Week 5 · Web Metrics and Google Analytics I

Kopi Kembali ran the Iced Range Launch from 6 to 19 July 2026 across five channels. The director wants to know where the funnel leaks and what one action to take — and everything you hand her should be true.

At a glance

Time	~70–75 minutes, in pairs
Work with	A partner
You need	An agent (Claude or Manus) + a spreadsheet
You hand in	A one-page campaign brief — the Campaign Metrics Audit and Brief
Graded on	Whether every number in the brief is actually verified, not just trusted

The scenario

The export is kk-campaign-funnel.csv: 210 rows, one per day, channel, and device (mobile, desktop, tablet). Two different systems produced these columns — ad platforms count clicks, site analytics counts views and purchases — and they never agree perfectly.

Data dictionary

Column	Meaning
date, campaign, channel, device	The grain: one row per day, channel, device
impressions	Ad or post impressions served
viewable_impressions	Independently measured viewable impressions. Populated only where viewability measurement ran
clicks	Clicks recorded by the ad platform
landing_page_views	Landing page views recorded by site analytics
product_page_views, cart_adds, purchases	The rest of the funnel, site-side
spend_sgd	Spend for that row
purchases_last_click	The same purchases, credited by last-click attribution
purchases_data_driven	The same purchases, credited by the data-driven model
avg_engagement_time_sec	Average engaged time per session for that row’s traffic
new_user_pct	Share of that row’s sessions from first-time users

Get your data

• Download kk-campaign-funnel.csv

Step 1 — Profile with the agent · 10 min

Before you touch the agent, open the file in a spreadsheet and scan the first ten rows yourself. Are all the columns there? Any missing values? Do the numbers look plausible (clicks less than impressions)? Note it down: I noticed…

Then ask the agent:

Profile this campaign dataset. For each column, tell me: the column
name, its data type, the number of missing values, and anything that
looks inconsistent. For example, are there clicks greater than
impressions? Do any rows have zero spend but lots of conversions?
List everything you notice; do not analyse yet.

Sanity check

Does the agent’s row count, date coverage, and totals per funnel stage match what you saw yourself? If the agent says there are no missing values but you saw blanks, it may not have parsed the file correctly.

Step 2 — Spot-check one number by hand · 10 min

Pick one metric: total impressions, total conversions, conversion rate, or cost per acquisition. Build one formula in your spreadsheet (for example, =SUM(purchases)/SUM(impressions)*100 for conversion rate). Then ask the agent for the same metric and compare.

They match — good, the agent read the data correctly. They don’t match — stop, and find out why before going further. This is a supervision win either way.

Step 3 — Walk the funnel · 15 min

Show me the funnel for this campaign. Calculate the percentage at each
stage: clicks as a percentage of impressions, landing page views as a
percentage of clicks, product page views as a percentage of landing
page views, cart adds as a percentage of product page views, and
purchases as a percentage of cart adds. Show the raw counts and the
percentages, overall, by channel, and by device. Do not analyse; just
show me the numbers.

Require an absolute number next to every percentage. Identify the single biggest drop-off — for example, in a funnel of 100,000 impressions → 1,000 clicks (1%) → 800 landing page views (80% of clicks) → 400 product page views (50%) → 80 cart adds (20%) → 16 purchases (20%), the biggest drop is impression-to-click: 99% of people never clicked. That’s the leak.

Step 4 — Audit the numbers and find the leak · 20 min

Before any figure enters your brief:

• Compare clicks with landing_page_views per channel. Where they diverge badly, which number do you trust for reach, and why?
• Check viewable_impressions against impressions where both exist. What does that ratio do to any cost-per-impression claim?
• Plot or scan clicks by day per channel. Any day that looks nothing like its neighbours needs an explanation or an exclusion, stated in the brief. Read avg_engagement_time_sec and new_user_pct for that day before you decide — engaged time of a few seconds and near-100% new users is the signature of non-human traffic, not of a successful ad.
• Sum purchases, purchases_last_click, and purchases_data_driven. Reconcile the totals, then compare per-channel credit. Which channels does each model flatter, and which story would each tell the director?

Then ask the agent:

This campaign has a major drop-off at [the stage you identified]. What
could cause that? Generate three possible explanations for this
drop-off.

Pick the explanation that makes the most sense, then ask the agent what data would be needed to confirm it — it can’t answer that without further research or testing, and that’s the boundary of what an agent can do here.

Step 5 — Check against context, then write the brief · 20 min

Ask the agent for total spend and cost per acquisition (CPA), then how it compares to industry benchmarks (typical e-commerce CPA runs roughly SGD $15–50, depending on category). If the CPA looks too good to be true, verify it was calculated across the full dataset, not a subset — a CPA of SGD $2 against a $25 benchmark is either a lucky result or a calculation error.

Then write the one-page brief:

CAMPAIGN BRIEF: Iced Range Launch

SITUATION
Two-sentence summary of what you did.

FINDING
The biggest thing you learned. One sentence.

EVIDENCE
Three or four key figures that support the finding.

WHERE WE LOSE PEOPLE
Which stage of the funnel has the biggest drop-off.

RECOMMENDATION: ONE THING TO CHANGE
A specific action to improve — not "improve the landing page,"
but the exact change and what you'd measure.

WHY THIS FIRST
Why this action, not something else.

CAVEAT (optional)
Anything the data doesn't tell you.

Attach a supervision note: which agent platform you used, one number you verified and how, one thing the agent got wrong or that you corrected, and one question it couldn’t answer without further research.

Verify & submit

• Every percentage carries its absolute base
• The clicks-versus-views discrepancy is resolved or declared, not averaged away
• Any anomalous day is excluded with a stated reason or included with a caveat
• Attribution is presented as a chosen lens, with the choice justified — not one model’s credit presented as fact
• The single recommended action follows from the leak you identified
• I calculated one key number by hand and it matched (or I explained why it didn’t)
• I checked for bot or traffic-quality red flags (very short engagement time, ~100% new users)
• The brief is one page, and the supervision note is attached

This is the required Campaign Metrics Audit and Brief — evidence section 4 of your Individual Analytics Portfolio. Submit it through xSiTe with links to the exact Sheet tab or calculation.

Unit 10 app · lesson plan