Optimising for the Wrong Event: Why Cheap Conversions Attract Buyers Who Never Buy

A pattern that shows up constantly and is almost never diagnosed correctly. The free sample offer converts brilliantly on cold traffic, cost per lead is excellent, the funnel numbers look healthy, and practically none of those people ever buy the actual product. Organic visitors who request the same sample convert at a normal rate. Same page, same offer, same follow-up emails. Only the entrance differs.

The algorithm did exactly what you told it to

Meta's delivery system is an extremely effective pattern matcher pointed at one target: the event you nominated. It finds people who look like the people who fired that event, and it is very good at it.

If the event is "requested a free thing", it learns the characteristics of people who request free things, and goes and finds more of them. People who enjoy claiming free samples are an enormous, cheap and very easy to reach audience. The cost per event will be excellent because the algorithm is succeeding at the task it was given.

The task was just the wrong one. You asked for sample requests and got sample requests. Nobody asked for buyers.

This is why the same offer converts differently depending on the entrance. Your organic visitor requested a sample because they were already considering a purchase; the sample is a step in their buying process. Your paid visitor requested a sample because a free thing appeared in their feed. Identical action, completely different intent, and the event you are optimising toward cannot tell them apart.

The trade-off nobody escapes

The obvious fix is to optimise for purchase instead, and often that is the right answer. But it is a genuine trade-off, not a free upgrade, because of volume.

Optimisation needs examples to learn from. The deeper down the funnel you point it, the fewer examples exist. A store getting 400 sample requests a week and 20 purchases has plenty of signal on the first and very little on the second. Switching to purchase optimisation on 20 weekly events puts you well under the threshold where delivery stabilises, and you get erratic spend and poor results for a different reason. The mechanics of that floor are in the 50 conversions per week rule.

So the real question is not "which event is most valuable". It is the deepest event in your funnel that still produces enough weekly volume to learn from. That is the sweet spot, and it moves as you scale.

Four ways out, roughly in order of effort

1. Optimise for purchase and accept slower learning. If you are anywhere near the volume floor, this is the cleanest answer. Budget for a genuinely slow start, do not panic and restructure during the learning phase, and give it a fair run. Most accounts that tried this and concluded it does not work actually changed something on day four.

2. Add a value rule so not all conversions count equally. Rather than treating every sample request as equivalent, attach a value and let the system prefer higher-value ones. This only works if you are genuinely sending differentiated values, and it needs enough purchase history behind it to mean anything. On thin volume it does very little, which is the honest answer to a question people ask often.

3. Define a stricter custom conversion.Instead of optimising toward every sample request, optimise toward the subset that correlates with buying. If people who configure a product before requesting a sample buy at five times the rate, that configuration step is a far better optimisation event than the request itself. You are moving the target from "took a free thing" to "showed purchase intent", which is the actual distinction you care about. This is usually the highest-leverage option and it is underused.

4. Keep the free offer out of cold traffic entirely. Send cold traffic straight to the product and optimise for purchase or add-to-cart. Reserve the sample offer for retargeting and email, where it works on people who have already shown interest. This is the structural fix, and for high-consideration products it is often simply correct.

Send the real outcome back, not just the front-end event

There is a step underneath all four options that is worth more than any of them, and most accounts skip it.

If the outcome you care about happens somewhere the pixel cannot see it, the algorithm will never learn from it. A sample request that becomes a £900 sofa order three weeks later, a quote request that becomes a signed contract, a lead that becomes a booked job: if none of that flows back to the platform, you are permanently optimising toward the front of your funnel.

Sending those downstream outcomes back, with their real values, changes what the system is able to learn. It is the difference between teaching it "find people who fill in forms" and "find people who eventually pay us". The mechanics are in the offline conversion tracking guide and, for lead businesses specifically, conversion leads and your CRM.

This is also where the ordinary tracking problems bite hardest. If your server-side events are dropping, or your match quality is poor enough that the platform cannot connect a purchase back to the person who clicked, you can implement all of the above correctly and still see no effect. Worth confirming the plumbing works first with improving event match quality.

A worked example, because the arithmetic is the argument

A made-to-order furniture brand sells sofas at around £900. The entry point is a free fabric sample pack. Two channels, same page, same pack.

• Organic and direct. 200 sample requests a month, of which 24 become sofa orders. A 12% sample-to-sale rate.
• Meta cold traffic optimised for the sample request. 600 sample requests a month at £4 each, of which 6 become orders. A 1% sample-to-sale rate.

The paid cost per sample looks excellent and the cost per actual customer is £2,400 against a £900 product. Meanwhile the numbers on the dashboard are the best in the account, because the dashboard is counting samples.

The twelve-fold difference in downstream conversion rate is the finding. It is not a creative problem, a landing page problem or an audience problem, and no amount of testing hooks will close a gap that size. Two populations were recruited using the same bait for completely different reasons, and only one of them was ever going to buy furniture.

Note also that the paid channel produced three times the sample volume of organic. Success at the nominated metric was total. That is what makes this failure mode so durable: every intermediate number says it is working.

Lead generation has a harder version of this

Ecommerce at least ends in a purchase event the platform can eventually see. Lead generation often does not, and the same dynamic gets worse.

Optimise for form fills and you will get form fills, including from people who fill in every form they encounter. The sales team quietly learns that paid leads are not worth calling, stops working them properly, and the resulting lack of closed deals confirms the belief. By the time anyone examines it, the channel has been written off on evidence it partly manufactured.

The fix is the same in shape and harder in practice: get the qualifying outcome back to the platform. Not "submitted a form" but "was qualified by sales", or "booked a meeting that happened", or "became a customer", with its value attached. That usually means wiring your CRM up rather than tweaking a campaign setting, which is why it gets deferred, and why the accounts that do it tend to pull ahead of the ones that do not. The practical route is in conversion leads and your CRM.

How to tell which problem you actually have

Before changing anything, establish whether this is an optimisation problem or something else wearing its clothes. Three checks.

1. Compare downstream conversion rate by source. Of people who took the front-end action, what percentage eventually bought, split by where they came from? If paid converts far below organic on the same action, you have the problem this article is about. If they convert similarly, your issue is volume or offer, not targeting.
2. Check whether the purchase event is even arriving.A suspiciously clean "lots of leads, zero sales" picture is sometimes not a targeting failure at all. It is a purchase event that stopped firing, which produces exactly the same shaped report. Rule it out, using clicks but no sales.
3. Look at the time lag before concluding. If your product genuinely takes six weeks from sample to purchase, a cohort three weeks old has not had time to buy. Plenty of campaigns get killed for failing to produce revenue that was still on its way.

The principle worth carrying into every account

Optimisation systems are literal. They will pursue precisely the thing you nominated, with more patience and at greater scale than you could manage yourself, and they will not notice or care that the thing you nominated is a poor proxy for what you actually want.

So the question to ask before launching anything is not "what can I easily measure". It is "if the algorithm becomes perfect at maximising this event, do I get rich or do I go broke?"

For a free sample event, a perfect algorithm finds every person alive who likes free things and hands them your samples. That is the outcome the setup was asking for. The fix is not better creative or a bigger budget. It is nominating a target whose perfect version is a business you would want to own.