Two lenses. One insight.
Insight needs both. A number nobody understands and a story nobody can size are the two ways research fails.
The question each lens asks.
- 01 / 02Quantitative
The question it asks
Whether an insight is proven
Its own signal
Confidence
On probability. Every quantitative number is a probability statement, not a count. “62% agree” is 62 ± 5 with 95% probability, for this sample. “40% share of voice” is the probability a mention was captured, times the probability it was about you. “Cited in 3 of 10 LLM answers” changes on the next run. A count of seats is a count; usage is a probability per seat per week. The Quantitative lens is the discipline of knowing which of your numbers are probabilities (all of them) and how wide. See the impression exhibit below.
- 02 / 02Qualitative
The question it asks
Whether an insight is understood
Its own signal
Meaning
Q × Q = SI(a, t)
- Q × Q
- — Quantitative × Qualitative. Multiplication, not addition: either at zero and there is no insight.
- SI(a, t)
- — Scalable Insight, as a function of its actionability and its transferability.
The insight scales as a whole, and a and t are read separately.
a — Actionability
How comprehensively, versus narrowly, the insight can be activated.
Measure: surfaces it can direct ÷ surfaces available. Surfaces are the decisions in a campaign plan: audience, message, creative, the four POET channels (Partnered, Organic, Earned, Targeted), partner, timing — nine in the canonical set. An insight that changes only the message scores 1/9; one that changes audience, message, partner and two channels scores 5/9. Gate before scoring: it must name an action, not only a truth.
t — Transferability
How far the insight travels beyond where it was found.
Measure: contexts it held in ÷ contexts tested. Contexts are markets × audiences × time. Found in one market with one audience, t = 1/1: honest, and small. Held in 4 of 6 segments across two markets, t = 0.67. The honesty rule is the point: transfer can only be claimed over contexts actually tested. This is why segmentation is non-negotiable, why more cells buy more t, and why the time dimension is what carries an insight from Build into Prove.
Both lenses feed both properties. The tendency: Quantitative establishes t (it held across cells); Qualitative establishes a (the voice tells you what to do with it).
Actionability against transferability.
| Actionability | Low t | High t |
|---|---|---|
| High a | Narrow insight | Scalable insight |
| Low a | Observation | Platitude |
A platitude transfers everywhere and changes nothing. A narrow insight works once, in one place. An observation is a number on its own. Q2 is the upper right.
Capture / Protect filter. Every insight in the table is also tagged Capture (upside to take) or Protect (downside to prevent), so a Scalable insight reads as "scalable capture" or "scalable protect". Observations and Platitudes are discarded in Assessment; only Narrow and Scalable insights proceed to Strategy.
Optional score: a × t, 0 to 1, if a Q2SI number is ever wanted. The Standard remains a checklist; the score is a dial read after the checklist is passed.
Research in the probability era.
Traditional research counted things. A census counted people. A completed interview was a person who answered. A sale was a sale. The move from traditional to digital research and measurement changed the unit: almost every digital number is an estimate of an event that probably happened, not a record of one that did. The impression is the clearest example, and it is one of many.
What the digital KPIs actually are
| KPI | Reported as | What it really is |
|---|---|---|
| Impression | Times your ad was seen | Probability a screen had the ad on it |
| Viewable impression | Times it could be seen | ≥50% of pixels for ≥1s (2s video); opportunity, not attentionS1 |
| Reach / unique users | People | Cookies, devices and panel models standing in for people |
| Attribution | Which touch caused the sale | A model's allocation, from last-click to probabilistic MMM |
| Sentiment | Share positive / negative | A classifier's confidence, above a threshold someone chose |
| Share of voice | Your share of the conversation | Share of what the monitoring captured, times relevance |
| Search volume | Times people searched | Sampled, rounded, bucketed estimates |
| LLM citation | Whether the model cites you | Changes on the next run; a rate, never a fact |
| AI-coded open-ends | Theme counts | Classification probabilities summed |
Reported as counts, every one of them is a probability with a width the report rarely shows.
Worked example: the impression, per 100 served
Rounded; four-leg sourcing
| Step | ≈ |
|---|---|
| Served — a screen loaded the ad | 100 |
| Viewable (MRC) — opportunity to see | ~70 |
| Any human attention — at least one eye fixation | ~26 |
| Two seconds of attention | ~17 |
| Past the ~2.5s attention-memory threshold | ~15 |
Divide by frequency to get people; the average views-per-person then hides a distribution in which most people saw it zero times.
Sourcing stands on four legs, with the vendor last: the MRC/IAB standardS1; IAS scale data, with Lumen's 64% desktop rate in line with IASS2; Nelson-Field's academic work — 30% not MRC compliant, 44% viewable with zero attention, 17% viewable with two seconds, ~85% below the 2.5s memory threshold across 130,000 views and 1,150 brandsS3, S4, S5; and the Ehrenberg-Bass dissent that attention metrics may mislead and reach is reachS6. Lumen's own 100 → 70 → 35 funnel is illustrative corroboration and is labelled as such by LumenS7. Citing the dispute is the point: even the probability of a probability is contested.
What the probability era demands
- 01
Report the width, not just the point. Every number in a research report is a probability; the report's job is to say how wide.
- 02
Treat digital KPIs as opportunities and estimates, and buy attention and outcome measures separately when the decision needs them.
- 03
Put a voice behind the probability. The probability says how often the event happened; only the qual says what happened when it did. Interview thirty of the people reached and you learn what a 1.5-second fixation left behind, which is the only thing the million was for.
Qualitative in the age of AI.
Qual's importance is rising, not falling, because AI can now do three things with it that were previously impractical.
- 01
Personas with personality.
Quant describes a segment; qual gives it a voice, a vocabulary, a set of reasons. Built on enough real voice (per-cell minimums), a persona becomes something you can put complex and creative ideas in front of and discuss, test, and iterate against, before a single respondent is recruited. The rule: a persona is only as real as the qual beneath it. Synthetic personas grounded in real qual are instruments for pretesting; they are not findings.
- 02
Decision inputs.
How internal experts actually decide — what evidence they trust, what they discount, what they need to see before they move — is itself qualitative data, and it is the most under-collected source in the grid (Asked: internal expert perspectives). Capturing it lets the system reason the way the organisation's best people do, and lets the organisation see its own decision logic.
- 03
Open-ends that reveal.
In the probability era, the open-end is where the person shows up. Design them to reveal the group, not just the answer: projective and creative prompts (“describe it to a friend who has never heard of it”, “what would you tell your boss”, “finish the sentence: the real reason I don't is…”) tell you more about a segment than a hundred scale points. Read for the segment, not the response. This is where persona personality actually comes from.
The through-line: quant tells you how many and how wide; qual tells you who, in their words, well enough to build a person you can talk to.
Sources cited on this page.
- [S1]
MRC/IAB viewable-impression standard; viewability ≠ noticed. https://www.publift.com/blog/ad-viewability-optimization
- [S2]
Lumen/JCDecaux VAC study; Lumen desktop viewability 64% in line with IAS. https://jcdecaux.com/node/1658
- [S3]
Nelson-Field, K. (2020). The Attention Economy and How Media Works. Springer.
- [S4]
Nelson-Field / AANA interview: 30% not MRC compliant; 44% viewable with zero attention; 17% viewable with 2s. https://aana.com.au/karen-nelson-field-new-world-new-metrics/
- [S5]
VCCP Media × Nelson-Field (2025), Hacking the Attention Economy: 1.5s can encode memory; 85% of placements below 2.5s; 130,000 views, 1,150 brands (earlier Amplified dataset). https://www.vccp.com/uk/news/2025/may/hacking-the-attention-economy-vccp-media-and-dr-karen-nelson-field-reveal-1-5-second-formula-for-effective-digital-advertising
- [S6]
Sharp / Ehrenberg-Bass dissent on attention metrics. https://www.mi-3.com.au/05-04-2023/reach-curves-have-gone-rogue-karen-nelson-field-warns-scrollable-media-attention-decay
- [S7]
Lumen funnel (illustrative): 100 served → 70 viewable → 35 seen, 1.5s. https://communications.seenthis.co/information/lumen-E4
Prove it. Then bring it to life.
An insight is scalable when it passes all five tests of the Q2SI Standard. Three tests for confidence, two for meaning.