Machine learning is a branch of artificial intelligence in which software learns patterns from examples and then uses those patterns to make predictions or decisions. Instead of a person writing every rule, the system works out the rules from data, and keeps adjusting them as new data arrives.
How machine learning works
A model is trained on historical examples where the outcome is known, such as thousands of website visits labelled “bought” or “did not buy”. It looks for combinations of signals that tend to come before each outcome, is tested on examples it has not seen, and is then used to predict outcomes for new cases. There are three broad types:
- Supervised learning uses labelled outcomes to predict things like whether a visitor will convert or a customer will cancel.
- Unsupervised learning finds groups in data without labels, such as clusters of customers who behave alike.
- Reinforcement learning improves by trial and feedback, trying actions and keeping what earns a reward.
Deep learning is a subset that uses large neural networks, and it is what powers image recognition and large language models.
In marketing, machine learning is already running much of the work. Google Ads’ automated bidding predicts the chance of a conversion in each auction from signals such as device, location, time of day, search query and audience, and sets the bid to match. Meta’s ad delivery decides who sees each ad in a similar way. GA4 offers predictive metrics such as purchase probability when a property has enough data. Google Search has used machine learning to interpret queries at least since 2015, when it announced RankBrain.
Why it matters
For anyone running ads, the job has changed. The platforms make thousands of decisions you never see, and your influence comes mainly through the data and goals you give them. A model trained on accurate conversions with sensible values will find more of the right customers. A model fed a weak signal, such as page views or every form fill including spam, will find more of that instead.
The UK adds a practical wrinkle. Visitors who refuse analytics and advertising cookies are not observed in the usual way, so ad platforms fill some of the gap with modelled conversions, which are themselves estimates from machine learning. Clean consent set-up and good first-party data give those models more to work with.
Common mistakes
- Optimising for a soft goal, such as clicks or page views, so the system learns to find cheap, uninterested traffic.
- Changing targets, budgets and settings so often that the model never settles.
- Splitting spend across so many campaigns that none has enough conversions to learn from.
- Assuming the model knows your margins or which products you most want to sell.
- Trusting the platform’s own report of its success without a controlled test.
How to act on it
- Check your conversion tracking before touching bids. Every model downstream depends on it.
- Send conversion values that reflect what a sale or lead is really worth, and import offline sales where you can.
- Consolidate campaigns so each has enough conversions to learn from.
- Give changes time to settle before judging them.
- Test automated approaches against a control using campaign experiments rather than trusting before-and-after comparisons.
Setting up accounts so the bidding systems learn from the right signals is central to my PPC management work.