M10.L3 · Analytics & Experiments

UTM discipline & the honest limits of attribution

14 min

What you'll be able to doWrite a UTM naming convention, apply it to five real links, and state an attribution stance in plain language — including what it systematically over- and under-credits.

Learn

The most-asked, worst-answered question

"Where do our customers come from?" Half the answer is hygiene: links labelled so consistently that you can still read the report in six months. The other half is humility: accepting that every attribution model is a simplification that flatters some channel, and that a large part of every journey is invisible to all of them — the dark funnel you met in M1.L2.

Your M1.L6 baseline let you write "word of mouth, no idea how much" in the top-sources box, and called it a valid answer. It was — then. This is the lesson where "no idea how much" starts becoming a number, without ever pretending to become a precise one.

UTMs: analytics only knows what you tell it

A UTM is a label you append to a link. Five parameters exist; three do most of the work:

?utm_source=newsletter&utm_medium=email&utm_campaign=2026-05_intake-study

utm_source says who sent the click (newsletter, linkedin, google), utm_medium says the kind of channel (email, organic-social, cpc), utm_campaign says which effort. utm_content distinguishes variants; utm_term is a legacy field for paid keywords. Untagged links get guessed at — and the guess is usually "direct", the analytics word for "no idea".

Here is the part that fails in the real world: to a computer, "Newsletter", "newsletter" and "email" are three different channels. Six months of casually tagged links produces a report with eleven channels, four of which are the same newsletter. The cure is not effort but a convention — written down once, then obeyed:

Lowercase, always. No spaces, ever. Hyphens between words.
A fixed vocabulary for source and medium — a short list you choose from, never improvise into.
Dated campaign names: 2026-05_intake-study sorts, groups and still makes sense in a year. sale does none of that.
One page, where you actually build links. Lena's whole convention fits on a page: sources {linkedin, newsletter, forum, google, partner}, mediums {organic-social, email, referral, cpc}, pattern YYYY-MM_slug.

Two cautions: never UTM-tag internal links (clicking from your own homepage to your own product page would restart the session as a fake new arrival), and remember UTMs are visible in the address bar — they are labels, not secrets.

Attribution: every model is an opinion about credit

Tagging tells you the last link a customer clicked. Attribution is the decision about who gets credit — and every available rule is an opinion:

Last-click gives everything to the final touch. Simple, standard — and structurally biased towards the channels that stand at the end of journeys: brand search, email, retargeting. It systematically starves whatever created the demand.
First-click, linear, position-based redistribute the same clicks by a different rule. None of them observes causation; they re-slice the same partial evidence.
Data-driven (GA4's default) lets a model do the slicing. Still an estimate, now with the added property that you can't check its working.

And all of them share the same blind spots: consent refusers, cross-device journeys, links shared in WhatsApp and group chats, conversations at trade shows and kitchen tables. The dark funnel doesn't appear in any model, because no click ever happened.

Storkflow finally measures the story

The pack's oldest complaint is Lena's: "The trade show works and nothing else clearly does, but I can't prove any of it." Her click data agrees with the first half and can't see the rest — last-click files most trial starts under direct and google/organic, and the trade show appears nowhere, because nobody clicks a trade show.

So she added the crudest instrument in measurement, and the most underrated: a required "How did you hear about us?" dropdown on trial signup. First full month, all 41 trials answered:

a colleague or word of mouth 13 · the trade show 9 · search 8 · LinkedIn 6 · the intake study 3 · can't remember 2

Twenty-two of forty-one — more than half — named sources no click model can see. The trade show went from a story to a number: 9 of 41. Note what actually happened in the journeys: someone hears of Storkflow at a stand or from a colleague, later types "storkflow" into Google, and last-click hands the credit to search. The click report wasn't lying; it was answering a narrower question than the one Lena was asking.

Self-reported attribution has its own biases — it over-remembers memorable touches and under-reports ads (Ostara's checkout question found 5 self-reported ad orders against the dashboard's ~12 claimed; Module 9 left that gap deliberately open). Which is exactly why you run both, side by side, and never average them.

The only causal instrument: turning it off

Every model infers. The one honest causal question is: what happens if we stop? This is incrementality — holdout and geo tests at big companies, and at SME scale, the humble pause test.

Mateus ran the one Module 9 proposed: retargeting (the ≈€250/month layer) switched off for two weeks in quiet May. The dashboard's last-click logic implied about 6 orders would vanish. The pause fortnight delivered 43 orders against a typical ~45. The honest reading, now canon: retargeting's true incremental effect is small — somewhere between nothing and about three orders a fortnight — and a single fortnight can't be more precise. Consequence: retargeting stays as a small always-on layer, and its 3.1 ROAS is never quoted in a report again. The number the whole Program has been side-eyeing since Module 1 is finally, quietly, retired.

Your stance, in one paragraph

The deliverable of this lesson is not a tool setting. It is a paragraph you could say to your boss without flinching. Lena's:

"We use last-click reports for week-to-week operations — they're consistent and cheap. We know they over-credit search and email and cannot see the trade show or word of mouth, so budget decisions also use the signup question, which says over half our trials start in places clicks can't see. Where the two disagree, we say so, and for anything expensive we prefer a pause test to an argument."

That is what attribution honesty looks like: models named, biases admitted, dark funnel counted crudely rather than ignored. Now write yours.

Do

Exercise 10.3.1 — Your UTM convention and attribution stance

Write your UTM convention, tag five real links you will actually use this month, and state your attribution stance in words your boss would understand.

Write these down — in your plan document, or on the worksheet at the end of this lesson.

What to write	Guidance
Your UTM convention	Allowed `source` values (a list of at least 3) · allowed `medium` values (at least 2) · a `campaign` naming pattern that includes a date element, e.g. `2026-09_autumn-sale`. Lowercase throughout
Five tagged links	Real links you will actually use this month: destination URL plus the built link. Check each against your convention — no value outside your lists, no uppercase, no spaces
Your attribution stance	60–150 words. Name the model(s) you'll use and at least one bias of each — "last-click over-credits X" is the minimum bar
The self-report question	Will you add "How did you hear about us?" — yes (say exactly where: checkout, signup, booking call) or no (say why). 10–60 words

To close the exercise, reopen the top_sources answer from your M1.L6 baseline: which of those sources will your five links finally make countable, and which will stay dark?

Where this goes: Measurement §3 — Attribution — of your Marketing Plan. Use your convention for every link you build from now on.

Check

Quiz — 4 questions

1. Storkflow's last-click reports file most trial starts under "google / organic" and "direct", yet 9 of 41 new trials say they heard of Storkflow at the trade show. The best interpretation is:

a) The customers are misremembering
b) The trade show creates demand that later converts via brand search, which last-click credits to Google ✔
c) The trade show has no effect
d) The analytics is broken

Why: nobody clicks a trade show. The journey was stand → remembered name → brand search → trial; last-click paid Google for the trade show's work. Both numbers are right — they answer different questions.

2. Which UTM set follows good convention?

a) utm_source=Facebook&utm_medium=Social Media
b) utm_source=facebook&utm_medium=paid-social&utm_campaign=2026-09_autumn-sale ✔
c) utm_source=fb on some links and utm_source=meta on others
d) utm_campaign=sale alone

Why: lowercase, no spaces, a dated campaign name, one fixed vocabulary. (a) creates duplicate channels through capitalisation, (c) splits one source into two, (d) will be meaningless within a month.

3. In Lena's first month of self-reported data, 22 of 41 trials named a colleague, word of mouth or the trade show. What does this say about her click-based reports?

a) They should be switched off — they're wrong
b) They're correct and the survey is wrong
c) They're accurate about clicks but describe less than half the journey — both instruments are needed, read side by side ✔
d) The 22 respondents saw ads they forgot about

Why: click attribution answers "which link came last?", not "where do customers come from?". Self-report is crude and biased too — towards memorable touches, against ads. Two flawed instruments, disagreeing honestly, beat one flawed instrument believed completely.

4. Ostara paused its ≈€250/month retargeting for a fortnight: 43 orders instead of a typical ~45, where last-click implied ~6 would vanish. The soundest conclusion is:

a) Retargeting drives exactly 2 orders a fortnight
b) The pause fortnight proves retargeting is worthless
c) Retargeting's incremental effect is far smaller than its dashboard claimed, though one fortnight can't give a precise figure ✔
d) The test should have run during December

Why: turning a channel off is the only causal instrument, but two weeks of noisy order counts settles direction, not decimals. And never run a pause test into your peak season — that's why May.

Advance

Three of five. You have labelled links, a written convention, and an attribution paragraph that admits what it cannot know — which puts your reporting ahead of most agencies'.

Next: M10.L4 — A/B tests and statistical significance. You can now mostly trust where visitors come from. Next: whether the changes you make for them actually work — and why the answer usually arrives slower than anyone wants.

Mark your own work

	Good	Not yet
Convention is closed	Fixed vocabulary; nothing invented per-link	"I'll tag sensibly as I go"
Campaigns are dated	`YYYY-MM_slug` or equivalent	`sale`, `test`, `new`
Links validate	All five parse against the convention	A capital letter or space survives
Stance names biases	Each model's flattery named	"We use last-click because it's accurate"
Dark funnel counted	A self-report question, placed somewhere real	Word of mouth left permanently invisible

Worksheet

THE SCHOOL OF NET MARKETING
Lesson 10.3 — UTM convention & attribution stance

MY CONVENTION (lowercase, no spaces, hyphens)
  allowed sources:  ____________________________________
  allowed mediums:  ____________________________________
  campaign pattern: YYYY-MM_______________  (date required)

FIVE LINKS I WILL USE THIS MONTH
  1. _________________________________________________
  2. _________________________________________________
  3. _________________________________________________
  4. _________________________________________________
  5. _________________________________________________
  check each: lowercase? no spaces? every value from
  the lists above? campaign dated?

MY ATTRIBUTION STANCE (60–150 words — say it to your boss)
  Models I use: _______________________________________
  What each over-credits: _____________________________
  What none of them see: ______________________________
  _____________________________________________________

SELF-REPORT QUESTION
  "How did you hear about us?" goes:  ☐ checkout
  ☐ signup form  ☐ booking call  ☐ nowhere, because
  _____________________________________________________

REMEMBER
  Last-click pays the closer, not the opener.
  The dark funnel is real: count it crudely or not at all.
  The only causal question is "what if we turn it off?"

Next: Lesson 10.4 — A/B tests & significance.
theschoolofnetmarketing.com/learn/ab-tests-and-significance