While most sectors are cautiously dipping their toes into artificial intelligence (AI), ad tech jumped in with both feet two decades ago and has been swimming laps ever since. This is a sector built on the back of AI: it is all about processing vast amounts of data at lightning speed and on a massive scale; a combination only possible with the assistance of highly sophisticated algorithms. Tracing their evolution offers a revealing glimpse of automation’s next big leap.
2005: The battle over predictable clicks
The battle lines of the mid-noughties internet were drawn around who had the best pay-per-click (PPC) business. Google, Yahoo!, and Ask were competing for search advertising supremacy, with victory determined not just by user volume, but by how accurately their PPC model could predict whether a user would click on a link or, as the models evolved, whether they would take an action.
These initial machine learning models digested an all-you-can-eat buffet of data derived from search impressions, detecting subtle correlations between search terms and user behaviour that could tilt the odds of a click or conversion. For the first time, advertisers could buy impressions based on their likelihood to facilitate desired outcomes, while sellers could leverage this likelihood to boost the value of their inventory.
Such predictive models were not isolated to search, either. Facebook’s advertising arm grew thanks to an even richer supply of behavioural data drawn from its users, while companies such as Criteo launched similar models across display advertising. This meant building data collection pipes into domain-agnostic platforms rather than the closed loop of search or social media. This was an incredible feat for developers working without a blueprint.
2015: The privacy push towards probabilistic models
By the mid-teens, smartphones and apps had fragmented the online ecosystem, scattering audiences across a maze of platforms and devices. ID graphs, which deploy machine learning models to stitch together user profiles across devices and channels, were born out of necessity. Without them, there was no way to know whether an IP address, an email login, and a device ID all represented the same person.
But as these machine-assembled profiles increased in detail, they caught the attention of regulators, who stepped in to ensure users consented to data collection. In 2018, GDPR kicked in, which eradicated a number of questionable data practices and put consent front and centre. With only consented data to analyse and utilise, audience extension became far more important. This process involves taking a small, known group of consented users and plugging them into a probabilistic algorithm that can extrapolate them to a larger pool of prospects. Thanks to machine learning, this meant advertisers and publishers could achieve audience scale within all regulatory parameters.
2020: AI begins to see and read like we do
In the second decade of the century, AI started to add qualitative capabilities to its quantitative repertoire. CAPTCHAs were used to educate computer vision models which, in combination with large language models (LLMs), allowed machines to process, interpret, and categorise media and text with an increasingly human level of understanding – yet with superhuman speed and scale.
In digital advertising, this meant that contextual categorisation, and the targeting methods that depend on it, could be achieved algorithmically rather than relying on manual keyword tagging. This also meant that the performance of advertising creative could be predicted and refined, with models identifying its visual and textual qualities, then mapping them to outcomes by tracing prior campaign performance.
2022: Generative AI gives machines their mainstream moment
While prior AI evolutions focused on interpretation, 2022 introduced the world to AI that could create. Generative models used the same principles that powered computer vision and LLMs and flipped them to produce media rather than analyse it. Midjourney made waves (and memes) with its machine-generated imagery, but it wasn’t until the launch of ChatGPT in November 2022 that AI went mainstream.
This presented tremendous and exciting potential for agencies and brands working in digital advertising, enabling the creation of numerous assets for multiple campaigns at a time. Meanwhile, the ability to interface with digital advertising platforms through natural language prompts made them more accessible and efficient to operate. At once, tools that once felt technical became accessible, faster, and more collaborative.
2025 and beyond: Thinking machines challenge conventional thinking
AI has been advancing ad tech’s growth since its inception. The question for many is, what’s next? Will the sector continue to enjoy incremental improvements or are we on the cusp of a new industrial revolution? Those who argue for the latter are pinning their hopes on AI agents, autonomous models that can carry out tasks on the user’s behalf with little to no manual intervention.
At their core, AI agents work by breaking a goal into sequential tasks, each executed by a specialised component designed to deliver a specific outcome. What sets agentic AI apart is its ability to consider strategy rather than simply follow a set workflow; the true realisation of a thinking machine. It can explore multiple potential approaches, test them, and select the best-performing option in a supercharged version of the A/B testing that was once advertisers’ bread and butter.
However, these AI agents are currently isolated to specific platforms. For them to be able to realise their true potential, there needs to be cross-industry collaboration of the likes we have never seen before. Ad tech vendors, agencies, brands, publishers, media owners, and platforms need to have AI-friendly APIs that agents can hop between without hitting walls. Whether the next chapter in ad tech’s AI timeline gets written will depend less on technology, and more on the industry’s willingness to collaborate at an unprecedented scale
