Doing the math, a test with eight headlines, six subheads, seven hero images, three body copy blocks, three background colors, and five offer buttons on one landing page comes out to 15,120 unique combinations. It's a number marketing technologist Scott Brinker worked out back in 2008, long before most teams had a testing platform fast enough to make this mistake at scale. I think most marketers build tests like this by accident, one well-intentioned variant at a time, and don't notice what they've done until the test has run for three months with no clear winner. There's a name for the mechanism that turns a handful of reasonable-sounding choices into three lost months, and understanding it on purpose beats discovering it by accident.

Combinatorics is counting without counting. It's the branch of math that tells you exactly how many ways you can select or arrange a set of things without ever making you list them out one by one, and that shortcut is the actual mechanism behind why funnels stall, why tests never reach significance, and why "let's just add one more audience segment" quietly multiplies your workload rather than simply padding it. This guide walks through the six steps I use to diagnose a broken funnel with that math before I touch a single creative asset, ending in a checklist you can run in the same meeting where someone proposes the next test.

Step 1: Is your marketing test a permutation problem or a combination problem?

Everything below rests on one distinction, so it comes first. A combination counts selections where order doesn't matter. A permutation counts arrangements where order does.

Here's the standard breakdown of those two formulas, applied to your own send calendar. Choosing which 2 of 5 subject lines to feature in this month's newsletter is a combination problem, since "line A and line C" is the same choice regardless of which one you decided on first: n=5, r=2, so nCr gives you 10 possible pairs. Deciding what order to send 3 nurture emails in is a permutation problem, since email A before email B tells a different story than B before A: n=3, r=3, so nPr gives you 6 possible sequences. Same three inputs, two different questions, two very different answers, and notice that you never had to write out the pairs or the sequences to get either number. That's counting without counting: the formula hands you the total directly. Most funnel math goes wrong because teams answer the combination question when they meant to ask the permutation question, or the reverse.

Step 2: How many combinations are you actually testing?

Multiply the number of variations for each element you're testing together and you get your true number of combinations without ever building the actual page variants, which is exactly how multivariate testing (MVT) software calculates what it's about to run: variations of element A × variations of element B × variations of element C, and so on. A landing page with 4 images, 2 CTA colors, and 3 headlines works out to 4 × 2 × 3 = 24 combinations competing for the same traffic: nine individual items, quietly running as two dozen simultaneous experiments. Scale that up the way real campaigns do, adding 20 ad variations and 4 audience segments on top of a 15,120-combination page, and you land north of 1.2 million total variants, which is the exact math Brinker's original example walked through. This is the same kind of arithmetic I walked through in the math behind virality: a handful of small, reasonable-looking multipliers compound into a number nobody would have signed off on if they'd seen it up front.

That table only has five rows and it's already past a thousand. This is why a test that felt reasonable in the planning meeting turns into a slog with no statistically valid winner: you didn't plan one experiment, you planned thousands, and you're splitting finite traffic across all of them at once.

Step 3: Why doesn't more traffic fix a bloated test?

Each additional combination divides your existing sample size and multiplies your workload at the same time. With 24 variants, each one gets roughly 1/24th of your traffic, and reaching statistical significance requires every one of those slices to individually hit the sample size your confidence level and minimum detectable effect demand, the same inputs any A/B test sample size calculator asks for before it'll call a winner. Teams under pressure to show a result respond by checking the dashboard early and stopping as soon as something looks significant, which is its own trap. Statistician Evan Miller showed that checking results after every new visitor and stopping the moment you see significance can push your real false positive rate to 26.1%, even when your dashboard is telling you it's 5%. The fix is deciding your combination count and sample size before the test launches, then leaving the dashboard alone until you hit it.

Step 4: Does the order of your touchpoints change your attribution math?

It depends entirely on which attribution question you're asking, and that's the same permutation-versus-combination distinction from Step 1, applied to your CRM instead of your test plan. "Which channels touched this deal" is a combination question: paid search, a webinar, and a case study download form one answer regardless of sequence. "What sequence of channels moved this deal forward" is a permutation question, and it's the one that actually explains why a first-touch model and a time-decay model can disagree about which channel deserves credit for the same customer, a gap most multi-touch attribution guides gloss over. Four distinct touchpoints can occur in nPr = 4P4 = 24 different orders. If your attribution model only counts which channels were present and ignores the order they happened in, you're answering a combination question while your CFO is asking you a permutation one. That mismatch is a common reason multi-touch attribution reports don't match what sales says actually closed the deal.

Step 5: How many audience segments can your funnel actually support?

Every attribute you layer onto a segment (industry, company size, role, lifecycle stage, region) multiplies your total segment count the same way test variables multiply combinations, and most teams hit the wall well before they notice it. Four industries times three company sizes times three roles times three lifecycle stages is already 108 segments, and few funnels generate enough monthly volume to say anything statistically meaningful about all 108 at once. Marketing scientists describe this as the same curse of dimensionality that shows up in statistics generally: as you add dimensions, your data gets spread thinner across an exploding number of cells until most of them are functionally empty. Over-personalizing without the volume to back it up actively degrades the experience, since a "personalized" message built on a sample size of six barely qualifies as personalization at all. It's the same tradeoff I laid out in diminishing returns in marketing: a bigger audience isn't free, and here the cost shows up as segments too thin to trust.

Step 6: What's the framework for fixing a funnel with combinatorics?

Run these three moves in order, every time, and you'll catch a bloated test before it ever launches.

Count. Classify the question as a permutation or a combination (Step 1), then multiply out every variable you're planning to test simultaneously (Step 2). If the number is in the thousands, you're running a math problem disguised as a marketing plan.

Cut. Bring the combination count down until it matches your actual traffic. Test main effects (does this headline win) before interaction effects (does this headline win specifically paired with this image), pick an attribution model that matches the question you're actually asking (Step 4), and build only the audience cells you have the volume to support (Step 5).

Confirm. Decide your sample size and test duration in advance (Step 3), and don't touch the dashboard's verdict until you hit it.

A funnel with enough variables in play stops being a marketing problem and starts being a counting problem, and counting problems have exact answers whether or not you bother to do the math first. That's the whole discipline here: counting without counting, done on a whiteboard in the same meeting where someone proposes adding one more test variant, long before a test has run for three months with no winner to show for it.

If this connects to something you're trying to solve, book a complimentary consulting session. No pitch, just perspective.

Frequently Asked Questions

What is combinatorics, and why does it matter for marketing?

Combinatorics is the branch of math that counts how many ways you can select or arrange a set of things without listing them out one by one. It matters for marketing because it's the actual mechanism behind bloated A/B tests, statistical significance you'll never reach, and audience segments too small to trust.

What's the difference between a permutation and a combination in marketing?

A combination counts selections where order doesn't matter, like which channels touched a deal. A permutation counts arrangements where order does matter, like the sequence those channels happened in. Attribution models that blur this distinction can disagree about which channel deserves credit for the same customer.

How do you calculate the number of combinations in an A/B or multivariate test?

Multiply the number of variants for each element you're testing together. Four images times two CTA colors times three headlines works out to 4 × 2 × 3 = 24 total combinations, all competing for the same traffic at once.

Why does adding more audience segments hurt a marketing test?

Every attribute layered onto a segment (industry, company size, role, lifecycle stage) multiplies the total segment count the same way test variables multiply combinations. Most funnels don't generate enough monthly volume to say anything statistically meaningful once that multiplication passes a few hundred segments.

Does the order of marketing touchpoints affect attribution reporting?

Yes. "Which channels touched this deal" is a combination question, but "what sequence moved the deal forward" is a permutation question. Four distinct touchpoints can occur in 24 different orders, and a first-touch model and a time-decay model can disagree about credit because they're answering two different versions of that question.