Triangulation
Date: 2026-08-17
Using several independent methods to answer one question. When methods with different weaknesses agree, confidence rises much faster than either alone justifies — and when they disagree, the disagreement is usually the more valuable finding.
Triangulation combines multiple methods, data sources or researchers to investigate one question, on the principle that each method’s blind spots differ.
Why it works
Every method has a systematic weakness. Independent methods have independent weaknesses, so agreement between them is unlikely to be produced by either method’s bias.
ANALYTICS what, at scale
✗ never why
✗ blind to the
unobserved
USABILITY TESTING why, in depth
✗ tiny sample
✗ artificial task
SURVEYS attitudes at scale
✗ non-response bias
✗ stated ≠ actual
SUPPORT TICKETS real problems, free
✗ only the people who
complained
SESSION REPLAY real behaviour, at scale
✗ no intent, easy to
over-read
Note that “stated versus actual behaviour” is the weakness of surveys and interviews, and not of analytics or replay — which is exactly why pairing them is productive.
The worked shape
QUESTION why do people abandon at
delivery?
ANALYTICS 34% abandon at the delivery
step; worse on mobile
→ the size, and where
REPLAY they scroll back up after
seeing the cost
→ the moment
TESTING "I thought delivery was free
over £30 — it's £40"
→ the mechanism
SUPPORT tickets asking about the
threshold
→ confirmation, independently
────────────────────────
four methods, one
conclusion, different
weaknesses
Any one of those alone is a hypothesis. Together they’re a finding — and it’s the kind of case that gets a fix prioritised.
Disagreement is the valuable case
When methods conflict, don’t average them — investigate.
SURVEY "delivery speed is what
matters most"
ANALYTICS the cheapest delivery option
is chosen 80% of the time
NOT "people are lying"
BUT stated preference and revealed
behaviour differ, which is
normal
→ they SAY speed and CHOOSE
price
→ both are true; they apply
in different moments
Three explanations to work through before concluding either is wrong:
1 measuring different things
(intent vs behaviour, different
populations, different timeframes)
2 one method's known bias is
operating
3 the finding is genuinely
conditional — true for one segment
and not another
Explanation 1 is by far the most common, and it’s the same discipline as diagnosing a metric discrepancy — Tool Discrepancies.
The kinds
METHODOLOGICAL different methods
← the usual meaning
DATA different sources or
timeframes for the same
method
INVESTIGATOR different researchers
analysing the same data
— Qualitative Coding
THEORETICAL different frameworks
interpreting the same
findings
See: Qualitative Coding
The practical loop
QUANT find and size the problem
— Funnel Analysis
↓
QUAL explain it
— Usability Testing
↓
HYPOTHESIS
↓
EXPERIMENT did the fix work?
— A-B Tests
↓
QUANT new baseline
See: Funnel Analysis · Usability Testing · A-B Tests
Skipping the qual step is the most common failure — it produces tests of solutions to guessed-at causes, which is why win rates are low in programmes that only look at funnels — Hypothesis Design.
Where it goes wrong
- Methods that aren’t independent. Two surveys to the same list is one method twice
- Using it to justify a decision already made. Running methods until one agrees
- Treating agreement as proof. Correlated biases produce correlated errors — several methods relying on the same instrumentation share its faults
- Over-triangulating. Not every question needs four methods. A clearly-diagnosed problem doesn’t need corroboration to be fixed
- Averaging conflicting results rather than investigating the conflict, which discards the most informative outcome