An observation is something the team saw or heard during testing, written down exactly as it happened. It records what people did or said, without explaining why.
For example, during a concept test of the SecondBank rewards dashboard, 7 of 10 new cardholders scrolled past their points balance without reading it. That is the observation. Deciding what it means comes later.
Why it matters
When teams mix what they saw with what they think it means, the evidence gets bent to fit the story. A room full of people can watch the same session and walk away with different memories of it, and decisions end up resting on whoever retold it best.
Clean observations fix that. They give everyone the same raw facts to argue from, they make it easy to count how often something happened, and they let anyone check a conclusion against what was actually seen.
Write down who did or said what, during which task, in plain words. Keep the interpretation for later, where it can be judged on its own.
How we write one
Every observation is written as one sentence that names who, what they did or said, and what happened:
A person or a count did or said what happened.
7 of 10 participants scrolled past the points balance without reading it.
Writing it this way keeps the sentence about behavior or words, and the count shows at a glance whether it was one person or most of them.
How it differs from similar ideas
-
An assumption is a conclusion drawn from data and treated as true before anyone checks it. An observation is the data itself, with no conclusion attached.
-
A signal interprets evidence and says what it means for a test. An observation only reports what happened; a signal is built on top of it.
-
A UX metric is a measured value, such as task success of 42%. An observation is a specific moment or pattern, such as people scrolling past the balance.
-
An input is the raw material, such as a session recording or transcript. An observation is one thing pulled out of that input.
-
A hunch is what someone suspects. An observation is what someone saw.
A good observation
-
Describes behavior or words, not meaning
-
Says who or how many
-
Names the task or moment
-
Uses a direct quote when someone spoke
-
Can be traced to a session or recording
-
Comes from one round of testing
What happens to it
-
Seen or heard. Something happens during a test session, interview or review.
-
Recorded. A note taker, the researcher or Ray writes it down from notes, a transcript or a recording.
-
Captured in Findings. It's added to the Findings for that round of testing, alongside the metrics and signals.
-
Interpreted. A signal cites it as evidence and says what it means.
-
Kept as evidence. It stays linked to the signals and decisions that rest on it, so anyone can check them later.
For Ray
This part tells Ray how to find observations in an input, such as a test session transcript, a recording summary, a note taker's log or a research readout, and how to store them.
How to find one
-
Mark moments of behavior or speech. Read the input and mark every place where a participant did something or said something during a task.
-
Separate fact from interpretation. Strip out any explanation the note taker added. Keep "scrolled past the balance" and drop "because they didn't care"; the explanation may become a Hunch or Signal on its own.
-
Find who or how many. Record the participant ID, or count how many participants did the same thing out of how many were tested.
-
Find the task or moment. Note which task, test question or screen it happened on.
-
Keep the quote or behavior. Copy spoken words exactly in quotes. Describe actions in plain, physical terms: clicked, scrolled, paused, asked, gave up.
-
Link it to the test. Identify the test, concept and audience the session belongs to.
-
Write the sentence. Use the format: [Who or how many] [did or said] [what].
-
Fill the record, score it, link it and store it using the rules below.
Cues to look for
-
Action verbs tied to a participant: "P4 clicked", "she scrolled past", "they hesitated on"
-
Direct quotes from participants in a transcript or notes
-
Counts across sessions: "7 of 10", "most participants", "only two people"
-
Timestamps or task labels in a session log or recording summary
What doesn't count
<table xmlns="http://www.w3.org/1999/xhtml" style="min-width: 50px;"><colgroup><col style="min-width: 25px;"><col style="min-width: 25px;"></colgroup><tbody><tr><th colspan="1" rowspan="1"><p>If the statement...</p></th><th colspan="1" rowspan="1"><p>Store it as</p></th></tr><tr><td colspan="1" rowspan="1"><p>Explains why something happened, such as "they don't trust points"</p></td><td colspan="1" rowspan="1"><p>Hunch or Signal, linked to the observation it rests on</p></td></tr><tr><td colspan="1" rowspan="1"><p>Reports a measured value, such as task success or a SUS score</p></td><td colspan="1" rowspan="1"><p>UX Metric</p></td></tr><tr><td colspan="1" rowspan="1"><p>Draws a conclusion from analytics before anyone has tested it</p></td><td colspan="1" rowspan="1"><p>Assumption</p></td></tr><tr><td colspan="1" rowspan="1"><p>Suggests what to change</p></td><td colspan="1" rowspan="1"><p>Recommendation</p></td></tr><tr><td colspan="1" rowspan="1"><p>Describes the session file or transcript itself</p></td><td colspan="1" rowspan="1"><p>Input</p></td></tr></tbody></table>
The record
<table xmlns="http://www.w3.org/1999/xhtml" style="min-width: 75px;"><colgroup><col style="min-width: 25px;"><col style="min-width: 25px;"><col style="min-width: 25px;"></colgroup><tbody><tr><th colspan="1" rowspan="1"><p>Field</p></th><th colspan="1" rowspan="1"><p>What to fill in</p></th><th colspan="1" rowspan="1"><p>Values</p></th></tr><tr><td colspan="1" rowspan="1"><p>Statement</p></td><td colspan="1" rowspan="1"><p>The one-sentence observation</p></td><td colspan="1" rowspan="1"><p>[Who or how many] [did or said] [what].</p></td></tr><tr><td colspan="1" rowspan="1"><p>Who</p></td><td colspan="1" rowspan="1"><p>The participant or participants</p></td><td colspan="1" rowspan="1"><p>Participant IDs, or a count such as "7 of 10"</p></td></tr><tr><td colspan="1" rowspan="1"><p>Kind</p></td><td colspan="1" rowspan="1"><p>Whether it was an action or speech</p></td><td colspan="1" rowspan="1"><p>Behavior, Quote</p></td></tr><tr><td colspan="1" rowspan="1"><p>What happened</p></td><td colspan="1" rowspan="1"><p>The behavior described plainly, or the exact quote</p></td><td colspan="1" rowspan="1"><p>Short text</p></td></tr><tr><td colspan="1" rowspan="1"><p>Task or moment</p></td><td colspan="1" rowspan="1"><p>Where in the session it happened</p></td><td colspan="1" rowspan="1"><p>Task, test question or screen</p></td></tr><tr><td colspan="1" rowspan="1"><p>Audience</p></td><td colspan="1" rowspan="1"><p>Whose evidence it is</p></td><td colspan="1" rowspan="1"><p>Link to Audience</p></td></tr><tr><td colspan="1" rowspan="1"><p>Test</p></td><td colspan="1" rowspan="1"><p>The test and concept being evaluated</p></td><td colspan="1" rowspan="1"><p>Link to Test and Concept</p></td></tr><tr><td colspan="1" rowspan="1"><p>Findings</p></td><td colspan="1" rowspan="1"><p>The round of testing it belongs to</p></td><td colspan="1" rowspan="1"><p>Link to Findings</p></td></tr><tr><td colspan="1" rowspan="1"><p>Source</p></td><td colspan="1" rowspan="1"><p>The input and exact passage or timestamp</p></td><td colspan="1" rowspan="1"><p>Link to Input</p></td></tr><tr><td colspan="1" rowspan="1"><p>Status</p></td><td colspan="1" rowspan="1"><p>Where it stands</p></td><td colspan="1" rowspan="1"><p>Recorded, Captured, Interpreted</p></td></tr><tr><td colspan="1" rowspan="1"><p>Confidence</p></td><td colspan="1" rowspan="1"><p>How directly it was recorded</p></td><td colspan="1" rowspan="1"><p>High, medium, low</p></td></tr></tbody></table>
New observations start as Recorded. They move to Captured when linked to Findings, and to Interpreted once a signal cites them.
Scoring confidence
Confidence measures how directly the observation was recorded, not how important it seems.
<table xmlns="http://www.w3.org/1999/xhtml" style="min-width: 50px;"><colgroup><col style="min-width: 25px;"><col style="min-width: 25px;"></colgroup><tbody><tr><th colspan="1" rowspan="1"><p>Level</p></th><th colspan="1" rowspan="1"><p>When to use it</p></th></tr><tr><td colspan="1" rowspan="1"><p>High</p></td><td colspan="1" rowspan="1"><p>Taken from a transcript, recording or live notes, with who, the task and the exact behavior or quote</p></td></tr><tr><td colspan="1" rowspan="1"><p>Medium</p></td><td colspan="1" rowspan="1"><p>Taken from a summary or notes, with the count or the task missing</p></td></tr><tr><td colspan="1" rowspan="1"><p>Low</p></td><td colspan="1" rowspan="1"><p>Recalled after the fact, secondhand, or mixed with interpretation that can't be separated</p></td></tr></tbody></table>
If any part of the record can't be filled from the source, leave it blank and lower confidence by one level.
Linking
-
Check for duplicates first. If an existing observation from the same round describes the same behavior on the same task, update its count and add the new source instead of creating another.
-
Link to its test, concept and audience, and to its source input.
-
Link to the Findings for that round of testing.
-
Link to any signal that cites it as evidence.
-
Note contradictions. If one participant did the opposite of the pattern, store that as its own observation and link the two as contrasting.
When to store and when to ask
-
Store automatically when confidence is high.
-
Add to the review digest when confidence is medium or low, when the fact can't be separated from interpretation, or when it contradicts an open test's expected result.
When to flag
-
A signal has no observation or UX metric behind it
-
An observation carries an interpretation that was never separated out
-
An observation has no test or Findings linked to it
-
A pattern rests on a single participant but is described as common
-
Observations from one round contradict each other with no signal reconciling them
Worked example
Input: a note taker's log from the SecondBank rewards concept test, sessions 1 to 10.
"Task 2, find your points. P1, P3, P4, P5, P7, P8, P9 scrolled straight past the balance card, didn't stop on it. P4: 'Is that number money or something else?' Seems like nobody cares about points."
What Ray finds:
-
"P1, P3, P4, P5, P7, P8, P9" is seven participants out of ten tested.
-
"Scrolled straight past the balance card" is the behavior, and "Task 2, find your points" is the moment.
-
P4's question is a direct quote, so it becomes a second observation of kind Quote.
-
"Nobody cares about points" is the note taker's interpretation, so it's dropped from the observation and noted as a possible Hunch.
Stored record:
<table xmlns="http://www.w3.org/1999/xhtml" style="min-width: 50px;"><colgroup><col style="min-width: 25px;"><col style="min-width: 25px;"></colgroup><tbody><tr><th colspan="1" rowspan="1"><p>Field</p></th><th colspan="1" rowspan="1"><p>Value</p></th></tr><tr><td colspan="1" rowspan="1"><p>Statement</p></td><td colspan="1" rowspan="1"><p>7 of 10 participants scrolled past the points balance without reading it.</p></td></tr><tr><td colspan="1" rowspan="1"><p>Who</p></td><td colspan="1" rowspan="1"><p>7 of 10 (P1, P3, P4, P5, P7, P8, P9)</p></td></tr><tr><td colspan="1" rowspan="1"><p>Kind</p></td><td colspan="1" rowspan="1"><p>Behavior</p></td></tr><tr><td colspan="1" rowspan="1"><p>What happened</p></td><td colspan="1" rowspan="1"><p>Scrolled past the balance card without stopping</p></td></tr><tr><td colspan="1" rowspan="1"><p>Task or moment</p></td><td colspan="1" rowspan="1"><p>Task 2: find your points</p></td></tr><tr><td colspan="1" rowspan="1"><p>Audience</p></td><td colspan="1" rowspan="1"><p>Participants: new cardholders, first 90 days</p></td></tr><tr><td colspan="1" rowspan="1"><p>Test</p></td><td colspan="1" rowspan="1"><p>Abstract-points concept test; current dashboard</p></td></tr><tr><td colspan="1" rowspan="1"><p>Findings</p></td><td colspan="1" rowspan="1"><p>Rewards value test, round 1</p></td></tr><tr><td colspan="1" rowspan="1"><p>Source</p></td><td colspan="1" rowspan="1"><p>Note taker's log, sessions 1 to 10, October 2026</p></td></tr><tr><td colspan="1" rowspan="1"><p>Status</p></td><td colspan="1" rowspan="1"><p>Captured</p></td></tr><tr><td colspan="1" rowspan="1"><p>Confidence</p></td><td colspan="1" rowspan="1"><p>High: live notes with participant IDs, task and behavior</p></td></tr></tbody></table>
Next step: stored automatically and captured in the round 1 Findings, ready for a signal to interpret it.
Edge cases
-
Interpretation baked in. "Users were confused by the balance" mixes a feeling with a fact. Look for the behavior behind it; if none is given, store at low confidence and send it to the review digest.
-
One participant. A striking moment from one person is still an observation. Record it as 1 of 10 so nobody mistakes it for a pattern.
-
Counts that don't add up. If notes say "most" but list four of ten, use the listed count and flag the mismatch.
-
Observations outside a test. Something seen in a support call or a sales demo is an Input that may inform a Situation. Only store it as an observation when it comes from a test session.