Use AI to Decide Which Marketing Channel to Cut

To use AI to decide which marketing channel to cut, give it twelve months of spend and revenue by channel, then force it to rank channels twice: once on last-touch revenue and once on how often the channel appears anywhere in a converting customer's history. The gap between those two rankings is the entire analysis. Channels that look dead on last-touch and busy on assisted touch are the ones you must not cut, and they are exactly the ones a model will tell you to cut if you only hand it the first table.

Get the data into a shape worth analysing

You need less than people assume. Three files, twelve months, exported to CSV:

• Spend by channel by month, including the hours you spend on the unpaid ones, valued at some hourly rate. An unpaid channel is not free.

• New customers by month with the channel each is attributed to, however imperfectly your system does that.

• Revenue per customer over the twelve months after acquisition, not just the first order. Channels differ enormously on repeat behaviour.

That third file is the one most small businesses skip, and skipping it is how a channel that brings in cheap one-time buyers keeps beating a channel that brings in customers who stay for three years. If you genuinely cannot produce lifetime revenue, use six month revenue and say so in the prompt so the model qualifies its conclusion rather than overstating it.

If the customer list is messy, which it will be, fix that first rather than asking the model to reason over duplicates and inconsistent channel names. Cleaning up a customer list with AI is a 20 minute job that changes the answer you get here.

The prompt that avoids the attribution trap

The default failure is that you paste a spend and revenue table, the model computes cost per acquisition per channel, and it recommends cutting whichever channel has the worst number. That is arithmetic, not analysis, and it reliably kills the top of your funnel. Structure the request to make the model do the harder thing:

You are analysing marketing channel performance for a small business.

Data: three CSVs (spend by channel by month, customers by acquisition
channel, 12-month revenue per customer).

Produce, in this order:

1. A table of each channel: total spend, customers, cost per acquisition,
   12-month revenue, and return on spend. Use last-touch attribution.

2. A SECOND ranking that ignores last-touch. For each channel, state how
   many converting customers had ANY recorded contact with it, and what
   share of total converting customers that is.

3. List every channel where ranking 1 and ranking 2 disagree by more than
   two places. For each, explain which is more likely to be the truth
   and what evidence would settle it.

4. Only then recommend one channel to cut, and state explicitly what you
   expect to happen to the other channels' numbers if it is cut.

Rules: if the data does not support a conclusion, say so rather than
estimating. Do not invent industry benchmarks. Flag any channel where
fewer than 30 customers make the numbers unreliable.

Step 4 is the one that earns its place. A model asked to recommend a cut will always produce one. Asking it to predict the knock-on effect forces it to reason about whether the channel was feeding others, and it frequently reverses its own recommendation at that step. That is the same behaviour we rely on when prompting AI to argue against your own idea.

A worked example

A six person home services company runs five channels. The last-touch table looks damning for one of them:

On these numbers you cut community sponsorship, and possibly the van livery too. Now run the second ranking. Of 257 converting customers, 104 mentioned seeing the vans or knowing the company locally when asked at intake, and 61 of those also came through local search ads. Sponsorship appears in the history of 34 customers who converted through other channels.

Van livery goes from worst-but-one to a channel touching 40 percent of conversions, and it costs 1,800 a year. Cutting it to save 1,800 while it assists 104 conversions is not a saving. Community sponsorship still looks weak on both rankings, and that is now a defensible cut rather than a guess.

Note what made the difference: an intake question asking how the customer heard about the company. If you do not have that field, add it today. It costs one line in a booking form and it is the only cheap source of assisted-touch data a small business will ever have.

Where AI helps decide which marketing channel to cut

The last one in the bad column matters most. No model can tell you the counterfactual, and any confident number about what you would have lost is invented. The way to get it is to cut the channel for a defined period and watch, which is a decision about tolerance for risk rather than an analysis problem. Treat the output as a ranked hypothesis list, and see how to measure AI ROI for a small business for the discipline of actually checking afterwards.

Before you cut

1. Pause rather than cancel, for 60 to 90 days. Contracts and seasonal effects both punish permanent decisions made in a slow month.

2. Write down what you expect to happen to the other channels. If overall enquiries drop more than the cut channel's share, it was assisting.

3. Keep the intake question running through the pause. It is your only measurement instrument.

4. Redeploy the money into the channel the analysis ranked highest on assisted touch, not the one with the best cost per acquisition.

One thing worth stating plainly: this method tells you which channel is weakest relative to your others, not whether your marketing is working overall. A business with five poor channels gets a confident ranking of poor channels. The broader question of what to automate and what to spend on belongs with how much a small business should spend on AI tools and the wider guide to AI for small business.

Frequently asked questions

How much data do I need before AI can analyse my marketing channels?

Twelve months and at least 30 customers per channel. Below 30 the per-channel numbers swing so much on individual customers that any ranking is noise, and a good prompt will make the model flag that rather than producing a confident answer from a handful of conversions.

Can AI tell me which marketing channel is bringing the most revenue?

It can compute that from your data reliably. What it cannot do is see contacts your data never recorded, which is why a channel with no last-touch conversions can still be doing most of the work. Add a how-did-you-hear-about-us question at intake to close that gap.

Should I cut the channel with the highest cost per acquisition?

Not on that number alone. Cost per acquisition ignores repeat revenue and ignores assisted conversions entirely, so it systematically favours channels that catch people already ready to buy. Rank on twelve month revenue and assisted touch before deciding anything.

What should I ask AI about a channel that has no conversions at all?

Ask what evidence would distinguish between the channel being ineffective and being invisible to your tracking. For offline channels such as signage or sponsorship, the answer is almost always that you have no measurement instrument, not that nobody saw it.

How long should I wait after cutting a channel before judging the result?

At least 60 days, and longer if your sales cycle is more than a couple of weeks. Cutting an upper-funnel channel produces a delayed drop, so a fast read on the first fortnight will tell you the cut was free when it was not.