Whether, chapter thirteen
Measurement you can defend
The link
| Source / medium | Campaign | Links |
|---|
If the redesign doesn't measurably move all three, it has failed regardless of how good it looks.
From my brief for the UVA Food Assist web redesign, 22 May 2026.
The idea
I think measurement starts before launch. Decide what counts as success while you can still be honest about it, tag every link the same way, and know which question each report answers. If you decide afterwards, there is always some number that makes the work look good.
Attribution models don't tell the truth. They tell different truths, because each one is a rule about which touch deserves the credit. So I choose one on purpose and say which. When data is missing, I say it's missing instead of filling the gap with a guess.
Tag it
A UTM tag is a label on the end of a link that analytics reads when someone clicks. The tools group by the exact text, so a capital letter or an underscore where a hyphen should be starts a new row and splits the numbers. The builder below checks a link against a short house list of mediums and against the Kiln links already in the log. It starts with two mistakes in it.
Credit it
This is one Ledgerline buyer's path to a demo request over thirty days, five touches long. The buyer is invented. Switch the model and watch the credit move. Nothing about the buyer changes; only the rule for sharing out the credit does.
- Day 1 Paid social: an ad in their feed
- Day 9 Organic search: a search for job costing software
- Day 15 Email: a webinar invitation
- Day 26 Referral: a review site
- Day 30 Paid search: a search for Ledgerline by name, then the demo request
| Channel | First touch | Last touch | Linear | Time decay | Position based |
|---|---|---|---|---|---|
| Paid social | 100% | 0% | 20% | 2.7% | 40% |
| Organic search | 0% | 0% | 20% | 6.0% | 6.7% |
| 0% | 0% | 20% | 10.9% | 6.7% | |
| Referral | 0% | 0% | 20% | 32.3% | 6.7% |
| Paid search | 0% | 100% | 20% | 48.1% | 40% |
Letting the paths decide
The five models above are rules a person picked. A data-driven model reads many paths instead, the ones that ended in a demo and the ones that didn't. The one I find easiest to explain is a Markov chain with a removal effect: for each channel, how much conversion would be lost if that channel disappeared from every path.
Here are twelve invented Ledgerline paths. Six end in a demo request.
| Touches, in order | Ended in |
|---|---|
| Paid social, organic search, email, referral, paid search | Demo request |
| Organic search, email, paid search | Demo request |
| Referral, email | Demo request |
| Paid social, email, email | Demo request |
| Organic search, referral, paid search | Demo request |
| Email, referral | Demo request |
| Paid social | Nothing |
| Paid social, organic search | Nothing |
| Organic search | Nothing |
| Paid social, paid social | Nothing |
| Paid search | Nothing |
| Organic search, email | Nothing |
| Channel | Removal effect | Last touch |
|---|---|---|
| 27.3% | 33.3% | |
| Organic search | 20.4% | 0% |
| Paid search | 20.4% | 50% |
| Referral | 20.2% | 16.7% |
| Paid social | 11.7% | 0% |
Which one to use
None of them is right. Last touch is fine for a quick shop where people buy on the first visit. For a long B2B sale like Ledgerline's I'd report position based for the monthly story, keep the removal effect as a check once there are enough paths to trust it, and write the model's name on every chart. Google Analytics 4 dropped first touch, linear, time decay and position based in 2023 and kept last click and its own data-driven model, so most teams now see one model by default, often without anyone saying a choice was made.
Where I used it
- The Giving Tuesday 2025 plan for Resilience Education set its revenue target before the campaign: $75K, modelled in Excel against several years of giving history, with base, target and stretch scenarios. Work for a nonprofit; described in my own words.
What I'd do now
In a group project, the Google Analytics prediction section of our report went out as placeholder slides. It wasn't my section, but it's the gap I'd close first. Here it is for Kiln, written as a prediction to check against the first month of real data, not as a result. Kiln is invented, and so is everything below. I haven't taken the Google Analytics certification yet.
- Audience
- People who already buy good coffee online, plus café regulars the shop sends to the site. I'd expect returning visitors to be a bigger share than a new shop usually sees, because the café does the first introduction.
- Acquisition
- Instagram and the email list bring most first visits in launch week. Search for single-origin and roaster terms grows slowly behind them, and every link from the first two is tagged.
- Behaviour
- Product pages get the time. The brew guides get read and left, so I'd watch whether anyone who reads a guide goes on to a product page.
- Conversions
- Two goals: a first order, and a subscription started. A newsletter sign-up is tracked but doesn't count as success.
- First-order conversion rate, by source / medium. It says which tagged channel brings buyers rather than visits.
- Subscriptions started per 100 first orders. A roaster makes its money on repeat bags.
- Customers who order again within 60 days. The second bag is the real test of the coffee.
- Revenue per session, by landing page. It shows which page earns the traffic it gets.
- Share of sessions with no channel assigned. My check on the tagging itself. If it climbs, the four numbers above are wrong and I'd say so.