caps that stop the system spending where it should not.
5. • Bids and budgets. Target values and allocation between campaigns, adjusted slowly.

Most accounts that underperform have spent their effort on layer five while layer one is wrong.

What the platform controls and what you control

Who controls which optimisation lever in an automated ad account

Lever	Platform AI decides	You decide	What to check
Bids per auction	The bid for each search or impression under Smart Bidding or Meta's delivery system	The bid strategy, the target CPA or ROAS, and which conversions count	Is the target based on margin, and is the conversion action the one that matters?
Audience	Who sees the ad, increasingly including people outside any audience you suggest	Audience signals, customer lists, exclusions and geography	Are existing customers excluded where they should be, and are locations set to presence rather than interest where needed?
Ad assembly	Which headlines, descriptions, images and videos appear together	Which assets exist, what they say, and any pins	Does every asset stand on its own and say something different?
Search queries	Which searches match, especially with broad match and AI Max	Keywords, negative keywords and brand settings	Search terms report reviewed weekly; wasteful themes added as negatives
Landing page	With final URL expansion switched on, which page to send people to	Whether expansion is on, URL exclusions, and the page itself	Are the pages Google may choose ones you would choose?

Layer one: fix the conversion signal

Automated bidding optimises towards whatever you count. Google's own description of Target ROAS is that it predicts the value of each potential conversion from the values you report and sets bids to hit the average return you ask for (Google Ads Help). If the value you report is wrong, the system becomes very efficient at the wrong thing.

Check these before anything else:

• • One primary conversion per campaign goal. Newsletter sign-ups, page views and add-to-baskets should be secondary, or the system will buy cheap micro-actions.
• • Values that reflect profit, not just revenue. If product margins differ widely, reporting revenue treats a low-margin sale as equal to a high-margin one. Some advertisers report gross profit as the conversion value instead.
• • Lead quality fed back. For lead generation, a form fill is not the outcome. Import qualified or won stages from the CRM using Google's enhanced conversions for leads or offline conversion import (Google Ads Help) and Meta's Conversions API (Meta Business Help Center).
• • Consent and tagging. Google's Tag Manager Help describes consent signals as required for its measurement and remarketing features on EEA traffic (Tag Manager Help); check the current requirement for your countries in Google's help centre before you rely on it. Missing signals shrink the data the bidding learns from.

A useful test: pick ten recent conversions and trace each one to a real order or a real sales conversation. If you cannot, fix tracking before touching anything else.

Layer two: structure for learning, not for tidiness

Automated systems need enough conversions per campaign to learn. An account split into dozens of small campaigns by product, device and region starves each one. Consolidate where the goal, margin and landing experience are the same, and split only where one of those genuinely differs: different target return, different country with different economics, or a brand campaign that should not share budget with prospecting.

For Performance Max, asset groups should follow themes a customer would recognise (a product category, a service line or an audience with a distinct need) rather than internal team boundaries. See the Performance Max creative assets guide for how to build them.

Layer three: supply creative that gives the system real choices

When the platform assembles ads, it can only choose among what you give it. Twelve headlines that rephrase the same promise give it one real option. The better unit of variety is the angle: price, speed, proof, risk reduction, a specific use case, a specific objection. Meta's engineering blog described its retrieval system, Andromeda, in December 2024 as built to match many ads to many people at speed (Engineering at Meta), which rewards genuinely different concepts rather than near-duplicates.

AI writing tools are useful here because they make it cheap to draft many angles. They are not useful at deciding which claims are true. Every generated asset still needs a human check against the offer, the evidence and advertising rules. The PPC manager's guide to AI ad creative covers that workflow.

Layer four: guardrails

Guardrails are how you keep broad automation honest:

• • Negative keywords, which Google announced in 2025 for Performance Max at campaign level (Google); check what your own account offers today.
• • Brand exclusions, so prospecting campaigns are not credited with people who were already searching for you.
• • Customer exclusions or new-customer goals, where acquisition is the point.
• • URL exclusions if final URL expansion is on, so the system does not send paid traffic to careers pages or blog posts.
• • Budget caps per campaign, so one experiment cannot absorb the month.

Layer five: bids and budgets, changed slowly

Target changes reset part of what the system has learned. Move targets in modest steps, one campaign at a time, and give each change at least one full conversion cycle before judging it. If your typical customer takes ten days from click to purchase, a change judged after three days is being judged on noise.

The weekly and monthly rhythm

Weekly (about an hour per account):

• • Spend against plan, and any campaign more than 20% over or under its expected pace.
• • Search terms: add negatives for irrelevant themes, and note new relevant themes for keywords or assets.
• • Conversion counts against the CRM or shop platform: any gap that has widened?
• • Creative: any asset or ad with enough impressions to judge, rated low by the platform or clearly weaker on your own metrics.
• • Anything the platform changed automatically (auto-applied recommendations should be reviewed, not left on by default).

Monthly:

• • Profit or pipeline per channel, not platform-reported ROAS alone.
• • Target values against current margins.
• • Creative refresh: retire tired angles and add new ones.
• • Channel reports in Performance Max (Google Ads Help) to see where spend actually went.

Where AfM fits

AfM runs this loop as part of its managed marketing service: the AI operating system keeps account context and drafts the analysis and creative, and the founder makes the budget, target and creative decisions and reviews what goes live. If you want an outside view first, the free written breakdown uses public information and any figures you choose to share; it does not need access to your ad accounts.

FAQ

Should I switch off auto-applied recommendations?

Review them rather than accepting them wholesale. Some are harmless housekeeping, others broaden match types or raise budgets. Turn off the categories that change targeting, bidding or budget, and apply those by hand after checking them.

How many conversions does a campaign need for automated bidding to work well?

The platforms publish guidance that changes over time, so check the current Google Ads and Meta help pages. The practical rule is to consolidate until each campaign sees conversions regularly enough that a week's results are not dominated by one or two sales.

Is platform-reported ROAS good enough to judge optimisation?

It is a useful directional signal, but each platform credits itself generously. Compare it monthly with revenue or pipeline in your own systems, and separate brand search from prospecting before drawing conclusions.

What is the single most common optimisation mistake?

Counting the wrong conversion. A campaign optimising towards page views or unqualified form fills will look efficient on the dashboard while producing little revenue.