Why most paid tests say nothing

Three things make a short paid test unreadable, and they stack. The first is sample size. With four purchases a week, the week-to-week cost per purchase swings by a factor of five on the same spend, which is counting error, not a trend. Nobody would call a coin biased after six flips; paid channels get judged on fewer conversions than that every week.

The second is the platform’s learning phaseThe first days of a campaign or a bidding change, when the platform spends to find out which clicks convert before it can bid on that knowledge.. Google Ads and Meta both bid on predicted conversion probability, and the prediction is built from conversions the campaign has already produced. In the first week or two the algorithm is spending your money to find out who converts, so the CPA of those weeks is the price of the warm-up, not the price of the channel. Judge in week two and you are judging the warm-up, the trap the Demand Curve piece names.

Noise or signal

Four purchases a week swings; eighteen settles

Six weeks of the same campaign at two budget levels. Each bar is one week’s purchases; the figure under it is that week’s cost per purchase. The dashed line is the 15 a week floor. Hypothetical example

$300 a weekUnderfunded

24 purchases in six weeks, about 4 a week

15 a week

2$150wk 1

7$43wk 2

1$300wk 3

6$50wk 4

3$100wk 5

5$60wk 6

Weekly cost per purchase runs from $43 to $300. The six-week figure is $75, and nothing in the data says whether that is the number or an accident.

$800 a weekAbove the floor

108 purchases in six weeks, 18 a week

15 a week

16$50wk 1

19$42wk 2

17$47wk 3

20$40wk 4

18$44wk 5

18$44wk 6

Weekly cost per purchase runs from $40 to $50. The six-week figure is $44, and the weeks agree with it.

The small test is not a cheaper version of the large one. It is a different experiment that cannot produce a verdict at any price per week.

The third is that the budget never bought enough conversions to begin with. A $3,000 test against a page that converts paid clicks at 1% on a $4 click produces about seven purchases in total, not per week. That test could not have succeeded whatever the channel did, so its failure carries no information. This is the case most often reported as “paid does not work for us”, and the link between conversion rate and CPA is why it keeps happening.

The conversions you need, and the budget that buys them

Demand Curve’s rule of thumb turns the floor into money: multiply your target cost per acquisition by roughly 60 to 80 to get the minimum monthly spend for a conclusive result. The arithmetic behind it is the weekly floor. Fifteen conversions a week is about 65 a month and 20 is about 87, so 60 to 80 conversions a month at the target CPA is the budget that reaches it. If that figure is above what you can spend for four to six weeks, the newsletter’s conclusion is blunt: you are not ready to test paid yet.

The rule assumes the campaign reaches the target CPA. In a first test it usually does not, because the CPA you actually pay is the click price divided by the conversion rate of the page the clicks land on, and the target was set from margin, not from the page. So the budget is the real cost per conversion times 60 to 80, and the real cost per conversion is CPC divided by landing page conversion rate. Two worked examples, with click prices from WordStream’s 2026 Google Ads benchmarks (13,474 US search campaigns, April 2025 to March 2026) and conversion rates as hypothetical inputs:

1. Online store, conversion event: purchase. The Shopping, Collectibles and Gifts category averages a $4.14 click. At a 2% purchase rate on the landing page, a purchase costs $207 in clicks, and 60 to 80 purchases a month cost $12,420 to $16,560. A $3,000 test is a fifth of the way there.
2. B2B SaaS company, conversion event: qualified demo request. Business Services averages a $5.87 click. At a 3% demo request rate, a request costs $196, and 60 to 80 requests cost $11,740 to $15,653 a month. Count only requests that sales accepts; a form fill that never becomes pipeline is not the event.
3. Compare the figure with what you can spend for four to six weeks without touching the campaign. If the budget is there, commit the whole period before launch. If it is not, the next section shows the only input that brings the figure down without lowering the conversion count.

The budget calculation

Conversions needed times what each one really costs

Demand Curve’s rule is target CPA times 60 to 80 a month. Replace the target with the CPA your page actually produces, click price divided by conversion rate, and the budget becomes a figure you can hold the test to. Hypothetical example

Conversions needed

60 to 80 a month

15 to 20 a week for 4 to 6 weeks, the floor below which the number is noise.

×

Real cost per conversion

CPC ÷ conversion rate

The click price is the platform’s. The rate is the landing page’s.

=

Minimum monthly test budget

CPC ÷ rate × 60 to 80

Commit it for the whole period before launch, or do not launch.

Online store, purchase$4.14 ÷ 2% = $207 a purchase, so 60 to 80 purchases a month cost $12,420 to $16,560. Click price: WordStream 2026, Shopping, Collectibles and Gifts.

B2B SaaS, qualified demo request$5.87 ÷ 3% = $196 a request, so 60 to 80 requests a month cost $11,740 to $15,653. Click price: WordStream 2026, Business Services.

Two inputs are outside your control before launch: the click price and the conversion count. The third, the page’s conversion rate, is the one that moves the total.

How conversion rate changes the price of a trustworthy test

Hold the click price and the conversion count fixed, move only the landing page’s conversion rate, and the budget moves in inverse proportion. The table shows the same two tests at three rates each. Every click price is a WordStream 2026 category average; every conversion rate is a hypothetical input, chosen around the 2026 paid search averages Ruler Analytics reports for retail and e-commerce (2.1%) and software (8.2%) across more than 5 million tracked conversions.

Hypothetical example Click prices: WordStream, 2026 Google Ads benchmarks. Conversion rates and budgets are illustrative inputs and arithmetic, not customer data.

Where the budget goes

You pay for 1,000 visits. 970 leave.

Ads, search, AI answers and social are four ways to earn a visit, four budgets and four promises. Then one page, identical for all of them, and roughly 97 of every 100 visitors gone at the industry’s 2 to 3%. The ads lane is the only one with a price on every visit.

Ads1,000 clicks at $3.40

970 leave30 convert

SEOmonths of content and links

970 leave30 convert

AI answerscitations earned, not bought

970 leave30 convert

Socialposts, reach, replies

970 leave30 convert

$3,400spent on 1,000 ad clicks at $3.40

$3,298of it spent on visitors who left

$102the remainder, on the 30 who converted

The test budget pays for all 1,000 clicks. Only the 30 count toward the floor. Raising the 30 is the only way to make the same budget produce a readable test.

Reading the result without fooling yourself

Write down what “it worked” means before the first click, as the Demand Curve piece insists: the CPA or purchase rate that counts as a pass, the conversion count that makes the number valid, and the end date. A test judged against criteria set afterwards always passes or always fails, depending on who is reading it.

Then compare against something. Before-and-after is the weakest reading a paid test can get, because the test weeks carry seasonality, a product launch and a competitor’s sale that the campaign gets credited or blamed for. A control groupVisitors from the same source over the same weeks who keep seeing the unchanged experience, so the test has something to compare against. on the same traffic over the same weeks removes all of that. The third requirement is to count only humans. Clicks that never scroll, never move and leave in under a second inflate spend and sessions and never convert, so a CPA that includes them is wrong by the bot share.

The site, not the bid, decides how expensive a trustworthy test is.

How to read a paid test

Half see the change, half never do, bots out before the count

Visitors from the campaign are screened for non-human behaviour first, then split at random. Both halves run over the same weeks with the same ads and bids, so whatever differs between them is the change, not the season.

Visitors from ads arrive

Bot screenClicks with no human behaviour are removed before anything is counted

Random 50/50 split

Half see the personalised page

The page opens on the promise the ad made

Half never do

The page as it is today, the control

Compared over the same weeks. Purchases (or accepted demo requests) per real visitor in each half, read as a live A/B test, not before and after.

The same structure reads a channel test honestly: pass criteria written first, a control on the same traffic, and a count that excludes clicks no person made.

In practice this is the method Pathmonk uses to measure its own effect on paid traffic. Half the visitors see a page personalised in real time for the source and the ad that brought them, half see the page as it is, and the two halves run over the same weeks as a live A/B test rather than a before-and-after. Bot traffic is detected and excluded, so experiences, experiments and the 50/50 test only count real visitors. Campaigns, bids and budgets stay in the ad account you own; only what the page shows changes. It belongs here because of the rate: a page that opens on the ad’s promise converts more of the same clicks, which is the input that brings the budget in the table down.

When not to test paid at all

The calculation gives a floor, and some companies should read it as a stop sign rather than a budget line. The decision flow below ends in three outcomes. Fund is for a test that reached the floor and met the criteria written before launch. Extend is for a test too short or too small to read but affordable to carry to the floor. Stop covers two cases: a test that reached the floor and missed, a verdict on the channel, and a test that missed the floor because the page converted below the rate the budget was sized on, a verdict on the page, where the page is the thing to fix before any retest.

Fund, extend or stop

Three questions before the budget meeting

Walk the test down the chart. The first question decides whether there is a result at all. The second reads it against the criteria written before launch. The third tells a failed channel from a failed page.

A paid test has ended, or the money is about to run out

Question 115 to 20 conversions a week for 4 to 6 weeks?

No

OutcomeExtend if you can fund it to the floor. If you cannot, stop and fix the page first: there is no result to read.

Yes

Question 2Did it meet the pass criteria written before launch?

Yes

OutcomeFund. Scale in steps and keep the control running, so the CPA is checked as the budget grows.

No

Question 3Did the page convert the clicks at the rate the budget was sized on?

No

OutcomeStop the test, not the channel. Fix the page, then retest at the budget the new rate implies.

Yes

Stop the channel. It was measured properly and it missed.

Most tests that get called a channel failure exit at question one. Only a test that passes question one and three has said anything about the channel.

Key takeaways

• A paid test below roughly 15 to 20 conversions a week for four to six weeks has produced noise, not a verdict, whatever the CPA says.
• The first week or two measure the platform’s learning phase. Judge nothing before the end date set at launch, and do not pause and restart.
• Minimum monthly budget is cost per conversion times 60 to 80, and the real cost per conversion is your click price divided by the landing page’s conversion rate.
• The page sets the price of the test: at a $4.14 click, 60 to 80 purchases a month cost $24,840 to $33,120 at 1% and $8,280 to $11,040 at 3%, with the same ads and bids.
• Lowering bids to make a test affordable buys fewer, lower-intent clicks and leaves the rate untouched. Raising the rate is the only input that cuts the budget without cutting the count.
• Write the pass criteria before launch, compare against a control on the same traffic over the same weeks, and exclude clicks no human made.
• A test that misses the floor because the page converted below the rate the budget assumed is a verdict on the page. Fix it, then retest at the budget the new rate implies.
• Keep the conversion events apart: purchases and revenue per visitor for a store, accepted demo requests and pipeline value for SaaS. Never add them.

FAQs on paid test budgets and conversion counts