Why none of the numbers agree

Your analytics, your advertising platform, your email tool and your own accounts will all report different numbers for the same month. Weeks get lost trying to reconcile them. They cannot be reconciled, and knowing why is more useful than trying.

The reasons, which are structural

• They count different things. An advertising platform counts a conversion if somebody saw or clicked an advert and later converted, within a window it chose. Your analytics counts the last source before the visit. Neither is wrong; they are answering different questions.
• They use different windows. One attributes a sale to a click from three weeks ago, another to the visit yesterday.
• They see different people. Anything depending on device storage misses whoever declined, used a blocker, or switched device.
• They de-duplicate differently. One person on a phone and a laptop is one customer, two visitors, and possibly two conversions.
• They are measured at different points. A platform counts a click, your server counts a request, your shop counts a paid order, and people drop out between each.

None of these is a bug and none will be fixed. They are consequences of measuring from different places.

The offset, measured once a month

1. 01Count what actually arrived: orders, enquiries, bookings, revenue, from your own records.
2. 02Note what each tool claimed for the same period.
3. 03Work out the ratio for each and write it down with the date.
4. 04Re-check when anything changes: the site, the consent banner, a platform update, a new payment flow. The ratio moves and a stale one is worse than none.

Illustratively, if the advertising platform consistently claims about a third more conversions than you can find in your own records, that is useful information: you can keep using its figures to compare one advert against another, while using your own for deciding whether the channel pays at all.

Attribution, honestly

Somebody reads an article, forgets about it, sees an advert, searches your name a fortnight later and buys. Every model assigns that sale to one or two of those touches and every assignment is a convention rather than a fact. The attribution question for a small business is not which model is correct; it is whether the channel is worth continuing, and that is answered by what happens to total sales when you change the spending.

Which leads to the one method that actually works at this size: turn something off for a few weeks and see what happens to the total. Crude, uncomfortable, and more reliable than any dashboard.

The one question customers can answer

Since no tool can tell you what caused a sale, it is worth asking the only party who was present. One optional field at the point of order, asking how they heard of you.

The answers are unreliable individually. People forget, and they credit whatever they saw most recently rather than whatever started it. In aggregate over a few months they are informative in a way no dashboard is, because they are the only data that sees across every channel at once.

1. 01Keep it optional and keep it one field. A required question costs you orders, and the cost is larger than the value of the answer.
2. 02Use a free text box rather than a list, at least at first. A list only returns the options you already thought of, which is the thing you are trying to learn.
3. 03Read them monthly and look for anything that is not a channel you are measuring: a person, a community, a podcast, a friend.
4. 04Compare the distribution with what your tools claim. Where they disagree badly, the tools are usually missing something structural rather than the customers being wrong.

This is the cheapest measurement available to a small business, works regardless of consent, and almost nobody does it because it does not produce a chart.