How Much Should You Spend on AI-Run Ads Before Judging Results?

Plan to spend enough to get 30 to 50 of the results you're paying for, at your expected cost per result, over at least two to four weeks. If a meat box order should cost about $25 to win, that's $750 to $1,250 before you judge the campaign. Spend much less and what you see is mostly luck.

The platforms' own thresholds are higher still. Meta says an ad set generally needs about 50 optimisation events within seven days of its last significant edit to leave the learning phase, and Google's guidance talks of up to about 50 conversions or three conversion cycles for its bidding to calibrate. At $25 an order, leaving Meta's learning phase in a week would take $1,250 a week. When that isn't affordable, change what the campaign optimises for, not how quickly you judge it.

Follow me on Instagram@sagnikteaches

The formula, and three budget scenarios

Test budget = expected cost per result × number of results you need. Thirty results is the least that gives a usable average; fifty is comfortable. The "result" is whatever the campaign is told to optimise for: a purchase, a lead form, a booking.

Connect on LinkedInSagnik Bhattacharya

Spread the test over at least two weeks even if you could spend it faster, because results vary by day of the week, and over at least two conversion cycles (the usual time between someone clicking and buying). For an online order that's often a day or two; for a wedding enquiry it can be weeks.

Subscribe on YouTube@codingliquids

Estimate your cost per result before spending anything

You won't know your real cost per result until you run ads, but you can work out what you can afford, which is more useful:

• Affordable cost per first order = average order value × gross margin. Spend more than this and each first order loses money.
• Affordable cost per customer = that first-order profit plus the profit on repeat orders you can reasonably expect in a year. Only use this if you have repeat data.
• Affordable cost per lead = profit per booking × the share of enquiries that become bookings.

An AI assistant is handy for running these, as long as you give it your numbers rather than asking it to guess:

Help me set a test budget for an ad campaign.
Product: [what]. Average order value: $[x]. Gross margin: [x]%.
Share of customers who reorder within 12 months: [x]%, ordering
[x] more times on average. (If I haven't given a figure, say so;
don't invent one.)
Calculate: affordable cost per first order, affordable cost per
customer over 12 months, and a test budget for 30 and 50 orders.
Then tell me the daily budget for a 4-week test.

Illustrative output for a butcher's meat box: "Affordable cost per first order: $85 × 30% = $25.50. With 40% of customers reordering an average of 2.5 more times, 12-month profit per customer is $25.50 + (0.4 × 2.5 × $25.50) = $51. A 30-order test at $25.50 is about $765; 50 orders about $1,275. A 4-week test of 50 orders needs roughly $45 a day." That matches the arithmetic, and the reply sensibly used the first-order figure for the test rather than the rosier customer value.

A butcher's four-week meat box test

Here's the illustrative test in full. The butcher runs a Meta campaign optimised for purchases, at $45 a day, with three ad variations made from photos of real boxes. Conversion tracking has been checked by placing a real test order.

The verdict: $26.25 per first order is roughly break-even against the $25.50 affordable figure, and clearly profitable once reorders are counted. So the campaign continues at the same budget, with new creative tested in week five and the budget raised in modest steps only after another fortnight of steady results.

Notice two things. The ad set never reached 50 purchases in a week, so by Meta's definition it stayed in learning, and it still produced a clear answer. And week one alone would have suggested killing it. Had the owner judged after seven days, at $35 an order, he'd have stopped a campaign that worked.

The same test on Google Search, worked from clicks

Search campaigns are easier to estimate in advance because you can see rough click prices before spending. Google's Keyword Planner gives a range of likely cost per click for the terms you'd bid on. From there it's two steps: cost per order = cost per click ÷ the share of clicks that buy, and test budget = cost per order × 30 to 50.

For the butcher, illustratively: meat box searches cost around $1.20 a click, and his site converts about 3% of visitors who arrive from search. Cost per order: $1.20 ÷ 0.03 = $40. That's well above his $25.50 affordable figure, so before spending anything he knows a generic search campaign probably won't pay on first orders. Two changes improve the sum: bidding only on high-intent phrases ("grass-fed meat box delivery" rather than "meat box") and improving the landing page. If those lift conversion to 5%, cost per order falls to $24, and a 40-order test needs about $960. Doing this arithmetic first saved him from a test he'd have abandoned in week two.

Short seasons: a delicatessen's hamper ads

A seasonal product gives you almost no time to learn. An illustrative delicatessen sells most of its gift hampers in the six weeks before Christmas. A four-week test that starts in late November ends as the season does, so whatever it learns arrives too late to use.

Its approach: test in October, on a cheaper proxy, then scale in November. In October it ran ads for its everyday gift boxes, optimised for purchases, at $30 a day for three weeks: about $630, which produced 26 orders at roughly $24 each. That answered the questions that matter (which images and audiences worked, and whether the checkout tracked properly) before hamper demand arrived. In November it launched hamper ads using the winning creative, with a budget set from the October cost per order. The test was really about the setup, not the season, and that's the only kind of test a short season allows.

When results arrive weeks after the click

For businesses where people enquire now and book later, the booking is the real result but it arrives too late to judge a campaign on. An illustrative catering company finds weddings are typically booked three to five weeks after the first enquiry. A four-week test judged on bookings would count only the enquiries from its first week.

The practical answer is to judge in two layers. At four weeks, judge on cost per qualified enquiry, using a simple definition written in advance (date, guest count and budget given, within the service area). At ten to twelve weeks, go back and count which of those enquiries became bookings, and work out cost per booking. If the second number is poor while the first looked fine, the ads are attracting the wrong enquiries, and the next change should be to the ad wording or the conversion values, not the budget.

When the maths doesn't work: a food truck's private-hire ads

An illustrative food truck wants more private-hire bookings for parties and weddings, and can spare $300 a month. A qualified enquiry is likely to cost around $60, so a 30-enquiry test would need $1,800. That's six months of budget to reach a verdict. Three options make sense instead:

1. Optimise for an earlier step. Tell the campaign to optimise for "viewed the private-hire page" or "opened the enquiry form", which might cost $3 to $8 each. You'll reach 50 events on a small budget. The risk: the AI finds people who click and browse but never book, so check the enquiries that do arrive.
2. Use tightly targeted search ads instead. A Google Search campaign on exact phrases such as "food truck hire for wedding" spends only when someone is actively looking. Judge it on cost per enquiry over six to eight weeks, and don't expect the bidding to learn much at this volume.
3. Run a capped, fixed-length test and accept a rough answer. $300 a month for three months, judged on bookings, not clicks. If three months produce two bookings worth $1,500 each, the verdict is clear enough without statistics.

The mistake would be spending $300 for one month on purchase-style optimisation, getting four enquiries, and concluding "ads don't work for us". Four is not a result; it's an anecdote.

Subscriptions change the sum: a farm shop's veg box test

When the product is a subscription, the first order is the wrong yardstick. Take an illustrative farm shop whose weekly veg box costs $18 and earns about $4.50 of gross profit a week. Its own records show subscribers stay about 20 weeks on average, so a subscriber is worth roughly $90 of profit. The affordable cost per sign-up is therefore far higher than the first box suggests, and a test showing $38 per sign-up, which looks alarming against a $4.50 first-week profit, is actually comfortable.

The catch is that people won by ads don't always behave like people who found you by walking in. So the farm shop judged its test twice:

• At four weeks, on cost per sign-up. 42 sign-ups from $1,600 of spend: $38 each. Against a $90 lifetime value, continue.
• At twelve weeks, on retention of those 42. If ad-acquired subscribers cancel after six weeks rather than twenty, each is worth about $27, and $38 a sign-up loses money. In this illustration, 31 of the 42 were still subscribed at week twelve, close to the shop's usual pattern, so the campaign stayed on.

That second check is the one most owners skip. Tag every subscriber with how they arrived (a separate sign-up code for ad traffic is enough) so you can run it.

Three ways owners misread the numbers

• Trusting the platform's "results" column. On Meta, "results" means whatever the campaign was set to optimise for. If someone chose traffic or engagement by mistake, results will look plentiful and cheap while sales stay flat. Check the objective and compare with orders in your shop system.
• Crediting ads with sales that would have happened anyway. Ads shown to people already searching for your business name, or retargeting people who were about to buy, look highly profitable. Look at how many results come from new customers.
• Comparing test weeks with a different season. A barbecue-season test against a February baseline flatters the ads; a January test against December flatters nothing. Compare with the same weeks last year where you can, or run a small holdout.

Costs beyond the ad spend

Writing ad text with AI is cheap and fast; writing Facebook and Instagram ad copy with AI covers prompts that don't produce generic lines. For comparing variations without burning the test budget, see testing ad variations with AI on a small budget.

What to look at on day 3, 7, 14 and 28

On the platforms' side of the fence, Meta's learning phase restarts after significant edits such as changing targeting, the optimisation event or the creative, or large budget changes. Google's bidding recalibrates after strategy or setting changes. Every edit made on day 5 moves your real start date.

Write your stop rules before you start

Decide in advance what would make you stop, continue or scale. Writing it down stops you panicking in week one or hoping in week six. The butcher's version, which you can adapt:

TEST: Meta meat box campaign, optimised for purchases
BUDGET: $45/day for 28 days (max $1,260)
AFFORDABLE COST PER FIRST ORDER: $25.50

STOP EARLY IF:
- tracking is broken and can't be fixed within 2 days
- after 14 days, cost per order is above $50 AND fewer than 10 orders
CONTINUE AT SAME BUDGET IF:
- after 28 days, 30+ orders at $20 to $32 each
SCALE (raise budget in modest steps, one change a week) IF:
- after 28 days, 30+ orders under $20 each
STOP AND RETHINK IF:
- after 28 days, cost per order above $32 or under 20 orders
CHANGES ALLOWED: one batch of edits per week, logged with date

For how Meta's automated campaigns behave for small budgets, whether a small business should use Meta Advantage+ campaigns is the companion piece, and for Google, whether AI can manage your Google Ads without an agency. Whatever the platform, check that purchases are being counted correctly before the first dollar goes out; setting up conversion tracking first is the step that makes every number above trustworthy.

Test-budget questions owners ask

Can I test with $5 a day?

You can run ads on $5 a day, but you can rarely judge them. At $5 a day you spend $140 in four weeks, which buys only a handful of sales for most products. Use a budget that small to test an early step such as landing-page visits or quote-form starts, and treat any sales it produces as a bonus rather than evidence.

Should I split my test budget across Meta and Google?

Not at the start, if the budget only just covers one test. Two half-funded tests produce two sets of noise. Pick the platform that fits how customers find you: Google when people search for what you sell, Meta when they need to be shown it. Run the second test once the first has given a clear answer.

What if the first test fails?

Separate the causes before deciding. If people clicked but didn't buy, look at the landing page, price and offer. If few people clicked, look at the creative and the audience. If tracking recorded fewer sales than you actually received, fix measurement first. A failed test that tells you which of these went wrong is still useful.

Further reads

• How Meta and Google Use AI to Run Your Ads: A Plain-English Guide — How the platforms' AI decides who sees your ads.
• Performance Max for Small Businesses: 9 Costly Mistakes — The costly Performance Max errors small accounts make.
• How to Build a Landing Page With AI That Converts — If people click but don't buy, fix the page first.
• How to Keep Your Brand Safe When AI Places Your Ads — Controls that stop automated ads appearing in bad places.
• How to Measure Whether AI Is Improving Your Marketing Results — How to tell whether AI is improving your marketing overall.
• How Much Should a Small Business Budget for AI? — Fit ad testing into a wider AI budget.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Meta Business Help Centre (learning phase); Google Ads Help (learning period, bidding algorithms and conversion cycles, budget overdelivery).