Web analytics: responsibilities and practical exercise
Data analysis turns a question into defined measures, checks collection and explains results within their scope. It distinguishes observed events from the business action being investigated. A report is useful when units, exclusions and limitations support an explicit decision.
These sheets describe activities and teaching exercises. They do not constitute job offers, certifying courses or recruitment promises.
Turn website measurements into decision-supporting observations, starting with the question, metric definitions and collection limits.
Typical activities
Define events and verify their journey triggers.
Distinguish sessions, enquiries, conversions and duplicates.
Compare periods with campaign, consent and content changes in context.
Responsibilities to agree with the team
A metric does not automatically establish causation. Explain collection losses, scope changes and insufficient data. Tracking implementation requires appropriate review.
A first piece of work to demonstrate
Present a question, tested measurement, proportionate finding and an action whose effect can be observed.
This is an editorial description of common activities. Job titles and scope vary across teams; it is not a certification framework.
Distinguish consultation of form, attempt to send and request received.
Work on a copy or demo environment. Define the data used and the expected result before you begin. Keep your decisions and the difficulties encountered in order to be able to explain the process.
Define events and what should be excluded from the data.
Check triggers with test cases.
Document periods, populations and collection breaks.
Link a reading of the results to a limited and verifiable decision.
Deliverable and checks
Provide the measurement plan for viewing, attempting submission and receiving an enquiry. Specify trigger, unit, exclusions and validation source for each event. Present demonstration data with period, filters, duplicates and collection breaks identified.
Repeat a successful submission, an error and a retry, comparing events with the expected business result. Explain discrepancies before calculating a rate. Check that readers can distinguish missing data, zero and an event that does not apply.
An increase may be due to a change in collection or public. The measure must specify what it does not allow to conclude.
Collaboration to explore: SEO and visibility.
Fictional deliverable excerpt : A measurement dictionary distinguishes a visit, contact click and enquiry actually received. It specifies unit, period, filters and collection limits before interpreting a change.
What passes between the roles : Share changed pages and their dates alongside measurement definitions. An observed increase alone does not establish the effect of an SEO action; other changes and data limitations need examination.
Fictional teaching material: this excerpt illustrates part of a deliverable. It is not a client reference, a complete specification or evidence of professional qualification.
Sample excerpt
Fictional dataset: 100 distinct enquiries reviewed.
Qualified: 45.
Outside scope: 35.
Status still unknown: 20.
Finding: 45% qualified among the 100 reviewed enquiries.
Fictional follow-up: the 20 unknown enquiries are reviewed; 10 become qualified and 10 outside scope.
New record: 55 qualified, 45 outside scope, none unknown; 55/100 = 55%.
Three distinct statuses in the fictional dataset
Qualified
45
Outside scope
35
Unknown
20
Base: 100 enquiries. The 20 unknown statuses remain part of this total.
What the author should be able to explain
The three categories are distinct and total 100.
The 20 unknown statuses do not automatically become failures.
45 / 80 = 56.25% would describe only enquiries with a known status: a different scope.
Can 56.25% replace 45% in the conclusion?
Only by explaining the changed denominator and the exclusion of 20 enquiries. Both calculations are possible, but answer different questions. Keep unknowns visible in the summary. Retain both classification dates and the same set of 100 enquiries. Here, the change from 45% to 55% comes from completed information, without new enquiries or evidence of an improved journey. Do not silently replace the result published before this review.
Another case to examine
Compare the audience mix as well
A second fictional dataset, separate from the 100 enquiries above: each period contains 110 enquiries, split differently between phone and desktop. Rates are rounded to one decimal where needed.
Scroll the table horizontally if needed. With a keyboard, focus the table and use the arrow keys.
Fictional data: groups improve, the overall rate falls
Period
Audience
Qualified
Total
Rate
Before
Phone
1
10
10%
Before
Desktop
80
100
80%
Before
Combined
81
110
73.6%
After
Phone
20
100
20%
After
Desktop
9
10
90%
After
Combined
29
110
26.4%
Both groups improve, yet the overall rate falls: the second period contains many more enquiries from the lower-rate group. A total does not describe this mix.
Calculate totals from counts: 81/110 then 29/110. Do not take an unweighted average of the group percentages or attribute these differences to a site change without further evidence.
Assess the quality of the work
For the exercise, two tables appear to report the same enquiry count but use different periods and counting units. Analysis should expose these differences before comparison. Examine identifiers, unknown values and calculation rules: a polished chart does not correct a poorly defined population.
Scroll the table horizontally to read every column. With a keyboard, focus the table area and use the arrow keys.
Assess the quality of the work
Skill
Evidence to examine
Failure to spot
Metric and population
Define what is counted: events, records or people, with period, relevant timezone and exclusions. Retain the calculation rule and data origin.
Repeated events are reported as distinct people, or a rate uses a denominator from another period.
Matching and unknowns
Prepare a fictional dataset with a duplicate, missing identifier and unknown value. Explain matching rules and the treatment of excluded rows.
A join multiplies rows unchecked, or an unknown value becomes zero and artificially changes a mean or proportion.
Reporting
The chart retains unit, period and a textual interpretation.
An average conceals an important group difference.
Denominator and period
Attach definition, population, period and missing-value handling to each ratio.
20 enquiries from 500 visits is 4%; collection changes are examined before any commercial explanation.
A focused next step
Manually recalculate a metric on a small documented dataset, then reconcile each row with the automated result. Explain differences before interpreting a trend. If collection changed between periods, state the comparison limit; a displayed increase may reflect counting methods rather than visitor behavior.
Product images are not embedded in the CSV. During import into another store, their URLs must still be accessible. Retain the original store until product, variant and image associations have been checked; inventory has a separate export.
Dates indicate when references were consulted. These facts are neither a site audit nor an overall compliance statement.
A review case to reproduce
A rate lacks comparable populations
Situation to examine
Two dashboards show conversion rate, but one uses visitors and another sessions. Their denominators differ.
Proposed check
Describe event, population, period and filters for each calculation. Recalculate a known small example and document exclusions. Explain definition differences before conclusions; missing measurements are not zero.
Evidence to retain
A metric record allows recalculation. It distinguishes real variation, methodological change and insufficient data.
Fictional finding example
Same period: 100 sessions, 80 distinct people, 8 people each sending one enquiry in 8 distinct sessions.
Interpret this outcome
Sessions with an enquiry: 8/100 = 8%. People with an enquiry: 8/80 = 10%. These rates describe different populations and are not interchangeable.
Three complementary checks
Fictional cases: adapt these expectations to your test contract and retain the limits of the check.
1Defined population
2Qualified value
3Bounded comparison
Scroll the table horizontally if needed; with a keyboard, focus it and use the arrow keys.
Scenario variations and expected evidence
Variation to test
Expected result in this case
Evidence to hand over
An observed value is zero
Keep zero as a known measurement
Source value and event definition
A value is absent
Flag the unknown without converting it to zero
Missing state and effect on calculation
The base changes from sessions to people
Show separate rates with their denominators
Population, period, filters and formula
Decision to hand over : Present differences in definitions before interpreting a rate change.
In-depth exercise: Analyze a small set of synthetic data on a research path. Define events and units before calculating an indicator. Introduce missing lines, duplicate and period variation. Compare the results before and after correction. Separate an event from a visitor and a volume of a rate. A numerical evolution does not prove its cause: formulate the possible explanations and indicate the verification that would make it possible to decide between them.
What to deliver with your work
Dictionary of fields and units
Demo data with identified anomalies
Reproducible calculations and cleaning rules
interpretation accompanied by the limits
Acceptance criteria
These criteria are a working basis to adapt to the context. For each, note the action taken, the expected result, the observed result and the information remaining to be confirmed.
The period and the denominator are indicated for each rate.
Duplicates and missing values have documented processing.
A calculation can be reproduced from the game provided.
The conclusion distinguishes between observation, hypothesis and proof of causality.
Deliver the test game, transformations and formulas with the result. Explain the exclusions and their effect on totals. Another person must be able to redo the calculation and find the same number. Charts should keep units and periods visible. If a production data has not been checked, do not replace this absence with artificial precision.
Analysis starts with a possible decision. Compare ongoing indicators with a one-off study according to definition stability and the depth of the question.
Indicator tracked over time
When to choose it
Useful for a workflow with a stable definition: action, population and period. Document exclusions and collection conditions.
What to plan for
Check changes to tools or tracking. A measurement discontinuity is different from a change in usage.
One-off analysis of a problem
When to choose it
Suitable for anomalies or clearly bounded situations. State the question and data required before building charts.
What to plan for
Check missing data, duplicates and incompatible scopes. Present possible explanations without confusing association with causation.
A risk to anticipate
An apparently precise dashboard can rest on ambiguous definitions. Include denominators, periods and limitations with important conclusions.
How to decide
Present a conclusion, supporting data and checks that could change it. Link it to an action or further investigation.
Compare questions, definitions and collection checks before dashboards. Tag installation differs from analysis supporting a verifiable decision.
Recurring operation
Plan quality monitoring, event revision and missing-period documentation. Separate tool costs from interpretation work.
Exit and handover
Request definitions, queries and authorised access scope. Transfer suitable results without unnecessary personal data; a chart alone cannot reproduce the calculation.
Define the event, unit, population and period, then show the counts used. Check duplicates and missing values on a small sample. For a rate, numerator and denominator must share the same definition; an unknown value does not become zero to fill a table.
Does an increased rate establish improvement?
Examine counts, traffic sources, measurement changes and seasonality too. An increase based on few enquiries may be fragile. State what the observations support and other plausible explanations before recommending a lasting change.
How can two periods be compared after a tracking change?
Identify comparable events and populations. Flag the discontinuity and avoid presenting the overall difference as a usage effect. If reliable comparison is impossible, explain the limit and rebuild a baseline with the new setup.