TL;DR
- Machine learning is a subset of AI; deep learning is a subset of machine learning that uses multi-layer neural networks.
- Google Ads Smart Bidding (Target CPA, Target ROAS, Maximize Conversions, Maximize Conversion Value) sets a bid in each auction from signals like device, location and time.
- Meta says an ad set exits its learning phase once it can deliver stably, which usually happens after about 50 results in the week after its last significant edit, so rare conversion events can keep it in learning.
- Thai businesses that close sales in LINE or Facebook chat need offline conversions or the Conversions API, and must handle customer lists under the PDPA.
Machine learning is a branch of artificial intelligence (AI) in which computers learn patterns from large amounts of data and then use those patterns to make predictions or decisions about new data, without a person writing every rule in advance. For marketers, machine learning is the system behind Google Ads Smart Bidding, Performance Max, Meta Advantage+ and the product recommendation engines on e-commerce sites.
This article explains machine learning for business people, without equations. It covers the difference between AI, machine learning and deep learning, the main types of learning, examples inside the marketing tools you already use every day, the data you need before the system can work well, and the limits worth knowing before you hand decisions to automation.
What machine learning is, explained for business people
Compare two ways of deciding which customers are likely to buy. The first is to write the rule yourself, for example "if someone visits the pricing page more than twice, treat them as interested." This is easy to understand, but the rule is set by the writer's gut feeling and does not adapt when customer behaviour changes. The second is to feed in historical website visits together with the outcome of who bought and who did not, and let the system work out which signals relate most to buying. The second approach is machine learning.
The result of the learning is called a "model", which is a set of statistical relationships the system has drawn from the data. When new data comes in, the model returns a prediction, such as the probability that this click becomes a sale, or the product category this user is likely to be interested in. The key thing to understand is that the model does not "know" anything beyond the data it was given. If the data is wrong, incomplete or skewed, the answers will be wrong or skewed too.
How machine learning differs from AI and deep learning
The three terms nest like three circles. AI is the largest circle and means any system that performs work that would normally require human intelligence, including systems that run on rules people wrote. Machine learning is the middle circle inside AI and means systems that learn from data instead of following fixed rules. Deep learning is the innermost circle, a machine learning technique that uses multi-layer neural networks and suits complex data such as images, audio and text.
| Term | Short meaning | Example marketers see |
|---|---|---|
| AI (artificial intelligence) | Systems that do work normally requiring human thinking, including fixed-rule systems | A chatbot that answers from a preset script |
| Machine learning | Systems that learn patterns from data and predict on new data | Automated auction bidding, customer segmentation |
| Deep learning | Machine learning using multi-layer neural networks | Image recognition, translation, large language models |
| Generative AI | Deep learning models that create new content such as text or images | Tools that help write ad headlines or create images |
In practice, when an ad platform says it uses "AI", it usually means machine learning used to predict outcomes such as the likelihood of a conversion. Features that write text or create images automatically are generative AI, which is a different job. Separating the two helps you ask the right question of a tool: are you asking the system to predict, or to create content?
The main types of machine learning in plain language
Supervised learning: learning from examples with answers
The system receives data along with the "right answer", for example ad click history with the outcome of which clicks led to a sale, and learns which signals lead to that answer. Most marketing work that involves prediction, such as conversion likelihood, lead scoring or predicting which customers are about to churn, falls into this group, because there are real outcomes for the system to check itself against.
Unsupervised learning: finding patterns with no answers given
The system receives data with no labelled answers and looks for hidden groups or structure. A common example is customer segmentation (clustering) based on purchase behaviour, which may surface groups the marketing team has never named, such as people who only buy during sales or people who reorder the same product on a cycle. Results like these still need a person to judge whether they mean anything for the business.
Reinforcement learning: learning by trying and getting rewarded
The system tries something, receives feedback on whether it was good or bad, and adjusts its next action to get a better result. The idea is close to testing several ads and gradually shifting budget toward the ones that perform, which requires balancing "trying something new" with "using what is known to work." Platforms do not disclose which technique each feature uses, so treat this as a principle rather than assuming a given feature uses a given method.
Machine learning in the marketing tools you use every day
Smart Bidding in Google Ads
Smart Bidding is the group of Google Ads automated bid strategies that use machine learning to set a bid in each individual auction (auction-time bidding): Target CPA, Target ROAS, Maximize Conversions and Maximize Conversion Value. At the moment someone searches, the system weighs a large number of signals, such as device, location, time of day, language, remarketing lists and the search query, predicts how likely that click is to convert or how much value it will bring, and then sets the bid in line with the goal you defined.
Smart Bidding only optimises for what you tell it counts as a conversion. If you count a tap on the call button as a conversion when most calls are general enquiries, the system learns to find people who tap call, not people who buy. When you change a strategy or target significantly, the strategy shows a "learning" status in the account, and results during that period can fluctuate. Changing targets too often means the system never gets to learn continuously.
Performance Max
Performance Max is a single campaign type in which Google's machine learning chooses placements, audiences and combinations of ad assets across Search, YouTube, Display, Discover, Gmail and Maps. The marketer supplies the assets (text, images, video) and audience signals as a starting point, and the system expands the search for people from there. The trade-off is less control and less visibility than separate campaign types, so it relies on accurate conversion measurement and on checking the asset and search term reporting the platform does make available.
Meta Advantage+
Meta uses the Advantage+ name for a group of automated features, such as Advantage+ audience, which lets the system expand beyond the audience you set when it expects better results, and Advantage+ campaigns, which hand more of the audience, placement and budget decisions to the system. Meta describes an ad set as being in a "learning phase" after it is created or significantly edited, and says ad sets exit that phase as soon as they can deliver stably, which usually occurs after about 50 results in the week after the ad set's last significant edit. This matters in practice: if you choose an event that rarely happens, such as a high-priced purchase, the ad set may stay in the learning phase. Some businesses therefore choose an event that happens more often but is still tied to buying, such as add to cart.
Product recommendation systems
The "you might also like" block on an e-commerce site, or the list of videos suggested next, comes from a recommendation system that learns from the behaviour of many users, for example that people who view product A often buy product B too. These systems can work on web pages, in email and in behaviour-triggered LINE messages. Worth knowing: recommendation systems need a fair amount of behaviour history, so new products or new customers with no data get less accurate suggestions. This problem is called cold start.
The data you need before machine learning works well
An automated system is only as good as the data you feed it. Before switching on any automated strategy, check these things in your own account.
- Conversion tracking is accurate and not duplicated: check that the tag fires once per purchase, that a refreshed thank-you page is not counted again, and that primary and secondary conversions are clearly separated.
- Choose a goal that reflects the real business: if the business earns money from leads that close, importing actual sales data back into the platform (offline conversion import) helps the system learn from valuable leads instead of from every form fill.
- Send the right values: if you use Target ROAS or Maximize Conversion Value, conversion values must match real revenue or profit, not a placeholder number that is the same every time.
- Enough data volume: campaigns with very few conversions give the system too little to learn from. Check each platform's current guidance in your account, and consider consolidating similar campaigns so data is not spread too thin.
- Your own customer data (first-party data): customer lists for Customer Match, or custom audiences on Meta, tell the system what a good customer looks like, but they must be collected and used within the consent you have.
- Enhanced conversions: a Google Ads feature that sends hashed customer data, such as email addresses, to help match conversions more completely when cookies are limited.
Limits and bias marketers should know
Machine learning helps a great deal, but it does not replace human judgement. The limits worth understanding are these.
- It learns from the past: the model predicts from patterns in historical data. When things change quickly, such as a big promotion, a festival or a website outage, the system can respond slowly or in the wrong direction. Google Ads offers seasonality adjustments and data exclusions so you can warn the system in advance or remove abnormal periods from the data.
- It only optimises for the goal you set: if the goal is cheap conversions, the system may find that existing customers or people searching your brand name are the easiest wins, and spend budget on people who would have bought anyway. Measuring new customers separately and running experiments with a control group help answer whether the system is creating additional sales.
- Bias in the data: if historical data reflects skewed decisions, the model learns that skew too. In marketing, one visible effect is that the system may show ads to some groups less and less simply because those groups saw few ads in the past and so generated little data.
- You cannot see the reasoning inside: the platform does not tell you why it set a given bid in a given auction, so marketers have to judge on aggregate results, such as cost per conversion, revenue and lead quality, rather than understanding every decision.
- Interests are not fully aligned: ad platforms earn money when budget is spent, so automated recommendations should be checked against your own business goals before you accept them.
An illustrative example (hypothetical numbers for explanation only): an online store sets Target ROAS but sends conversion values as gross sales before returns. If some product categories have a high return rate, the system sees those categories as high-revenue and shifts more budget into them, even though the real profit is lower. The fix is to send values closer to real profit, or to split high-return categories out and manage them separately.
What this means for Thai marketers
For businesses in Thailand, the first issue is data generated off the website. Many businesses close sales through LINE Official Account, Facebook chat or phone. If those sales are not sent back to the ad platform, the machine learning system sees only the click or the chat message, not who actually bought. Connecting sales data back into Google Ads or Meta through offline conversions or the Conversions API is therefore especially important in a market where buying through chat is normal.
The second issue is the use of customer data. Uploading customer lists for Customer Match or Meta audiences is a use of personal data and must follow Thailand's Personal Data Protection Act (PDPA), for example by having an appropriate legal basis or consent and stating the purpose clearly. The third issue is language. Thai ad assets that the system combines automatically should be checked to make sure they read naturally in every combination, because a headline that is fine on its own can read oddly when paired with certain descriptions.
FAQ about machine learning
What is machine learning, in short?
Machine learning is a way for computers to learn patterns from data and apply them to predict on new data, without writing every rule by hand. It is part of AI and is the core technology in Smart Bidding, Performance Max and Meta Advantage+.
How is machine learning different from AI?
Machine learning is a subset of AI: AI means any system doing work that needs intelligence, including fixed-rule systems, while machine learning means only systems that learn from data, and deep learning is a further technique within machine learning.
Can a small business use machine learning?
Yes, because most businesses already use machine learning through ad platforms without building models themselves. What they need to do is measure conversions correctly, choose goals that reflect real revenue, and avoid changing settings so often that the system cannot learn.
Why is Smart Bidding not working even after I switched it on?
Common causes are conversion data that is inaccurate or too thin, CPA or ROAS targets that are unrealistic, or strategy changes made during the learning period. Check the strategy status, conversion tracking and change history in the account before concluding the system does not work.
What are the risks of machine learning in advertising?
The main risks are the system optimising for the wrong goal, learning bias from historical data, and taking credit for sales that would have happened anyway. Setting goals that match the business and running controlled experiments help reduce these risks.
Summary
Machine learning means letting a system learn from data instead of writing the rules yourself, and it sits at the core of the ad tools marketers use every day. The quality of the results depends mainly on the quality of the data and the goals you give the system, not on switching on an automated feature.
If you want to set up conversion tracking and bid strategies so the system learns from accurate data, see our Google Ads service and Facebook Ads service. If you need to connect sales data from chat or a CRM back into your marketing systems, see marketing process automation, and for reworking data systems across the organisation, see the digital transformation service.
