How much data do you actually need? Performance campaigns, A/B tests and forecasts, with the numbers

TL;DR

Difficulty: Intermediate

Most marketing decisions estimate a proportion, and the margin of error shrinks only with the square root of the data. A 2% shop needs 21,110 visitors per variant to detect a 20% lift. A cost per acquisition built on 10 conversions is uncertain by 62%, on 100 by 20%, so 100 conversions per option is the first real decision. Spend to a decision is conversions times cost per conversion, and time is that spend divided by the daily budget. Small budgets can still kill losers early with the rule of three and Thompson sampling. Below about 3,000 visitors a month no conversion test works, and expertise has to stand in as the prior. Seasonal forecasts need two years of history.

Everyone promises growth, and then the project starts and the numbers are missing: seven conversions last month, no clean tracking, a budget sized for "a test" rather than for the observations a decision needs. This is the arithmetic to put on the table before anything is signed: how many visitors a test needs at your conversion rate, how many conversions a cost per acquisition needs before it means anything, what that costs and how long it takes on each platform, how to stop paying for losing ads early, what to do while the data is too small, and how much history a forecast needs.

The conversation nobody has before the contract is signed

Everyone promises growth. The pitch deck has a curve going up and to the right, the proposal says "data driven", and the kick off meeting is full of energy. Then week two arrives and someone asks the only question that matters: how many conversions did we actually get last month? The answer is seven. Seven conversions, a few hundred visitors, no clean tracking, and a budget sized for "a test campaign" rather than for the number of observations a decision needs. From that moment the project quietly runs on opinion while everybody keeps saying "data".

The disconnect is rarely about effort or talent. It is about nobody doing the arithmetic up front: how much data a given decision needs, what that data costs on each platform, how long it takes to arrive at a given daily spend, and what you can honestly do while you wait. This article is that arithmetic, in full, with the formulas, the tables and the uncomfortable numbers. It is the conversation I want to have before anything is signed, so that everyone knows what to expect.

It is long on purpose. Skim the tables if you are in a hurry, but read the section on expertise before you decide your numbers are too small to bother, because that is where most of the useful work for smaller businesses happens. If you want the short version for your own business, the interactive planner on my data science page takes your visitors, customers, history and budget and tells you which of the four stages at the end you are in. It runs on exactly the formulas below, so its numbers and this article's numbers are the same numbers.

The one formula everything hangs on

Almost every "how much data do I need" question in marketing is a question about estimating a proportion: a conversion rate, a click through rate, a churn rate, a win rate. A proportion estimated from a sample has a margin of error, and at 95% confidence that margin is (see confidence intervals in the Foundations)

MoE=z0.975p(1−p)n≈1.96p(1−p)n\text{MoE} = z_{0.975}\sqrt{\frac{p(1-p)}{n}} \approx 1.96\sqrt{\frac{p(1-p)}{n}}

where pp is the proportion you are estimating, nn is the number of visitors, clicks or impressions it was measured on, and z0.975≈1.96z_{0.975} \approx 1.96 is the point of the normal distribution that leaves 2.5% in each tail. Notice the square root. To halve the margin of error you need four times the data. To cut it to a tenth you need a hundred times the data. That single fact explains most of the budgets in this article.

How many visitors an A/B test needs

The second formula sizes a test. To detect a difference between two conversion rates with a given false positive rate and power, each arm of the test needs roughly (the derivation is in sample size in the Foundations)

nper arm=2 (z1−α/2+z1−β)2 pˉ(1−pˉ)δ2,pˉ=p1+p22n_{\text{per arm}} = \frac{2\,(z_{1-\alpha/2} + z_{1-\beta})^2\,\bar{p}(1-\bar{p})}{\delta^2}, \qquad \bar{p} = \frac{p_1 + p_2}{2}

where p1p_1 is the current conversion rate, p2p_2 the rate you hope the change produces, δ=p2−p1\delta = p_2 - p_1 the difference, α\alpha the false positive rate you accept (usually 0.05), 1−β1-\beta the power, the chance of spotting a real difference of that size (usually 80%), and z1−α/2≈1.96z_{1-\alpha/2} \approx 1.96 and z1−β≈0.84z_{1-\beta} \approx 0.84 the matching points of the normal distribution. The constant 2(1.96+0.84)22(1.96 + 0.84)^2 is about 15.7. Everything else depends on your baseline rate and on the smallest improvement you care about, the minimum detectable effect or MDE. Here is what that gives per variant for the baselines and relative improvements that come up in practice.

Baseline conversion rateDetect +10%Detect +20%Detect +30%Detect +50%
0.5%327,92385,86339,88615,600
1%163,09642,69419,8287,751
2%80,68321,1109,7993,827
3%53,21213,9156,4562,519
5%31,2358,1593,7811,472
10%14,7523,8431,776687

Read one row slowly. A shop converting 2% of visitors that wants to know whether a new checkout is at least 20% better (2.4% instead of 2.0%) needs 21,110 visitors in the old checkout and 21,110 in the new one. That is 42,220 visitors for one answer. A shop with 100 visitors a month would need 35 years. A shop with 300,000 visitors a month has the answer in about four and a half days of traffic, which is why the same test is routine for one business and impossible for another. Nothing about the skill of the people involved changes that. The traffic does.

Two things about the table that people get wrong in the same direction. First, "detect +20%" means 20% relative, not 20 percentage points: going from 2% to 2.4% is a 20% lift. Second, if the real improvement is smaller than the MDE you sized for, the test is not "slightly underpowered". It is mostly noise, and the green tick a testing tool shows after a week is what I described in the statistics every marketer actually needs: a confidence interval that still contains zero, dressed up as a result.

Decisions inside a performance campaign

A performance campaign is not one decision. It is a stack of decisions, each on a different metric with its own sample. From the top of the funnel down: is this creative worth showing (click through rate, measured on impressions). Does this ad bring people who do anything (landing page conversion rate, measured on clicks). Does this ad set or audience pay (cost per acquisition or return on ad spend, measured on conversions). And should the budget move (marginal return, measured across several budget levels, which belongs to media mix modelling). The higher up the funnel, the more plentiful the data, which is why creative decisions come fast and budget decisions come slowly.

Click through rate is a proportion measured on impressions. To know an ad's CTR to within 20% of itself, set the margin of error from the first formula equal to 0.2p0.2p and solve for nn (sample size):

nimpressions=z0.9752 (1−p)r2 pn_{\text{impressions}} = \frac{z_{0.975}^2\,(1-p)}{r^2\,p}

where pp is the click through rate and rr the relative precision you want (0.2 for plus or minus 20%). At a 1% CTR that is about 9,508 impressions per ad, at 0.5% about 19,111 and at 2% about 4,706. At an £8 CPM, 9,508 impressions cost around £76. That is why you can know within days which of five creatives people click, and why that answer is close to worthless on its own: clicks are not customers.

Conversions behave like counts, and for a count of nn conversions the relative margin of error at 95% confidence is simply (confidence intervals)

MoEn≈z0.975n\frac{\text{MoE}}{n} \approx \frac{z_{0.975}}{\sqrt{n}}

where nn is the number of conversions behind the figure and z0.975≈1.96z_{0.975} \approx 1.96. This gives a bracket for every cost per acquisition you will ever look at:

Conversions behind the numberRelative margin of errorWhat the CPA figure actually means
10plus or minus 62%A £30 CPA could be £11 or £49. Not a decision.
25plus or minus 39%Direction only. Enough to kill a disaster, not to pick a winner.
50plus or minus 28%Roughly the volume Meta asks for before delivery settles. Keep or kill, nothing finer.
100plus or minus 20%A real decision between options that differ by 30% or more.
400plus or minus 10%Fine comparison. Budget moves between ad sets become defensible.
1,000plus or minus 6%Bidding and allocation optimisation. Most small businesses never get here per ad set.

This is the honest answer to "which ad is winning". With ten conversions each, two ads at £30 and £40 CPA are statistically the same ad. With a hundred each they are probably different. With four hundred each you can move budget with a clear conscience. The planner uses the same brackets: under 10 conversions is no signal, 10 to 29 a signal, 30 to 99 a direction, 100 to 399 a decision and 400 or more optimisation.

What the platforms themselves ask for

The platforms know this arithmetic, which is why their learning phases exist. Their current help pages say the following.

Meta says an ad set exits the learning phase once it delivers stably, which "usually occurs after about 50 results in the week after the ad set's last significant edit".1 TikTok's help centre no longer names a fixed conversion count. It says volatility "starts to decline after about 25 campaign results or 7 days" from when a campaign enters learning, and warns that pausing and certain edits retrigger it.2 Google recommends judging a Target CPA strategy over the last 30 days "including at least 30 conversions",3 and for Target ROAS on Search and Shopping it requires at least 15 conversions in the past 30 days, with higher floors for other campaign types (30 over 30 days for video action campaigns, 50 over 35 days per campaign for Demand Gen).4 LinkedIn has no published learning phase threshold, but an ad set needs an audience of at least 300 members and LinkedIn suggests at least 50,000 to drive results.5

Now put Meta's figure into money. About 50 results a week per ad set at a £30 CPA is £1,500 a week, roughly £6,500 a month, just to let the algorithm settle on one audience. Four ad sets and the platform's own guidance implies a monthly budget around £26,000. The learning phase is not a technicality. It is the platform telling you what the formulas tell you.

What spend that implies on each platform

The usual question is "how much do I need to spend". The honest structure of the answer is: conversions needed for the decision, times the cost of one conversion on that platform, in your industry, against your competition. The first number comes from the bracket table. The second varies so much that any single figure misleads, so here are the planning assumptions I start from, as illustrative UK figures for late 2026, before your own account data replaces them.

PlatformPlanning CPMPlanning CPCPlanning landing conversion ratePlanning cost per conversion
Meta (Facebook and Instagram)£5 to £12£0.40 to £1.501 to 3% for ecommerce£15 to £60
TikTok£3 to £8£0.20 to £1.000.5 to 1.5%£20 to £80
Google Searchn/a£0.80 to £4.00 broadly, far more in legal, finance, insurance and B2B software3 to 7%£20 to £120, several hundred in competitive B2B
Google Shopping and Performance Maxn/a£0.30 to £1.001.5 to 3%£15 to £50
LinkedIn£25 to £60£5 to £112 to 5% on a landing page, more on a lead form£60 to £250 per lead

Treat these as a starting assumption, not a promise. Your own numbers will differ, mostly because of the following, roughly in order of how much each moves the result.

Industry and competition come first. Cost per click is an auction: if insurers, law firms and venture funded software companies bid on your keywords, a click costs many times what it costs a florist, and nothing you do changes the other bidders. Product and price point come next. A £25 impulse purchase can convert from a cold audience, while a £3,000 service converts to a call and the call converts weeks later, so the conversion you can count in the platform is a different, earlier event with a different rate. Geography and language matter because bigger, more crowded markets usually cost more per click for the same intent. Creative and landing page quality is the factor you control, and it is why best practice matters before data does: a page that converts at 4% instead of 2% halves every cost per conversion in the table. Season matters too, since the fourth quarter auctions fill up with retail advertisers and costs rise with them, so a test sized in March and run in November needs more budget for the same conversions. And a new ad account with no conversion history pays a premium for its first weeks while the platform works out whom to show the ads to.

Now multiply. To reach a given number of conversions per variant, this is the money involved:

PlatformCost per conversion used30 conversions (direction)100 conversions (decision)400 conversions (optimisation)
Meta£30£900£3,000£12,000
TikTok£40£1,200£4,000£16,000
Google Search, consumer£50£1,500£5,000£20,000
Google Search, competitive B2B£200£6,000£20,000£80,000
LinkedIn, lead£150£4,500£15,000£60,000

Per variant. Compare two audiences on LinkedIn properly and that is £30,000. This is the figure missing from most growth proposals, and it explains why so many campaigns are stopped at exactly the moment they were about to produce an answer.

Time, and why a small daily budget is a slow clock

Budget and time are the same quantity seen from different ends. If you collect conversions at a steady rate, the days until you have enough are (ratios and rates)

Tdays=n⋅CBT_{\text{days}} = \frac{n \cdot C}{B}

where nn is the number of conversions the decision needs, CC the cost of one conversion and BB the daily budget. At a £30 cost per conversion:

Daily budgetDays to 30 conversionsDays to 100 conversionsDays to 400 conversions
£10903001,200
£3030100400
£100930120
£30031040
Bar chart of the days needed to collect 100 conversions at seven daily budgets: 300 days at £10 a day, 150 at £20, 100 at £30, 60 at £50, 30 at £100, 15 at £200 and 10 at £300. Tap or click to open full screen and move around

The shape is a hyperbola, and it matters. Going from £10 to £30 a day saves two hundred days. Going from £100 to £300 saves twenty. On a small budget every extra pound a day buys a lot of calendar time. On a large one it buys almost none, and the money is better spent on more variants.

Spending £10 a day is a reasonable way to keep an account warm and learn about creative. It is not a way to make conversion decisions this quarter, and "the campaign needs another two weeks" at that budget is a request to flip a coin a few more times. Three other clocks run alongside the arithmetic. Weekly cycles: B2B converts midweek, consumer ecommerce on Sunday evenings and paydays, so any decision needs at least one full week and preferably two, so that each variant has seen the same mix of days. Conversion lag: if people buy a week after clicking, the last week of any test is understated, so add the lag to the test length or you will kill the variant that was still converting. Learning resets: significant edits to budget, targeting or creative restart learning on Meta and TikTok,12 so a small budget edited every few days never leaves it. On a tight budget, set it and leave it for the full window.

Stopping a losing ad early on a tight budget

This is the useful question for a business with £300 a month rather than £30,000: are there statistically sound rules that let you cut a losing ad early and keep the money for the ones that work? Yes, and they come from three ideas.

The rule of three

If an ad has delivered nn clicks and zero conversions, the 95% upper bound on its true conversion rate is approximately (confidence intervals)

pupper≈3np_{\text{upper}} \approx \frac{3}{n}

where nn is the number of clicks and pupperp_{\text{upper}} the highest conversion rate still compatible with seeing none. Turned round, it gives a kill rule: if the ad needs a conversion rate pp to pay, kill it after 3/p3/p clicks without a conversion. If your landing page needs 2% for the ad to pay and an ad has delivered 150 clicks with no conversion, you can say with 95% confidence its true rate is below 2%.

Conversion rate the ad must reach to payClicks with zero conversions before you can kill it
0.5%600
1%300
2%150
3%100
5%60

At a £1 cost per click, that is £60 to £600 to kill an ad with confidence rather than a hunch. Compare that with the £3,000 it takes to crown a winner on Meta and you see the asymmetry: proving something bad costs far less than proving something best. A tight budget should spend most of its time proving things bad.

Sequential looks without fooling yourself

Classical tests assume you look once, at the end. Everybody actually looks every day, and every look is another chance for noise to cross the line. I simulated 100,000 tests with no real difference and checked them against the usual 5% threshold: looking 10 times produced a false "winner" in about 19% of them, and looking daily for 30 days in about 28%, instead of 5% (the logic of the test is in hypothesis tests). Sequential methods fix this by making the threshold strict early and relaxing it as data arrives. A common choice is the O'Brien and Fleming type spending function of Lan and DeMets: keep the overall α\alpha at 0.05, plan four equally spaced looks, and the nominal p value thresholds come out at roughly 0.0001 at the first look, 0.005 at the second, 0.02 at the third and 0.04 at the last (my own simulation of one million null tests, which spends exactly the 5%). Stopping early is still allowed when a result is overwhelming. Stopping on a result that merely looks nice is not.

Letting the budget follow the evidence

For creative tests on small budgets, Thompson sampling6 usually beats a fixed split. Each ad's conversion rate is described by a Beta distribution (distributions):

pi∼Beta(a0+ci, b0+ki−ci)p_i \sim \text{Beta}(a_0 + c_i,\ b_0 + k_i - c_i)

where pip_i is ad ii's unknown conversion rate, cic_i its conversions so far, kik_i its clicks so far, and Beta(a0,b0)\text{Beta}(a_0, b_0) the prior, what you believe before the first click. For every unit of budget, draw one value from each ad's distribution and give the budget to the ad with the highest draw. Ads that are clearly losing stop being drawn and stop being funded, and ads that are close keep being compared. The prior is where expertise comes in: if pages like yours convert around 2%, a prior worth 200 clicks and 4 conversions, Beta(4,196)\text{Beta}(4, 196), stops the algorithm over reacting to one lucky day.

To show what this does, I simulated four ads whose true conversion rates are 2.6%, 2.2%, 1.5% and 1.0%, with 50 clicks a day for six weeks, that prior, and 2,000 repetitions.

Simulated share of clicks for four ads with true conversion rates of 2.6, 2.2, 1.5 and 1.0% at 50 clicks a day, averaged over 2,000 runs. The best ad rises from 30 to 51% of clicks by week six, the second holds about 30%, and the two weaker ads fall from 23 and 20% to 12 and 7%. Tap or click to open full screen and move around

Over the six weeks the two weaker ads received 28% of the clicks instead of the 50% a fixed four way split gives them, so 44% less money went to the losers, while the leading pair kept being compared. The budget never had to reach 100 conversions each for all four, because two of them never deserved it.

What these rules cannot do

They cannot turn a small budget into a big one. They stop you wasting money on obvious losers and keep two or three candidates alive for longer. They do not let you tell a 2.0% ad from a 2.2% ad on £10 a day. That remains a question for a bigger budget or for a different kind of decision, which brings us to the website.

How many tests you can run at once

Two things cap how many tests run at the same time: traffic and false positives.

Traffic first. Every A/B test needs two arms of the size in the first table, and every multivariate cell needs the same again. The number of tests that fit into a window on independent parts of the funnel is at most (sample size)

kmax⁡=⌊V⋅w/4.33a×nper arm⌋k_{\max} = \left\lfloor \frac{V \cdot w / 4.33}{a \times n_{\text{per arm}}} \right\rfloor

where VV is monthly visitors, ww the window in weeks, 4.334.33 the average number of weeks in a month, aa the number of arms or cells per test and nper armn_{\text{per arm}} the visitors per arm from the table. At 300,000 visitors a month, a 2% baseline and a 20% MDE, a two week window holds about 138,600 visitors: room for three A/B tests, or one 2 by 2 multivariate test with four cells. At 30,000 visitors a month the same window holds nothing, and a four week window holds one A/B test only if you look for a 30% lift rather than 20%. At 3,000 visitors a month a single 30% test takes 29 weeks, and the arithmetic has effectively stopped.

Then false positives. Each test at α=0.05\alpha = 0.05 has a 5% chance of a false winner, and with several running the chance that at least one lies to you is (hypothesis tests)

P(at least one false winner)=1−(1−α)kP(\text{at least one false winner}) = 1 - (1-\alpha)^k

where α\alpha is the false positive rate of each test and kk the number of independent tests.

Tests runningChance of at least one false winner
15%
210%
314%
523%
1040%

Run ten tests a month without correcting for this and a real share of the "improvements" you ship will be noise, and you will wonder why the overall conversion rate never moves. The fix is a correction: Bonferroni divides α\alpha by kk (strict and simple), Holm's method does the same stepwise and wastes less power, and for large programmes the false discovery rate approach of Benjamini and Hochberg controls the share of shipped winners that are false rather than policing each test. Whichever you pick, the stricter threshold needs more visitors per test, which feeds back into the traffic cap.

Interaction is the third limit. Tests on the same step of the funnel are not independent: a new headline and a new price on the same page can each win alone and lose together, which is the point of why testing one thing at a time is costing you money. The practical rule is that tests on different steps (ad creative, landing page, checkout, email) can overlap if visitors are randomised independently into each, while tests on the same step must be one factorial design or run one after the other.

What the website needs before you can test on real user data

The first table is the whole answer to "how many visitors do I need", but it helps to translate it into the situations that actually occur. All figures assume a 2% baseline and the same formula the planner uses, counting whole weeks.

Monthly visitorsMonthly orders at 2%What is honestly possible
1002Nothing statistical. Two orders a month cannot tell you anything about a change. This is a best practice and direct sales business (see the section on channels).
1,00020No purchase tests. A micro conversion with a 10% base rate (add to basket, a click to the contact page) can catch a 50% lift in about six weeks, a 30% lift in about four months.
10,000200One A/B test at a time: about nine weeks for a 30% lift, nineteen weeks for 20%. Choose tests with big expected effects: offer, price presentation, checkout flow.
50,0001,000One properly sized test for a 20% lift every four weeks, or two in parallel on different steps. Enough for a real programme.
300,0006,000Three concurrent A/B tests or a 2 by 2 multivariate test every two weeks, with corrections for multiple comparisons. Segment level tests become possible.
1,000,000 and above20,000A full programme: factorial designs, holdouts, bandits for personalisation and tests on 5 to 10% lifts. The constraint is now ideas and engineering, not data.

Multivariate tests multiply the demand: a 2 by 2 test has four cells and needs four times nper armn_{\text{per arm}}, and a 3 by 3 needs nine. Fractional factorial designs test more factors with fewer cells by giving up some interactions, a trade worth making at 50,000 visitors and rarely at 10,000. Four hygiene rules apply whatever the traffic: run for whole weeks; never peek and stop without a sequential plan; check the sample ratio (if 52% of visitors landed in variant B when you split 50 to 50, the randomisation is broken and so is the test); and keep tests short, because browsers clear the identifiers that keep returning visitors in the same arm. Safari, for example, caps cookies set by script at seven days,7 so a long test slowly mixes its arms.

One caveat on micro conversions, the usual escape hatch for low traffic sites. Add to basket at 10% needs a fraction of the traffic of a purchase test. But the thing you care about is purchases, and a change that lifts basket adds by 20% while making checkout slightly worse can lose money. A micro conversion is a reasonable proxy for a test about the step it measures and a poor proxy for anything downstream of it.

When you do not have the numbers yet, expertise is the data

The tables make it look as though a business with 1,000 visitors a month can only guess. That is wrong, for a precise statistical reason.

Every decision uses a prior. In Bayesian terms, the number you act on is a weighted average of what your data says and what you believed before, weighted by how much data you have. With 40,000 visitors in a test the prior barely matters. With 400 it is most of the answer. The question for a small business is not "data or no data" but "whose prior". A prior built from years of accounts, landing pages, checkouts and pricing pages across ecommerce, software, hospitality and professional services is a very different starting point from one built on a blog post and a hunch, and you can put a number on it. Expressing "pages like yours convert at roughly 2%" as Beta(9,391)\text{Beta}(9, 391) (centred on 2.25%) gives that belief the weight of 400 visitors. Your first 400 visitors move the estimate halfway from that belief towards whatever they actually do, and your first 4,000 all but replace it. That is how a business with no data gets decisions better than chance from day one, and gets them replaced by its own evidence as soon as the evidence exists.

The hierarchy of evidence for a small business, best first:

  1. Your own controlled test, sized as above. Rare at small traffic.
  2. Your own observational data read carefully: cohorts, funnels, repeat rates, with the margins from the bracket table attached. Available to anyone with clean tracking and a few hundred orders.
  3. Category benchmarks adjusted for your price point and geography.
  4. Best practice from businesses shaped like yours: what has worked repeatedly on sites and campaigns of the same kind. This is what experience actually is.
  5. Opinion, including mine and including yours.

A lot of marketing for small businesses runs on level five while describing itself as level one. Good work runs on levels two to four and says so. Concretely, the first months with a small shop or an early software product are about tracking that counts each order once and attributes it sensibly. The conversion fundamentals known to work before any test confirms them (one clear offer per page, a checkout with no surprises, pricing presented the way the category expects, a contact path that works on a phone). Channels that cost time rather than money. And a small paid budget run with the kill rules above, so it teaches you something without pretending to decide anything. The goal is to grow steadily into the traffic band where real testing becomes possible and to have clean data when you get there. Each step up the website table unlocks a faster way to learn, which is why growth compounds once it starts.

Index of the conversion rate, month 0 equals 100, assuming one win in three tests and a 6% real lift per win. One completed test a month reaches 127 after 12 months, three tests reach 201 and six tests reach 390. Tap or click to open full screen and move around

The curves follow one formula (compounding):

Im=100 (1+t⋅w⋅g)mI_m = 100\,(1 + t \cdot w \cdot g)^m

where ImI_m is the index after mm months, tt the completed tests per month, ww the share of tests that win (one in three here) and gg the real lift per win after shrinking for the winner's curse (6% here). One test a month compounds to about 27% in a year. Three tests a month double the conversion rate. Six come close to four times the starting rate. The curves are not promises; they show what test velocity does, and test velocity is set by traffic and budget. That is the whole argument for growing traffic first, and for being honest about which curve you are on.

Is performance marketing right for your stage and budget?

Before any of the above comes a question that deserves a formula rather than a feeling. Paid acquisition pays when the cost of a customer is comfortably below what that customer is worth, and when you can afford the learning budget to find out. The ceiling on what you can pay per customer is (lifetime value)

CPAmax⁡=AOV×m×fR\text{CPA}_{\max} = \frac{\text{AOV} \times m \times f}{R}

where AOV is the average order value, mm the contribution margin as a share, ff the expected orders per customer over the period you count, and RR the target ratio of lifetime value to acquisition cost (3 to 1 is the usual convention).

For an invented Leeds hamper shop, hampers.example, with a £68 median basket, a 45% contribution margin and 1.6 orders per customer over a year, aiming for 3 to 1:

CPAmax⁡=68×0.45×1.63≈16\text{CPA}_{\max} = \frac{68 \times 0.45 \times 1.6}{3} \approx 16

So about £16. Meta at £15 to £60 per conversion is marginal at best, and so is Google Shopping. The shop's first job is not more ads but a higher basket, a higher repeat rate or a higher margin, and only then ads. For an invented Manchester software product at £2,000 a year with an 80% gross margin and a 30 month average life, the ceiling is 2000×0.8×2.5/3≈13332000 \times 0.8 \times 2.5 / 3 \approx 1333 pounds, and LinkedIn leads at £150 that close one in five cost £750 per customer, comfortably inside it.

Then there is the learning budget, the line nobody puts in the plan. The planner works with two figures: 50 conversions to learn whether a channel works at all, and 200 (two approaches at 100 each) to compare within it. At the planner's illustrative cost per conversion that gives:

Business typePlanning cost per conversionLearning budget (50 conversions)Comparison budget (200 conversions)
Ecommerce£30£1,500£6,000
Local trades and services£40£2,000£8,000
Hospitality£35£1,750£7,000
Professional services£120£6,000£24,000
B2B software£150£7,500£30,000
Manufacturing and industrial£180£9,000£36,000

If that budget is more than you can lose without regret, performance marketing is the wrong first channel, however good the creative.

The alternatives when paid is not yet right

For many businesses, particularly early B2B ones, the alternative is to go to customers directly, and the options are more specific than "do some outreach".

LinkedIn Sales Navigator and personal outreach cost £94.99 a month per seat on the Core plan, or £959.88 a year,8 plus your time. That suits anything with a contract value above roughly £5,000 and a definable buyer: software, consultancy, agencies, manufacturing, recruitment. It does not suit consumer products, or anything below about £1,000 a year, where the time per conversation never pays back.

Email outreach depends on the country. In the UK, the Information Commission (formerly the ICO, renamed under the Data (Use and Access) Act 2025)9 says you can email or text any corporate body without consent, as long as you identify yourself and offer an opt out. Sole traders and some partnerships count as individuals and need consent.10 In Germany, the Unfair Competition Act treats advertising by email without the recipient's prior express consent as an unacceptable nuisance, with no exception for businesses.11 Austria's Telecommunications Act 2021 likewise forbids direct marketing email without prior consent.12 So in Germany and Austria the same tactic becomes LinkedIn, a carefully judged phone call, a letter or an event. Do not let a UK playbook run in German speaking markets.

Referral and partner programmes are the lowest cost channel for professional services and agencies, and the slowest to build. I run one myself. The terms are on the partner programme pages. Search is for intent only: Google Search works when people already look for what you sell, but it does not create demand, and a new category with no search volume needs Meta, TikTok or LinkedIn to make people want it, with the budgets above. Marketplaces and directories (Amazon, Etsy, booking platforms, comparison sites) let you rent traffic for a commission, which at small scale is usually a better deal than building your own paid acquisition. Content and SEO take six to twelve months to pay and then have a low marginal cost, right for anything people search for and wrong as the only channel for a business that needs revenue this quarter.

By type of business, the channels that are usually relevant and the ones that usually are not worth a learning budget first:

BusinessUsually relevantUsually not worth the learning budget first
Consumer ecommerce, basket under £100Meta, TikTok, Google Shopping, marketplaces, email to existing customersLinkedIn, search on generic terms
B2B softwareGoogle Search on problem keywords, LinkedIn, outbound via Sales Navigator, partnerships, contentTikTok, Meta beyond consented retargeting
Professional services and consultanciesReferrals, LinkedIn organic and outbound, SEO, speakingAny paid channel until referrals are systematic
Local trades and servicesGoogle Maps and local search, Meta locally, reviewsLinkedIn, TikTok, national search
Hospitality and tourismMeta, Google Hotel ads and Search, booking platforms, email to past guestsLinkedIn, generic TikTok
Manufacturing and industrialOutbound, trade shows, LinkedIn, SEO on product terms, distributorsMeta, TikTok

Take the shape rather than the individual cells, which move every year: the fewer and more valuable your customers, the more the right first channel is a person, and the more numerous and less valuable they are, the more it is an auction. Performance marketing is a volume instrument. If you have twelve customers a year and need twenty, it is the wrong instrument, and data science will not make it the right one.

Forecasting: who can do it, and on how much data

Forecasting is where the gap between promise and data is widest, because a forecast looks the same whether it is built on three years of weekly history or on a spreadsheet of wishes. So here is what each amount of history buys, as the rules of thumb I work to, in weekly data unless noted.

What you want to forecastMethodMinimum historyComfortable history
Next month's level, no seasonalityMoving average or simple exponential smoothing8 weeks13 weeks
Level and trendHolt's method, regression on time13 weeks26 weeks
Seasonality within the yearSeasonal decomposition, Holt Winters, seasonal regression2 full years (104 weeks)3 full years
Autocorrelated series with shocksARIMA (Box and Jenkins)about 50 observations100 or more
The effect of ad spend on salesRegression with adstocked spend, ARIMAX20 to 30 weeks with real variation in spend52 weeks or more with variation
Daily demand with a weekday patternAny of the above on daily data8 weeks (56 points)26 weeks

The planner maps history to the same ladder: under six months supports a level, six to twelve months a trend, one to two years a seasonal forecast, and a seasonal forecast with a campaign term only with one to two years of history plus over a year of campaign data with varying spend, or with over two years of history plus at least three months of campaign data.

Two years for seasonality is the requirement that catches most young businesses. Without two Decembers you cannot separate "December is big" from "we grew" (the mechanics are in seasonality in the Foundations). With two you can estimate it. With three you can estimate it with a margin you would show a bank. Daily data gives more points but each is noisier, so it buys less than the count suggests. A weekly series of 104 points is usually a better basis than a daily series of 180.

Accuracy should be judged as mean absolute percentage error on a holdout, never as a fit statistic on the training data. These are the planning bands I use:

SituationError at a four week horizonWhat that means for planning
Mature business, 3 or more years, stable channels5 to 12% monthly, 10 to 20% weeklyPlan stock and cash against the band and it holds most months
1 to 2 years of history, some channel changes12 to 25%Plan against the lower edge, treat the upper edge as upside
Under a year of history, or a channel mix that changed last quarter25 to 40% or worseA direction, not a plan. Re forecast monthly and say so.
New product, no historyNo model. Benchmarks and a prior.Scenarios, not a forecast.

The error band also widens with the horizon. For a series that behaves like a random walk with drift, the forecast error grows with the square root of the horizon (spread):

σh=σ1h\sigma_h = \sigma_1 \sqrt{h}

where σ1\sigma_1 is the standard error one step ahead and σh\sigma_h the standard error hh steps ahead. A forecast 16 weeks out carries twice the uncertainty of one 4 weeks out, and a twelve month forecast from weekly data is honest only as a wide band. A single line twelve months out without a band is a drawing. The full build, including the holdout and the band, is in forecasting demand: autocorrelation, seasonality and honest error bands, and the wider map of methods in the complete guide to forecasting.

Factoring current campaigns into a forecast

This is the common request that most often cannot be met honestly, for a reason that has nothing to do with modelling skill. To estimate what a pound of ad spend adds to sales, the model must have seen different levels of spend. If you spent £1,000 a month, every month, for a year, the data contains no information about what £2,000 would do, and no technique can extract it. The requirements I set before putting a campaign term into a forecast are three. First, at least 20 to 30 weekly observations in which spend actually varied, with a coefficient of variation of roughly 0.3 or more. Weeks of zero spend are gold, because they show the baseline. Second, about 10 to 20 observations per parameter the model estimates: one channel with adstock and a saturation curve is three parameters, so 30 to 60 weeks, and three channels is nine parameters, roughly two years. Third, spend that was not itself set by sales. If the budget rule was "spend 10% of last month's revenue", sales cause spend as much as spend causes sales, and a single regression overstates the ad effect. The fix is in simultaneous equations, and it needs even more data.

Where that history does not exist, there are two honest alternatives. An incrementality test switches a channel off in one region or for one audience for four weeks and compares, which needs comparable regions and at least a few hundred conversions in the period. Or a prior return on ad spend from benchmarks and experience goes into the forecast as a scenario with a wide band, not as a coefficient. Both are legitimate. Pretending the coefficient came out of your data when it came out of someone's head is not.

For "what will next quarter's campaign do", the rule I work to is three comparable campaigns or 12 weeks of continuous spend with variation, whichever is more. Below that, a campaign forecast is a benchmark range narrowed by your creative's early CTR: useful, but not a forecast of your campaign. Above a year of varied spend, the campaign term becomes an estimated coefficient with an interval, and the forecast can show what pulling budget forward into October does to December. That is the version worth paying for, and it can only be bought with history.

What you have, and what can honestly be done with it

Pulling it together, this is the table I would rather show before a contract than discover after one. It is also the logic of the planner on my data science page, which places you on it from your visitors and orders: both thresholds must be met to move up a stage.

StageWhat you haveWhat is honestWhat I do with you
FoundationsUnder 3,000 visitors a month or fewer than 50 orders, little or no historyNo statistical tests. Forecasts are scenarios. Paid media is for learning, with kill rules, not for deciding.Tracking that counts once. Conversion fundamentals applied without waiting for proof. Channels that cost time. A small paid budget run on brackets. Target: the next stage within two to four quarters.
Directional3,000 to 20,000 visitors and 50 to 400 orders, 6 to 12 months of dataLarge effect tests, often on micro conversions. Cohort and funnel analysis with margins attached. Level and trend forecasts with wide bands. Paid decisions at the 30 to 50 conversion bracket.A monthly read of the numbers with intervals, not a dashboard of averages. One large test at a time. Creative tests with Bayesian allocation. A first seasonal forecast once two cycles exist.
Testing20,000 to 100,000 visitors and 400 to 2,000 orders, over a year of dataOne properly sized A/B test every few weeks. Seasonal forecasts at 12 to 25% error. Paid decisions at 100 conversions per variant.A test roadmap ranked by expected value. Sequential testing so you can stop early honestly. Campaign terms in the forecast where spend has varied. Budget moves between channels with an interval on each.
Programme100,000 visitors and 2,000 orders a month or more, two or more years of varied spendConcurrent tests with multiple comparison control. Factorial designs. Media mix modelling. Forecasts at 5 to 12% monthly error with campaign scenarios.The full experimentation and forecasting system: tests on several funnel steps at once, bandits for allocation, incrementality tests, a forecast that answers budget questions, and the steepest growth curve the traffic allows.

The stages are not a judgement. Most small businesses start in the first row, and the work is to get them to the third. What the table does is stop a project pretending to be in row four while the data is in row one, because that pretence is where budgets vanish and trust goes with them. Your raw data stays yours, and the terms for code and models are in the analytics and data science terms.

If you want this for your own numbers, the free workshop is where we put your visitors, orders, history and budget through exactly this arithmetic and agree which row you are in, what the next row needs and what it costs to get there, and the first two weeks of work after it are on me. The commercial terms are on the pricing page. Nothing in the workshop assumes you already have the data. It assumes you want to know what you can do with the data you have.

Questions to ask anyone who promises you growth

How many conversions did the recommended decision rest on, and what is the margin of error at that count? What minimum detectable effect is this test sized for, and how many visitors per variant does that need at our baseline? At our daily budget and cost per conversion, how many days until that number of conversions exists? What is the kill rule for a losing ad, and after how many clicks does it trigger? How many tests are running at once, and what correction is applied for multiple comparisons? What is the learning budget for this channel before we know whether it works, and what happens if we cannot afford it? How many weeks of history is this forecast built on, what was its error on a holdout, and does the band widen with the horizon? Did our ad spend vary enough for the model to estimate its effect, or is the campaign coefficient a benchmark? And which row of the table are we in, and what would move us to the next one?

If you want this arithmetic run on your own numbers rather than on my examples, book the free workshop: we put your visitors, orders, history and budget through these formulas and agree which stage you are in and what the next one needs, and the first two weeks of work are on me. For a first answer right now, the planner on my data science page runs the same calculation in your browser.

Sources

  1. 1.

    Meta Business Help Centre, About the learning phase, https://www.facebook.com/business/help/112167992830700

  2. 2.

    TikTok Ads Manager Help Centre, About the learning phase, https://ads.tiktok.com/help/article/learning-phase

  3. 3.

    Google Ads Help, About Target CPA bidding, https://support.google.com/google-ads/answer/6268632

  4. 4.

    Google Ads Help, About Target ROAS bidding (conversion requirements by campaign type), https://support.google.com/google-ads/answer/6268637

  5. 5.

    LinkedIn Help, Audience size requirements in Campaign Manager, https://www.linkedin.com/help/lms/answer/a423690

  6. 6.

    William R. Thompson, On the likelihood that one unknown probability exceeds another in view of the evidence of two samples, Biometrika 25, 1933, https://doi.org/10.2307/2332286

  7. 7.

    John Wilander, Intelligent Tracking Prevention 2.1, WebKit blog, February 2019, https://webkit.org/blog/8613/intelligent-tracking-prevention-2-1/

  8. 8.

    LinkedIn Sales Solutions, Compare Sales Navigator plans and pricing (GBP: Core £94.99 a month or £959.88 a year), https://business.linkedin.com/sales-solutions/compare-plans

  9. 9.

    Data (Use and Access) Act 2025, section 118, https://www.legislation.gov.uk/ukpga/2025/18/section/118

  10. 10.
  11. 11.

    Gesetz gegen den unlauteren Wettbewerb (UWG), § 7 Abs. 2 Nr. 2, https://www.gesetze-im-internet.de/uwg_2004/__7.html

  12. 12.

    Telekommunikationsgesetz 2021, § 174 Abs. 3, Rechtsinformationssystem des Bundes, https://ogd.ris.bka.gv.at/Dokumente/Bundesnormen/NOR40238632/NOR40238632.html