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

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 that was sized for "a test campaign" rather than for the number of observations a decision needs. From that moment the project is quietly running on opinion while everybody keeps saying "data".

I have sat in that meeting more times than I would like to admit, on both sides of the table, 15+ years of growth, performance marketing and data science work. The disconnect is never about effort or talent. It is about nobody having done 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 with every client before anything is signed, so that we all know what to expect and what can be done.

It is long on purpose. Skim the tables if you are in a hurry, but read the section on expertise before you decide that your numbers are too small to bother, because that is exactly where most of my work with smaller businesses happens. And 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 stages below you are in.

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 $n$ observations has a margin of error, and at 95% confidence that margin is

\text{MoE} = z_{0.975}\sqrt{\frac{p(1-p)}{n}} \approx 1.96\sqrt{\frac{p(1-p)}{n}}

where $p$ is the proportion you are estimating and $n$ is how many visitors, clicks or impressions it was measured on. The important thing to notice is 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.

The second formula is the one that sizes a test. If you want to detect a difference $\delta$ between two conversion rates $p_1$ and $p_2$ with a false positive rate $\alpha$ and power $1-\beta$ , each arm of the test needs roughly

n_{\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}

With the conventional $\alpha = 0.05$ and power of 80%, $z_{1-\alpha/2} = 1.96$ and $z_{1-\beta} = 0.84$ , so the constant $2(1.96 + 0.84)^2$ is about 15.7. Everything else is driven by your baseline rate and by the smallest improvement you care about, usually called the minimum detectable effect or MDE. Here is what that produces, per variant, for the baselines and relative improvements that actually come up in practice.

Baseline conversion rate	Detect +10%	Detect +20%	Detect +30%	Detect +50%
0.5%	327,922	85,863	39,886	15,600
1%	163,096	42,694	19,828	7,751
2%	80,683	21,110	9,798	3,826
3%	53,211	13,915	6,455	2,518
5%	31,235	8,159	3,781	1,472
10%	14,752	3,842	1,775	687

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 can have the answer in about four 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 more things about the table, because people get them 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. Going from 2% to 22% would be a 1,000% lift and nobody has ever seen one. 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 your testing tool shows after a week is the thing I wrote about in the statistics every marketer actually needs: a confidence interval that still contains zero, dressed up as a result.

Layer one: decisions inside a performance campaign

Performance campaigns are not one decision. They are a stack of decisions, each sitting on a different metric, and each metric has its own sample. From the top of the funnel down:

Is this creative worth showing? The metric is click through rate, and the sample is impressions.
Does this ad bring people who do anything? The metric is landing page conversion rate or cost per conversion, and the sample is clicks.
Does this ad set or audience pay? The metric is cost per acquisition or return on ad spend, and the sample is conversions.
Should the budget move? The metric is marginal return, and the sample is conversions across several budget levels, which is a different thing again and 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. Let me put numbers on each.

Click through rate is a proportion measured on impressions. To know an ad's CTR to within plus or minus 20% of itself, you need $n = 1.96^2\,p(1-p)/(0.2p)^2$ impressions, which at a 1% CTR is about 9,500 impressions per ad, at 0.5% about 19,100, and at 2% about 4,700. On Meta at a £8 CPM, 9,500 impressions cost around £76. That is why I can tell you within a few 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 the relative margin of error at 95% confidence is simply

\frac{\text{MoE}}{n} \approx \frac{1.96}{\sqrt{n}}

This is the formula I use most in client meetings, because it gives you a bracket for every cost per acquisition you will ever look at:

Conversions behind the number	Relative margin of error	What the CPA figure actually means
10	plus or minus 62%	A £30 CPA could be £11 or £49. Not a decision.
25	plus or minus 39%	Direction only. Good enough to kill a disaster, not to pick a winner.
50	plus or minus 28%	The platforms' own learning threshold. Keep or kill, nothing finer.
100	plus or minus 20%	A real decision between options that differ by 30% or more.
400	plus or minus 10%	Fine comparison. Budget moves between ad sets become defensible.
1,000	plus or minus 6%	Bidding and allocation optimisation. Most small businesses never get here per ad set.

This bracket table 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 platforms know this, which is why their learning phases exist. At the time of writing, Meta asks for roughly 50 optimisation events per ad set within seven days before it leaves learning, TikTok for about 50 conversions in the same window, Google's Smart Bidding has historically wanted around 30 conversions in 30 days before target CPA behaves and more before target ROAS does, and LinkedIn has no formal learning phase but needs an audience of at least a few hundred and a meaningful daily budget before its delivery stabilises. Notice that 50 conversions in a week per ad set is an enormous number for a small business. At a £30 CPA that is £1,500 a week per ad set, or £6,500 a month, just to let the algorithm settle on one audience. Run four ad sets and the platform's own recommendation is a £26,000 monthly budget. The learning phase is not a technicality. It is the platform telling you the same thing the formulas do.

Layer two: what spend that implies on each platform

The question I am asked most 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 above. The second varies so much that any single figure is a lie, so I will give you ranges and then explain exactly what moves a business within them.

Platform	Typical CPM	Typical CPC	Typical landing conversion rate	Typical cost per conversion
Meta (Facebook and Instagram)	£5 to £12	£0.40 to £1.50	1 to 3% for ecommerce	£15 to £60
TikTok	£3 to £8	£0.20 to £1.00	0.5 to 1.5%	£20 to £80
Google Search	n/a	£0.80 to £4.00 broadly, £8 to £40 in legal, finance, insurance and B2B software	3 to 7%	£20 to £120, several hundred in competitive B2B
Google Shopping and Performance Max	n/a	£0.30 to £1.00	1.5 to 3%	£15 to £50
LinkedIn	£25 to £60	£5 to £11	2 to 5% on a landing page, 8 to 13% on a lead form	£60 to £250 per lead
Microsoft Ads	n/a	20 to 30% below Google Search	Similar to Google Search	Similar, slightly lower

These are UK and DACH ranges from accounts I have worked on and from published benchmarks, as of late 2026, and they are illustrative. Yours will differ, and here is why, in roughly the order of how much each factor moves the number:

Industry and competition. Cost per click is an auction. If insurers, law firms and SaaS companies with venture money bid on your keywords, a click costs ten times what it costs a florist. Nothing you do on your side changes the other bidders.
Product and price point. A £25 impulse purchase converts at 3% from a cold audience. A £3,000 service does not; it converts to a call, and the call converts weeks later. The "conversion" you can count in the platform is a different, earlier event, with a different rate.
Geography and language. UK English clicks cost more than Austrian German clicks for the same intent, because more advertisers compete for them. DACH LinkedIn inventory is smaller and often more expensive per click than the UK's.
Creative and landing page quality. This one you do control, and it is why best practice matters before data does. A landing page that converts at 4% instead of 2% halves every cost per conversion in the table.
Season. Fourth quarter CPMs on Meta routinely run 30 to 60% above spring. A test sized in March and run in November needs a bigger budget for the same number of conversions.
Account history. A new ad account with no conversion history pays a premium for the first weeks while the platform has no idea whom to show your ads to.

Now multiply. To reach 100 conversions per variant, the point at which two options are genuinely comparable, you are looking at this kind of money per variant:

Platform	Cost per conversion used	30 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 that is missing from nearly every growth proposal I have ever read, and it is the figure that explains why so many campaigns are stopped at exactly the moment they were about to produce an answer.

Layer three: time, and why a small daily budget is a slow clock

Budget and time are the same quantity seen from different ends. If you spend $B$ per day and a conversion costs $C$ , you collect $B/C$ conversions a day, and reaching the $n$ conversions your decision needs takes

T_{\text{days}} = \frac{n \cdot C}{B}

At a £30 cost per conversion:

Daily budget	Days to 30 conversions	Days to 100 conversions	Days to 400 conversions
£10	90	300	1,200
£30	30	100	400
£100	9	30	120
£300	3	10	40

The chart is a hyperbola, and that shape matters. Going from £10 to £30 a day saves you two hundred days. Going from £100 to £300 saves you twenty. For a small budget, every extra pound a day buys a lot of calendar time; for a large one, it buys almost none, and you are better off spending the extra on more variants.

Spending £10 a day is a perfectly reasonable way to keep an account warm and learn about creative. It is not a way to make conversion decisions this quarter, and anyone who tells you the campaign "needs another two weeks" at that budget is asking for a coin to be flipped a few more times. Three other clocks also run alongside the arithmetic:

Weekly cycles. B2B converts on Tuesday to Thursday, consumer ecommerce on Sunday evenings and paydays. Any decision needs at least one full week, and I prefer 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 always understated. The platforms' attribution windows hide this rather than solve it. Add the lag to the test length, or you will kill the variant that was still converting.
Learning phase resets. Every significant edit to budget, targeting or creative restarts the learning phase on Meta and TikTok. A small budget edited every few days never leaves learning. On a tight budget, set it and leave it alone for the full window.

Layer four: brackets for a tight budget, and how to stop spending on a loser early

This is the question I enjoy, because the answer is genuinely useful 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 you have seen $n$ clicks and zero conversions, the 95% upper bound on the true conversion rate is approximately $3/n$ . That gives a clean kill rule. If your landing page needs a 2% conversion rate for the ad to pay, and an ad has delivered 150 clicks with no conversion, you can say with 95% confidence that its true rate is below 2%. Kill it. The thresholds:

Conversion rate the ad must reach to pay	Clicks 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 £300 to kill an ad with confidence rather than with a hunch. Compare that to the £3,000 it takes to crown a winner, and you see the asymmetry: proving something is bad costs far less than proving something is best. A tight budget should spend most of its time proving things bad.

Sequential testing. 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. Looking daily for a month at a test with no real difference gives you something like a 25 to 30% chance of a false "winner" instead of 5%. Sequential methods fix this by making the threshold stricter early and relaxing it as data arrives. The simplest version I use with clients is a spending function: set the overall $\alpha$ at 0.05, allow yourself say four looks, and use roughly $\alpha = 0.001$ at the first look, $0.004$ at the second, $0.019$ at the third and the remaining $0.043$ at the last. Early stopping is still allowed when a result is overwhelming. Stopping on a result that merely looks nice is not.

Bayesian allocation. For creative tests on small budgets, I have stopped using fixed splits and started using Thompson sampling. Each ad's conversion rate is a Beta distribution, $\text{Beta}(\alpha_0 + \text{conversions},\ \beta_0 + \text{clicks} - \text{conversions})$ . Every day, draw one value from each ad's distribution and give tomorrow's budget to the ad with the highest draw. Ads that are clearly losing stop being drawn and stop being funded, automatically, and ads that are close keep being compared. The prior $\text{Beta}(\alpha_0, \beta_0)$ is where expertise comes in: if I know from similar accounts that landing pages like yours convert around 2%, a prior equivalent to 200 clicks and 4 conversions stops the algorithm from over reacting to the first lucky day. Over a month this typically spends 30 to 50% less on losers than a fixed split, and that saving is real money for a small account.

The lines are four ads in a simulated run on The Gift Bow, my demo hamper shop, with the two losers starved within three weeks while the leading pair kept being compared. The budget never needed to reach the "100 conversions each" level for all four, because two of them never deserved it.

What these brackets cannot do is turn a small budget into a big one. They stop you wasting money on obvious losers and they let you 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 altogether, which brings me to the website.

Layer five: how many tests can you run at once

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

Traffic. Every A/B test needs $2 \times n_{\text{per arm}}$ visitors from the table at the top, and every multivariate cell needs the same again. If your monthly traffic is $V$ and a test should finish within a $w$ week window, the number of tests you can run at once on independent parts of the funnel is at most

k_{\max} = \left\lfloor \frac{V \cdot w / 4.33}{\text{arms} \times n_{\text{per arm}}} \right\rfloor

At 300,000 visitors a month, a 2% baseline and a 20% MDE, a two week window gives you about 138,000 visitors and 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 supports nothing, and a four week window supports one A/B test only if you are willing to look for a 30% lift rather than 20. At 3,000 visitors a month the arithmetic simply stops.

False positives. Each test at $\alpha = 0.05$ has a 5% chance of a false winner. Run $k$ at once and the chance that at least one of them lies to you is $1 - 0.95^k$ :

Tests running	Chance of at least one false winner
1	5%
2	10%
3	14%
5	23%
10	40%

Run ten tests a month without correcting for this and you will ship four false improvements a year and wonder why the overall conversion rate never moves. The fix is a correction: Bonferroni divides $\alpha$ by $k$ (strict, 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 lets you say "no more than 10% of our shipped winners are false" rather than policing each test individually. Whichever you pick, the stricter threshold needs more visitors per test, which feeds back into the traffic cap.

Interaction. 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 whole point of why testing one thing at a time is costing you money. The practical rule: tests on different steps (ad creative, landing page, checkout, email) can overlap if visitors are randomised independently into each. Tests on the same step must either be one factorial design or run one after the other.

Layer six: what the website needs before you can optimise on real user data

The table at the top is the whole answer to "how many visitors do I need", but it is worth translating into the situations I actually meet.

Monthly visitors	Baseline 2%, monthly orders	What is honestly possible
100	2	Nothing statistical. Two orders a month cannot tell you anything about a change. This is a best practice and direct sales business, see the next section.
1,000	20	Still no conversion tests. You can test a micro conversion with a higher base rate (add to basket at 10 to 15%, or a click to the contact page) and catch only very large effects over two to three months.
10,000	200	One A/B test at a time looking for a 30% lift takes about two months. Looking for 20% takes four. Choose tests with big expected effects: offer, price presentation, checkout flow.
50,000	1,000	One properly sized A/B test every three to four weeks, or two in parallel on different steps with a longer window. Enough to run a real programme.
300,000	6,000	Three concurrent A/B tests or a 2 by 2 multivariate every two weeks, with corrections for multiple comparisons. Segment level tests (new versus returning, mobile versus desktop) become possible.
1,000,000 and above	20,000	Full experimentation 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 $n_{\text{per arm}}$ ; a 3 by 3 has nine. Fractional factorial designs let you test more factors with fewer cells by giving up on some interactions, and that is a trade I will often make at 50,000 visitors but almost never at 10,000. There are also four hygiene rules without which the numbers above mean nothing: run for whole weeks, never peek and stop early without a sequential plan, check the sample ratio (if 52% of visitors landed in variant B, your randomisation is broken and so is the test), and end within about six weeks before returning visitors' cookies expire and contaminate the arms.

One caveat on micro conversions, because they are the usual escape hatch for low traffic sites. Add to basket converts at 10% and needs a tenth 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. Micro conversions are a reasonable proxy for a test about the step they measure, and a poor proxy for anything downstream of it.

Layer seven: when you do not have the numbers yet, expertise is the data

Here is where I disagree with a lot of what gets written about "data driven marketing". The tables above make it look as though a business with 1,000 visitors a month has no way to improve except guessing. That is wrong, and it 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 you had the data, with the weights set by how much data you have. With 40,000 visitors in a test, the prior barely matters. With 400, the prior is most of the answer. The question for a small business is not "data or no data". It is "whose prior". A prior built from 15+ years of accounts, landing pages, checkouts and pricing pages across ecommerce, SaaS, hospitality and professional services is a very different starting point from a prior built from a blog post and a hunch, and you can put a number on it. If my experience says pages like yours convert between 1.5 and 3%, and I express that as $\text{Beta}(9, 391)$ , it carries the weight of about 400 visitors of evidence. Your first 400 visitors move me halfway from that belief towards whatever they actually do. Your first 4,000 all but replace it. That is how a business with no data gets decisions that are better than chance from day one and gets them replaced by its own evidence as soon as the evidence exists.

In practice the hierarchy of evidence I use for a small business looks like this, best at the top:

Your own controlled test, sized as above. Rare at small traffic.
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.
Category benchmarks adjusted for your price point and geography.
Best practice from accounts like yours: what has worked, repeatedly, on sites and campaigns with the same shape. This is what experience actually is.
Opinion, including mine and including yours.

Most agency work for small businesses runs on level five while describing itself as level one. Most of my work for small businesses runs on levels two to four, and says so. Concretely, that means the first months with a small shop or an early SaaS product are about: tracking that counts each order once and attributes it sensibly; the conversion fundamentals that are known to work before any test confirms them for you (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 from layer four so that it teaches you something without pretending to decide anything. The goal is to grow the business steadily into the traffic band where real testing becomes possible, and to have the data clean when you get there. Each step up the table in layer six unlocks a faster way to learn, which is why growth compounds once it starts.

The three lines assume a win in every third test, an average real lift of 6% per win after shrinking for the winner's curse, and one, three and six completed tests a month respectively. One test a month compounds to about 27% in a year. Six tests a month compound to almost four times the starting conversion rate. The lines are not promises; they are what the arithmetic of test velocity does, and test velocity is set by traffic and budget. The more we can test, the steeper the curve. That is the whole argument for growing traffic first, and for being honest about which line you are on.

Layer eight: is performance marketing even right for your stage and budget?

Before any of the above, there is a question that I think should be answered with 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

\text{CPA}_{\max} = \frac{\text{AOV} \times \text{contribution margin} \times \text{expected orders per customer}}{\text{target LTV to CAC ratio}}

For a hamper shop with a £68 median basket, a 45% contribution margin and 1.6 orders per customer over a year, aiming for the conventional 3 to 1, that is $68 \times 0.45 \times 1.6 / 3 \approx £16$ . Meta at £15 to £60 per conversion is marginal at best. Google Shopping at £15 to £50 is marginal. Which tells you the shop's first job is not more ads, it is a higher basket, a higher repeat rate, or a higher margin, and only then ads. For a SaaS product at £2,000 a year with 80% gross margin and a 30 month average life, the ceiling is $2000 \times 0.8 \times 2.5 / 3 \approx £1,333$ , and LinkedIn leads at £150 that close one in five are £750 a customer and comfortably inside it.

Then there is the learning budget, which is the thing nobody puts in the plan. Using the bracket table, a sensible minimum to learn whether a channel works at all is 30 to 50 conversions in it, and to compare two approaches within it is 100 each. On Meta or Google for a consumer product that means roughly £3,000 to £10,000 spent before you know anything reliable. On Google Search in a competitive B2B category or on LinkedIn, it means £10,000 to £30,000. If that learning budget is more than you can lose without regret, performance marketing is the wrong first channel, regardless of how good the creative is.

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

LinkedIn Sales Navigator and personal outreach. Around £80 to £100 a month per seat, plus your time. Right for anything with a contract value above roughly £5,000 and a definable buyer: SaaS, consultancy, agencies, manufacturing, recruitment, software for a specific trade. Wrong for consumer products and for anything below about £1,000 a year, where the time per conversation never pays back.
Email outreach. In the UK, business to business email to corporate addresses is allowed under PECR with an opt out, so it works for UK B2B lists. In Germany and Austria, unsolicited email needs prior consent even between businesses, so the same tactic there is LinkedIn, the phone with care, a letter, or an event. Do not let a UK playbook run in DACH.
Referral and partner programmes. 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 intent only. Google Search works when people already look for what you sell. It does not create demand. A new category with no search volume needs Meta, TikTok or LinkedIn to make people want it, and those channels need the budgets above.
Marketplaces and directories. Amazon, Etsy, booking platforms, comparison sites. You rent their traffic and pay their commission, which at small scale is usually a better deal than building your own paid acquisition.
Content and SEO. Six to twelve months before it pays, then low marginal cost. Right for anything searched for; wrong as the only channel for a business that needs revenue this quarter.

By industry, the channels that are usually relevant, and the ones that usually are not:

Business	Usually relevant	Usually not worth the learning budget first
Consumer ecommerce, basket under £100	Meta, TikTok, Google Shopping, marketplaces, email to existing customers	LinkedIn, search on generic terms
B2B SaaS	Google Search on problem keywords, LinkedIn, outbound via Sales Navigator, partnerships, content	TikTok, Meta beyond retargeting
Professional services and consultancies	Referrals, LinkedIn organic and outbound, SEO, speaking	Any paid channel until referrals are systematic
Local trades and services	Google Maps and Local Services, Google Search with location, Meta locally	LinkedIn, TikTok, national search
Hospitality and tourism	Meta, Google Hotel and Search, booking platforms, email to past guests	LinkedIn, generic TikTok
Manufacturing and industrial	Outbound, trade shows, LinkedIn, SEO on product terms, distributors	Meta, TikTok

The thing to take from the table is not the specific cells, which move every year, but the shape: 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 is not going to make it the right one.

Layer nine: forecasting, who can do it, and on how much data

Forecasting is where the gap between promise and data is widest, because a forecast is the one deliverable that looks the same whether it is built on three years of weekly history or on a spreadsheet of wishes. So let me be specific about what each amount of history buys you.

Minimum history by method. The rules of thumb I work to, in weekly data unless noted:

What you want to forecast	Method	Minimum history	Comfortable history
Next month's level, no seasonality	Moving average or simple exponential smoothing	8 weeks	13 weeks
Level and trend	Holt's method, regression on time	13 weeks	26 weeks
Weekly seasonality within the year	Seasonal decomposition, Holt Winters, seasonal regression	2 full years (104 weeks)	3 full years
Autocorrelated series with shocks	ARIMA, Box and Jenkins	about 50 observations	100 or more
The effect of ad spend on sales	Regression with adstocked spend, ARIMAX	20 to 30 weeks with real variation in spend	52 weeks or more with variation
Daily demand with weekday pattern	Any of the above on daily data	8 weeks (56 points)	26 weeks

The year of history for seasonality is non negotiable, and it is the one that catches most young businesses. Without two Decembers you cannot separate "December is big" from "we grew". With two you can estimate it, with three you can estimate it with a margin of error you would show a bank. Daily data gives you more points, and daily data is noisier, so it buys you less than the count suggests; a weekly series of 104 points is usually a better forecasting basis than a daily series of 180.

Accuracy you can honestly expect. I report mean absolute percentage error on a holdout, never a fit statistic on the training data, and these are the bands I see:

Situation	MAPE at a four week horizon	What that means for planning
Mature business, 3 or more years, stable channels	5 to 12% monthly, 10 to 20% weekly	Plan stock and cash against the band and it holds most months
1 to 2 years of history, some channel changes	12 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 quarter	25 to 40% or worse	A direction, not a plan. Re forecast monthly and say so.
New product, no history	No 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 standard error of a forecast $h$ steps ahead grows with $\sqrt{h}$ , so 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. Anyone showing you a single line twelve months out without a band is showing you a drawing. I walked through the full build, including the holdout and the band, in forecasting demand: autocorrelation, seasonality and honest error bands, and the broader map of methods is in the complete guide to forecasting.

Factoring in current campaigns. This is the request I get most and the one that most often cannot be met honestly, for a reason that has nothing to do with modelling skill. To estimate how much a pound of ad spend adds to sales, the model needs to have seen different levels of spend. If you have 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 I will put a campaign term into a forecast:

At least 20 to 30 weekly observations in which spend actually varied, with a coefficient of variation of roughly 0.3 or more. Periods where spend was zero are gold, because they show the baseline.
About 10 to 20 observations per parameter you ask the model to estimate. One channel with adstock and a saturation curve is three parameters, so 30 to 60 weeks. Three channels is nine parameters and you are talking about two years.
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, the honest alternatives are an incrementality test (switch a channel off in one region or for one audience for four weeks and compare, which needs regions comparable enough and at least a few hundred conversions in the period), or a prior ROAS from benchmark and experience, carried into the forecast as a scenario with a wide band rather than as a coefficient. Both are legitimate. Pretending the coefficient came out of your data when it came out of my head is not.

How much past campaign data for a reliable campaign forecast. For "what will next quarter's campaign do", the rule I give clients is three comparable campaigns or 12 weeks of continuous spend with variation, whichever is more. Below that, a forecast of a campaign is a benchmark range with your creative's early CTR used to narrow it, which is useful but is 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 you what pulling the budget forward into October does to December. That is the version worth paying for, and it is the version you can only buy with history.

What you have, and what I can honestly do 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 behind the interactive planner on my data science page, which takes your visitors, customers, history, campaign data and budget and places you on it.

Stage	What you have	What is honest	What I actually do with you
Foundations	Under about 3,000 visitors a month, fewer than 50 orders, little or no history	No statistical tests. Forecasts are scenarios. Paid media is for learning, with kill rules, not for deciding.	Tracking that counts once. Conversion fundamentals from 15+ years of accounts, 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.
Directional	3,000 to 20,000 visitors, 50 to 400 orders, 6 to 12 months of data	Micro conversion tests on large effects. Cohort and funnel analysis with margins attached. Level and trend forecasts, wide bands. Paid media 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. First seasonal forecast once two cycles exist.
Testing	20,000 to 100,000 visitors, 400 to 2,000 orders, over a year of data	One properly sized A/B test every three to four weeks. Seasonal forecasts at 12 to 25% error. Paid decisions at the 100 conversion bracket 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.
Programme	Over 100,000 visitors, thousands of orders, two or more years of varied spend	Concurrent 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 businesses I have grown started in the first row and the work was to get them to the third. What the table does is stop the project from 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.

If you want this for your own numbers, the free workshop is two hours 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. 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 decision you are recommending rest on, and what is the margin of error at that count?
What is the minimum detectable effect this test is sized for, and how many visitors per variant does that need at our baseline?
At our daily budget and our cost per conversion, how many days until that number of conversions exists?
What is the kill rule for a losing ad, and at what number of clicks does it trigger?
How many tests are running at the same time, 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 actually vary enough for the model to estimate its effect, or is the campaign coefficient a benchmark?
Which row of the table are we in, and what would move us to the next one?