Advanced · Lesson 3 of 4
AI in performance marketing: what works today
About 11 minutes · Yogender Kumar, Managing Partner
By the end of this lesson you'll be able to
- Explain where AI already runs inside ad platforms
- Choose tasks where AI tools save time without adding risk
- Test an AI-assisted change against a control
Why this matters
Every platform now sells AI features. Some of it already runs your campaigns, and some needs a person checking every line. This lesson helps you tell which is which.
The core idea
AI shows up in performance marketing in three places.
Inside the platforms. Automated bidding on Google and Meta uses machine learning to set a bid for each auction. Broad audience tools decide who sees your ads. You don't switch these on as AI. They're simply how the platforms work now. Your job is to feed them accurate conversion data and sensible targets.
As a drafting tool. Large language models can write first drafts of headlines and descriptions, group thousands of search terms by theme, summarise reports, and suggest test ideas. This is where most teams save time.
In analysis. Models can spot unusual patterns in data, such as a partner's sales clustering at odd hours. A person still decides what the pattern means.
The limits are just as clear:
- Models make things up. A drafted ad can include a claim, a price or a statistic that isn't true. The advertiser is responsible for what it says, and the CAP Code applies all the same.
- Models know nothing about your margins or your customers unless you tell them.
- Automated bidding optimises what you count. Bad tracking gets you confidently wrong results.
Worked example
Hale & Rye is an imaginary UK bakery chain selling gift boxes online, so the figures are illustrative. Its paid search CPA is £30.
It uses an AI tool to draft 20 new headlines from its product pages and reviews. A copywriter checks each one, removes 12 (four made claims the bakery couldn't prove), and edits the other 8.
Hale & Rye runs an experiment for six weeks, using the platform's experiment tool. Traffic is split at random between the new headlines and the old:
- Old headlines: £6,000 for 200 orders, £30 each.
- New headlines: £6,000 for 222 orders, about £27 each.
The copywriter also spent about 4 hours on the task instead of 10.
The honest limits. A gap of 22 orders on about 200 each could easily be chance, because counts like these swing by around 20 on their own. The test shows the new headlines are no worse, not that they're better. Hale & Rye keeps them for the six hours saved, and repeats the test next quarter.
My view: let AI draft and sort, and let people decide what goes live. The four false claims would have reached customers if nobody had read them.
Hale & Rye's headline test: cost per order
In practice
- Give automated bidding accurate conversion data and a realistic target.
- Use AI for first drafts, search term grouping and report summaries.
- Have a person check every claim, price and promise before anything goes live.
- Test AI-assisted changes against a control, like any other change.
- Never paste customers' personal data into a tool without checking it's allowed.
- Record the time saved as well as the results.
Common mistakes
- Publishing drafts unchecked. Models invent claims, and advertising rules still apply. Read every line.
- Blaming the algorithm for bad inputs. Automated bidding follows your tracking and targets. Fix those first.
- Treating a vendor's AI label as proof. Ask what it does, and test it against a control.
Try it yourself
An AI-assisted ad test spends £4,000 on each side. The control gets 160 orders. The AI-assisted version gets 176. What's the CPA on each side, and how much cheaper is the new one?
Show answer
Control: £4,000 ÷ 160 = £25. AI-assisted: £4,000 ÷ 176 = £22.73. That's £2.27 less an order, about 9% cheaper. The gap is small enough to be chance, so run the test longer before relying on it.
Quick check
Not scored. Pick an answer to see the explanation.
Which task is safest to hand to an AI tool today?
The control spends £3,000 for 120 orders. The AI-assisted side spends £3,000 for 150. What's the AI-assisted CPA?
Key takeaways
- AI already sets bids and audiences inside the platforms.
- It saves most time on drafting, sorting and summarising.
- People must check every claim before anything goes live.
- Test AI changes against a control, like any other change.