Who coined the term machine learning?
Arthur Samuel, an IBM researcher, popularised it in a 1959 paper about a program that got better at checkers by playing game after game. The idea has not changed since: instead of writing the rules, you supply examples and let the program find the rules.
Most machine learning learns from labelled examples, such as past ad clicks marked "bought" or "did not buy". Some finds groups in data nobody labelled, such as customers who behave alike. Some learns by trial and reward, which is how game-playing programs learn and is part of how chat assistants are trained. Large language models are machine learning on an enormous scale.
Where is machine learning already running your ads?
In the bidding and in the delivery. Google Ads Help describes Smart Bidding as bid strategies that use machine learning to set a bid in each auction from how likely that search is to convert. Meta's delivery system decides which people in your audience see each ad, and it needs a run-in period: the Meta Business Help Centre says an ad set usually needs about 50 of the results it is aiming for within 7 days to leave the learning phase, and results are less stable until it does.
Both systems learn from your conversions. Fewer conversions means less to learn from, and that is the small business problem in one sentence.
What can a dive shop's budget teach the algorithm? (example figures)
A dive shop in Cairns runs a Meta campaign told to find bookings for its $189 introductory reef dive. Each booking costs about $45 in ads. At $600 a week the campaign buys about 13 bookings, well short of 50, so the ad set sits in Learning limited and its results swing from week to week.
| Option | Weekly spend | Results a week | Leaves learning? |
|---|---|---|---|
| Keep aiming at bookings | $600 | About 13 bookings | No |
| Raise the budget to match | $2,250 | About 50 bookings | Yes, if the cost per booking holds |
| Aim at checkouts started, about $9 each | $600 | About 67 checkouts | Yes, though it learns from a weaker signal |
For most small budgets the third row, or something like it, is the realistic one. Let the system learn from an earlier step that happens more often, and judge the campaign on bookings in your own records. Waiting for 13 bookings a week to teach the algorithm what 50 would is the plan that does not work.
When is machine learning the wrong tool?
When there is too little data, when the rules are already known, or when someone needs to explain the decision. Nobody needs a model to learn that GST is 10 percent. A business with 20 sales a month gives a learning system very little to work with, and a decision about a person (a loan, a job application) needs reasons a human can check. For a five-person business, the useful question to ask of any "AI-powered" feature is how many examples it learns from.
What does CoraVision read from the machine-learning side of your ads?
CoraVision reads the results of Google's and Meta's machine learning every night. Once connected, it reads each Google Ads campaign's spend, conversions, conversion value and bidding type, and each Meta campaign and ad set's spend, results and their value. The dashboard's Google Ads and Meta cards show the last 7 days of cost per acquisition and sales, or leads if you win work through enquiries. CoraVision changes nothing in either account: when a bidding change is the right step, Cora teaches it on your own screen and you make it.
Questions
Is machine learning the same as AI?
Machine learning is one branch of AI, and by far the most successful one, so most of what gets called AI today is machine learning underneath. Earlier AI relied on rules written by hand, such as the expert systems of the 1980s.
Why do my ad results wobble after I make a change?
Big edits send the system back to learning. The Meta Business Help Centre says significant edits, such as a large budget change or new creative, can restart the learning phase, and Google Ads shows a Learning status on a bid strategy after a change. Make changes in fewer, larger steps and give each one time.