You will be able to list the blind spots in attribution and the other evidence that fills them.
Priya's GA4 reports said most of her new enquiries came from organic search and direct visits. Her Google Ads and Meta ads together claimed a few more. Then, at a trial class, she asked a mother how she had heard about the centre. "Oh, my friend's son goes here. She sent your number to our class WhatsApp group." Priya asked the next five parents. Four gave a version of the same answer.
None of those recommendations appeared anywhere in her reports. They showed up as direct visits, or as searches for the centre's name, credited to Google. Every attribution model in lesson 5.1 was busy sharing out credit among the touches it could see, while the touch that mattered most was invisible to all of them.
Attribution only works on what can be tracked, and in practice that means clicks on links. A large part of how people choose a business never involves a trackable click.
A friend's recommendation over lunch, a message forwarded into a parents' WhatsApp group, a mention on a podcast, a flyer at the void deck, a conversation at a wedding, a shopfront someone walks past every day: all of these can start or finish a buying decision. When the person finally visits the website, they type the name or search for it, and GA4 records direct or organic search. The model then gives the credit to whichever trackable touch came last, which is usually the one that did the least persuading.
This is not a small gap for businesses that run on reputation, which in Singapore includes most tuition centres, clinics, renovation firms, home bakers and financial advisers. For them, the channels attribution cannot see are often the biggest ones.
Even the trackable touches go missing. Lesson 4.2 explained that visitors who decline analytics cookies may not be recorded, or only as modelled estimates. Their visits never join a path, so they cannot be credited.
Switching devices breaks the chain too. A parent who clicks Priya's Meta ad on her phone during lunch and books a trial on the family laptop that evening looks like two different users to GA4. The ad click sits on one path with no result. The booking sits on another path that began with a direct visit, and the ad gets nothing.
So every model, data-driven included, is working with a partial record. It can still tell you useful things about the touches it sees, as long as you remember it is not seeing everything.
The cheapest fix is to ask. A self-reported source question, such as "How did you hear about us?", on an enquiry form, a booking form or a checkout page, catches what tracking misses. People are surprisingly willing to answer, especially when it is one short question with a dropdown.
A few rules make the answers usable. Offer a fixed list rather than a free-text box, so answers can be counted, and include an "other" option with a short text field for the rest. Use the words customers use, such as "a friend or family member" or "WhatsApp group", rather than your channel names. Keep the list short enough to read at a glance. Place the question near the end, after the important fields, so it does not discourage anyone from finishing.
Self-reported answers have their own weaknesses. People remember the last or most memorable touch, not every touch, and someone who saw three of your ads may say "Google" because that is where they typed your name. So treat it as a second view that sits beside your tracking data. Where tracking and self-reporting agree, you can be fairly confident. Where they disagree, as with Priya's referrals showing up as organic search, the disagreement itself is the finding.
The other approach answers a different question. Attribution asks which touches were present before a sale. A holdout test asks whether a channel causes sales that would not have happened anyway.
The idea is simple. Pause one channel for a defined period, or in one area if you serve several, and keep everything else the same. Then compare results from your source of truth with what you would have expected. If enquiries hardly change while the channel is off, it was mostly taking credit for customers who would have come anyway. If enquiries drop clearly, it was adding something.
Priya could pause her Meta ads for four weeks and watch trial bookings in her spreadsheet. The weakness is that other things change too: school holidays, exam season, a competitor's promotion. So she would pick a period that is normally steady, compare it with the same weeks last year as well as the weeks just before, and decide in advance how big a drop would count as a real effect. Bigger businesses run this more formally with matched regions or groups of customers. For a small business, a careful pause with a written plan is a reasonable start.
Holdout tests cost something, since you may lose sales while the channel is off, so save them for channels where you spend real money and the attribution numbers leave you unsure.
For most small businesses, though, the self-reported question is the place to begin. It costs nothing, it starts collecting evidence the day you add it, and it is the second view of your channels that module 8 builds into your unit economics. The work now is deciding where your question goes and what options it offers.
Add a how did you hear about us question to one form or checkout and write down the answer options you will offer.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).