Measuring AI Ad Creatives: The Metrics That Matter

When you start making many versions of an ad, the dashboard fills with numbers. It is easy to watch the ones that move most and miss the ones that matter. This post names the few numbers worth reading when you test AI ad creatives, says what each one tells you, and gives a weekly routine for reading them.

It gives no benchmarks, because what counts as good depends on your market, your price and your audience. It shows you how to compare your own ads against each other, which is what a creative test is for. Lesson 4.5 covers how to set that test up: Make ad creatives in minutes.

Compare ads against each other on a few numbers, in order: were they seen, were they clicked, what did a click cost, and did clicks turn into sales.

Activity numbers versus outcome numbers

Ads dashboards show two kinds of number.

Activity numbers describe what happened to the ad: how many people saw it, how many clicked. They are early signals.

Outcome numbers describe what happened to your business: orders, bookings, enquiries, sign ups. They are what you are really paying for.

The common trap is to track only activity. An image can win clicks and still lose sales. Always look at both, and when they disagree, trust the outcome.

The five numbers to read

1. Reach and impressions: was it seen?

Before comparing anything, check each ad was actually shown to enough people. If one ad was seen by a handful of people and another by thousands, you cannot compare them. Treat a low view count as "not enough information", not as "this ad is bad".

2. Link clicks: did people act?

Meta's help centre defines cost per link click (CPC) as the amount spent divided by link clicks (Meta Business Help). Link clicks are clicks that take people towards your site or page, which is different from every tap on an ad such as likes or opening the image. For a creative test, link clicks tell you which image or headline drew people in.

A simple way to compare: divide link clicks by the number of people who saw the ad. Doing that yourself gives you a rate you can compare across ads of different size. Your ad platform may show a click through rate directly. Check its definition in the help pages before you rely on it.

3. Cost per link click: what did attention cost?

Using the definition above, a lower cost per link click means each visit cost you less. It is useful for comparing versions of the same offer to the same audience. It is not useful for comparing different offers or audiences, because those carry different costs.

4. Frequency: is it getting tired?

Meta defines frequency as the average number of times each account saw your ad (Meta Business Help). If the same people see the same ad again and again, results can fade. In a test, uneven frequency between ads can distort the comparison, so look at it when two ads look oddly different.

5. Results and cost per result: did it pay?

Results are the outcome you chose when you set up the campaign, such as orders or leads. Cost per result is what you spent divided by results. This is the number that decides whether an ad is worth keeping. Early on you may have only a few results, so read it with care. A cost per result based on a very small number of results can swing wildly.

Numbers you can safely ignore at first

• Likes and reactions. Nice to see, weak link to sales.
• Total spend by itself. Spend matters only against results.
• Ranking ads by one day of data. Too early, too noisy.

A weekly reading routine

Pick the same time each week and follow the same order. Consistency matters more than depth.

1. Check the test is still cleanConfirm nothing was edited mid test and all ads delivered. If something broke, note it.
2. Check each ad had enough viewsIf not, extend the test or repeat it instead of judging.
3. Compare link clicks and cost per link clickWhich creative drew people in, and at what cost?
4. Look at frequencyIs any ad shown to the same people far more than the others?
5. Check resultsDid clicks turn into orders or enquiries? Does the click winner also win here?
6. Write one sentence"Close ups beat lifestyle scenes on clicks, and on orders." Then plan the next test.

How long before you read

Read at the end of the planned test, not before. Meta's A/B testing guidance recommends tests of at least 7 days, says tests under 7 days may be inconclusive, and notes that if your customers usually take more than 7 days to convert you should run longer (source). Google's Performance Max guidance suggests waiting 2 to 3 weeks before deciding to replace low performing assets (Google Ads Help). Both point the same way: patience protects you from false conclusions.

An illustrative reading

This is an invented example with no real figures, to show the thinking.

Say you tested five images with the same headline for a bakery. At the end you find that image 2 had the most link clicks, image 4 had the fewest, and images 1, 3 and 5 sat in the middle. But image 2 produced fewer orders than image 1. What do you do?

• You would not simply pick image 2. The click winner is not the sales winner.
• You would check whether image 2 and image 1 had similar views and frequency, to see if the comparison is fair.
• You would check the landing page. Maybe image 2 promised something the page does not deliver.
• You would write the finding down and run a second test on the two leading images.

The thinking matters more than any number. You are looking for a reason, not just a winner.

When the numbers look fine but sales are weak

Sometimes ads draw clicks and nobody buys. Look beyond the creative. The cause is often one of these:

• The page does not repeat the ad's promise.
• The page is slow or awkward on a phone.
• The offer or price is not compelling.
• You are reaching the wrong people.

Fixing the page can do more than another round of images. The lessons on a live website and a page that sells address that, and Lesson 4.3 covers being found in search.

Keeping a simple log

A spreadsheet with one row per ad is enough. Columns: date, what changed, views, link clicks, cost per link click, frequency, results, one sentence of what you learned. After three tests you will see patterns no dashboard shows. For a framework that uses this log each round, see a repeatable framework for AI ad creatives.

Where to go next

Choose one test, decide its length, and decide in advance which numbers you will read and in what order. Deciding first stops you from picking whichever number flatters your favourite ad.

To set up that test from scratch, with every step shown, take Lesson 4.5: Make ad creatives in minutes.

Questions people ask

What is a good click through rate?

There is no single answer, since it varies by industry, offer and audience. Compare your own ads to each other rather than to a number from someone else.

Should I pick the ad with the lowest cost per click?

Not by itself. Check that it also leads to results, because cheap clicks that never buy are not cheap.

How many results do I need before deciding?

There is no magic number. If results are very few, treat the finding as a hint and repeat the test.

Do I need to track anything outside the ad platform?

If your orders happen on your own site, check that sales are being recorded there too, so you are not relying on one dashboard.