Digital advertising with AI is not about switching on Performance Max and crossing your fingers. It is a systematic approach to generating, testing and optimising creatives, audiences and messages at a speed and scale no human team can match manually. This guide explains how it really works and how to implement it to achieve measurable results.
First, the reality: the AI in advertising platforms (Performance Max, Advantage+) has improved significantly over the last two years. For most advertisers with budgets below ten thousand euros a month, letting the algorithm optimise within the right parameters produces competitive results. The problem is that most advertisers don’t set the right parameters or provide the right inputs.
Google and Meta’s AI needs three things to work well: sufficient conversion data (a minimum of 50 conversions per month for Performance Max to learn), high-quality creatives in multiple formats (text, image, video, different ratios) and correct audience signals (not generic interest-based ones, but audiences built from your real data). Without those three inputs, the algorithm optimises towards metrics that don’t represent your actual goal.
The architecture of a well-designed AI campaign
Layer 1: Mass creative generation
Traditional creative testing is a slow process: the creative team proposes three or four variants, the client approves two, both are published, within two weeks there is enough data to know which performs better, and the winning one is scaled. The full process takes between one month and a month and a half for each optimisation cycle.
With AI creative testing, the full cycle takes days. The system generates fifty or more variants combining different headlines (working different angles: benefit, urgency, social proof, curiosity), different body copies, different CTAs and different images or videos. The variants are published simultaneously with a small test budget. A Bayesian model — which does not wait for statistical significance — starts allocating more budget to those that demonstrate better performance from the first few hours.
The result: in one month of systematic AI creative testing, more variants are tested than in a full year of traditional A/B testing. And the CPA drops consistently because the best creative is always being served to the largest share of the budget.
Our Creative Testing product implements exactly this system, with CPA reductions of 25–40% in the first sixty days.
Layer 2: Predictive audiences
Meta’s interest-based audiences or Google’s segments are useful as a starting point. Predictive audiences are qualitatively different: instead of defining who we want to reach (women aged 25–45 interested in fitness), the model learns from the patterns of users who have converted and finds similar users in the available inventory.
The most effective predictive audiences are built from:
- Current customer lists uploaded as Custom Audience to create Lookalike
- Web behaviour captured with the pixel: not just "visited the site" but "reached the pricing page and spent more than 60 seconds"
- CRM data: leads that converted vs those that didn’t have behavioural patterns the model can learn from
Combining predictive audiences with optimised creative testing produces the best results: the right message for the right person at the right moment.
Layer 3: Dynamic budget optimisation
Manual budget allocation across campaigns, ad groups and audiences is a process that requires constant monitoring and frequent adjustments. AI-powered dynamic budget optimisation does that work continuously: it reallocates budget in real time towards wherever the model predicts greater return, considering historical CPA, current trends, seasonality and bid competition.
Google Smart Bidding and Meta’s Advantage Budget Optimisation are the native versions of this. For accounts with sufficient conversion history, they work well. For new accounts or those with low conversion volume, proprietary rule-based + ML optimisation systems tend to outperform the platforms’ native algorithm.
How to use AI to create better creatives
The generation of creatives with AI is not just text. It is the complete process of concept, copy and visual.
For copy: define the brief (product, audience, main benefit, objections to address, CTA) and use the language model to generate twenty headline variants, ten main copy variants and five CTA variants. Then select the combinations with the most potential to test, rather than relying on the copywriter’s instinct to choose a single option.
For visuals: image generators such as FLUX or Midjourney trained on the brand’s aesthetic produce creatives that respect the visual identity without the need for a photo shoot. For e-commerce, this is especially valuable: product photos in different settings and compositions generated in hours rather than weeks.
For video: models such as Runway or Veo can generate short fifteen-second clips that work well as video ads on Meta or TikTok. For brands that have a production budget but want to multiply test variants, AI can generate the additional variants from the original production material.
We go deeper into this in our articles on AI Images and AI Video.
Google Ads with AI: what works and what doesn’t
Performance Max works well when: you have more than 50 conversions per month for the model to learn from, you provide high-quality assets in all formats (text, image, video), you have conversion tracking properly set up with attributed values, and you let the system learn for at least four weeks before making adjustments.
Performance Max does not work well when: conversion tracking is misconfigured (the system optimises towards the wrong metric), you provide low-quality or very generic assets, the budget is too low for the volume of competition in the sector, or frequent changes are made that restart the learning period.
The most costly mistake: adding Performance Max to an account where search campaigns are already working well, without understanding that PMax takes priority and cannibalises brand and branded terms traffic that was generating conversions at a very low CPA. Configuring brand exclusions in PMax is mandatory before activating it.
Meta Ads with AI: Advantage+ properly configured
Meta’s Advantage+ system (formerly Advantage Shopping Campaigns for e-commerce, now extended to more formats) works similarly to PMax: the algorithm controls the audience, placement and budget optimisation. The advertiser only controls the creatives and the objective.
For it to work well: Advantage+ audiences need between 1.000 and 10.000 conversions per year to learn correctly, product catalogues must be fully optimised for dynamic ads, and Meta’s Creative Quality Guidelines must be met across all assets.
A properly configured Meta pixel with Conversions API (not just the browser pixel) is a prerequisite for the system to work. The iOS 14.5 update significantly reduced the browser pixel’s ability to track conversions on Apple devices. The Conversions API sends data directly from the server, without browser limitations.
How to measure the true impact of AI on your campaigns
The most common mistake is measuring the impact of AI in campaigns using only platform metrics. Platforms have incentives to attribute conversions to themselves: Meta’s Multi-Touch Attribution and Google Attribution Model have well-known biases that favour those platforms.
The correct measurement combines:
- Platform data for tactical optimisation
- Incrementality testing to measure real impact (control groups that don’t see the ads)
- Revenue data from the CRM or e-commerce platform for real attribution
The marketing mix modelling (MMM) model is making a strong comeback precisely because cookie-based attribution models are becoming less and less reliable. For companies with marketing budgets above twenty thousand euros per month, a basic MMM is the most honest way to measure which channels are generating real incremental results.