What is an insight, and why most reports do not contain one

The quarterly deck that nobody reads

Every quarter, the five hotels of Ardcloon Hotels send their figures to head office in Galway, and every quarter a 38 page PDF comes back. Occupancy by hotel. Average daily rate by month. Revenue by channel. Guest satisfaction by question. Twelve slides on the loyalty programme. It is accurate, it is on time, and it is beautifully formatted.

Ardcloon is an invented small hotel group in Connemara, Co. Galway, and every number in this chapter is made up so I can show each step. But I have seen this deck, in slightly different fonts, at a dozen companies.

In the example, the managing director reads the first three pages and then skims the rest for anything red. Her hotel managers do not open it at all. One of them, quite reasonably, puts it like this: "It tells me what happened. I was there. I know what happened."

That sentence is the whole chapter. The report contains several hundred facts and not one insight, and the difference between the two is what the reader can do next. Most analytics output fails at exactly this point: it stops at the finding. This chapter is about the last step, the one that turns a number into a decision, and about why so few teams take it.

Where this sits in the series

This is the third and last chapter of part one of the series, how analytics helps. In the previous chapter, on consumer behaviour as the basis of marketing strategy, I argued that the behaviour of real customers, not the channel report, is the raw material of strategy. Here I close part one by asking what all that analysis is for, and the answer is not a chart.

From the next chapter, on modelling demand and price elasticity with regression, the series moves into part two, dependent variable techniques, where the modelling proper begins.

Reporting, analysis and insight are three different jobs

Reporting answers the question "what happened". Occupancy was 78%. Direct booking share was 41%. Reporting is necessary, it should be automated, and it should be boring.

Analysis answers "why did it happen". It cuts the number by hotel, by device, by guest type, by month, and looks for the place where the change is concentrated. But analysis on its own still leaves the reader with a shrug. You now know why direct share fell. Fine. What would you like me to do about it?

Insight generation answers "so what do I do, and what is it worth". This is the step almost everyone skips, because it asks the analyst to leave the safe ground of describing the data and take a position on the business.

My working definition: a finding is a true statement about the data. An insight is a finding that has been attached to a decision, priced, and had its risk stated out loud. If any of those three parts is missing, it is still a finding, however clever it is.

The four tests an insight must pass

An insight has to pass four tests, in this order, because the last one matters most.

1. New. The reader did not already know it. "Weekends are busier than midweek" fails this test at every hotel on earth.
2. Non obvious. It could not have been guessed from the surface of the numbers. It usually comes from a cut, a comparison or a model, not from the headline.
3. Explains why. It names a mechanism, not a correlation. "Guests who use the spa spend more" is a correlation. "Guests who book the spa at check in stay longer because the spa slot anchors their departure day" is a mechanism, and it can be wrong, which is exactly what makes it useful.
4. Actionable. It names a decision that someone in the room has the authority to take, with a size and a risk attached.

Three out of four is trivia. The fourth test is the one that separates an analytics function from a very expensive newsletter, and I wrote about that mindset in Start with the decision, not the dashboard.

An example: "Our Clifden hotel had its best August on record" is true, and the general manager already knew it, so it fails the first test before it reaches the others. "August was strong because a walking festival filled the town for a week" passes the first three tests. It only passes the fourth once it becomes "Ask the festival organisers now for next year's dates, and open a festival package on the hotel's own site eight weeks earlier than this year", because now someone can say yes or no.

Pricing the decision with expected value

If an insight must name a decision, it must also say what the decision is worth, and that needs one small piece of maths. The expected value of a decision is the probability weighted sum of what happens if you take it:

$$EV(d) = \sum_{i=1}^{n} p_i \, v_i$$

Here $d$ is the decision, $i$ runs over the $n$ outcomes you think are plausible, $p_i$ is the probability you assign to outcome $i$, and $v_i$ is the value of that outcome in money, usually per year. The probabilities $p_1$ to $p_n$ must add up to one. They are judgements, not measurements, and the honest way to present them is as judgements. Foundations: expected value.

The expected value on its own is not the number that matters. The number that matters is how much better off you are than if you did nothing, after paying for the change:

$$\Delta EV = EV(\text{act}) - EV(\text{status quo}) - C$$

$\Delta EV$ is the value of the decision, $EV(\text{act})$ is the expected value if you take the decision, $EV(\text{status quo})$ is the expected value if you carry on as you are, and $C$ is the cost of making the change, including the time of the people it takes. If $\Delta EV$ is positive, the decision is worth taking on expectation, though a small positive number with a large downside still deserves the risk line the brief carries later in this chapter. Foundations: expected value.

I covered the fuller version of this, including what it is worth to buy more information before deciding, in The three questions your data should answer, and the decision theory behind it in Decision theory for founders.

Worked example: a small decision priced

A hotel considers a €2,000 change to its breakfast booking page. The example assumes three outcomes: a 50% chance it adds €6,000 a year, a 30% chance it adds €2,000, and a 20% chance it adds nothing. If nothing changes, the status quo adds nothing either, so $EV(\text{status quo}) = 0$.

Step one, the expected value of acting: $EV(\text{act}) = 0.5 \times 6{,}000 + 0.3 \times 2{,}000 + 0.2 \times 0 = 3{,}000 + 600 + 0 = 3{,}600$.

Step two, the value of the decision: $\Delta EV = 3{,}600 - 0 - 2{,}000 = 1{,}600$.

The decision is worth about €1,600 in year one on expectation. The worst case is minus €2,000, the cost with nothing to show for it, and that sentence goes in the brief too.

Scoring candidates so the word stays honest

Candidate insights are scored on the four tests plus the size of the prize, with actionability weighted double:

$$S = N + O + W + 2A + P$$

$S$ is the score, $N$ is newness (0 to 2), $O$ is non obviousness (0 to 2), $W$ is whether it explains why (0 to 2), $A$ is actionability (0 to 3, counted twice) and $P$ is the size of the prize (0 to 3), for a maximum of 15. It is a weighted sum, the same building block as the expected value. Foundations: expected value.

Anything below 8 goes in the appendix. Anything with $A$ of zero goes nowhere, regardless of the total.

The rubric is mine, and the weights are a judgement, not a law. What matters is that the team agrees them before the scoring starts, so nobody tunes them to rescue a favourite.

The same numbers, written three times

Back to Ardcloon. Page 14 of the deck contains a perfectly good finding.

Worked example, version one: the finding

"Direct booking share fell from 44% to 41% in Q3 compared with the same quarter last year. Platform bookings rose correspondingly."

True. Formatted nicely. Ignored by everyone, because nobody knows what to do with it. Three points could be seasonal mix, a platform promotion, or noise.

Worked example, version two: the analysis

Cut by hotel, device and guest type, the story gets specific. The fall is entirely concentrated in two hotels, Ardcloon Clifden and Ardcloon Roundstone, which moved to the group's new booking engine in June. In those two hotels the mobile completion rate of the direct booking flow dropped from 3.1% to 1.9% after the new date and room selection step went live. Desktop completion did not move. The other three hotels, still on the old engine, are flat year on year.

The guests being lost are mostly returning guests, whose direct share fell from 70% to 61%. They start on the hotel site on their phone, give up at the new step, and finish on a platform app that already knows their card details.

Now we know why. That is analysis, and still nobody has to do anything.

Worked example, version three: the insight

"Fix the mobile date and room step on the new booking engine at Clifden and Roundstone, and give returning guests a stored profile plus a direct rate reminder three weeks before their usual rebooking window. Group room revenue is around €9.6 million a year, so three points of share is €288,000 routed through platforms at 15% commission, or about €43,200 a year in commission, plus the guest relationship. Expected recovery is about €25,200 a year against a cost of about €9,000, so the decision is worth roughly €16,200 in year one and the full recovery in every year after. The main risk is that the shift is partly a mix effect. The share of first time guests is flat year on year, so I do not think it is, but the group will know within one quarter."

The 15% is the example's assumption. Online travel agents typically charge commission of 15% to 25% on every booking, according to a Which? investigation in the UK reported in 2020,¹ so the example sits at the low end on purpose: a brief that needs the high end to work is not a good brief.

That paragraph is what an insight looks like. A decision, an owner (the group's web team), a prize with its working shown, a probability weighted expectation, a cost, a risk, and a date when the group will know if it was wrong. Here is the working, because an insight that hides its arithmetic asks to be trusted rather than checked.

Outcome	Probability	Commission saved per year	Weighted value
Full recovery of the three points	0.45	€43,200	€19,440
Partial recovery of one point	0.40	€14,400	€5,760
No measurable effect	0.15	€0	€0
Expected value of acting	1.00		€25,200
Cost of the change			€9,000
Value of the decision, year one			€16,200

Read it line by line. The first row says that if the fix does what the funnel data suggests, the group stops paying 15% commission on €288,000 of bookings, which is $0.15 \times 288{,}000 = 43{,}200$, and I give that slightly less than an even chance because booking engines are never fixed as quickly as the vendor promises.

The second row is the realistic partial outcome: the step is fixed but the returning guests have already formed the platform habit, so only one point comes back. One point of €9.6 million is €96,000, and 15% of that is €14,400.

The third row is honest failure, and I have never written a brief without one. Sum the weighted column, $19{,}440 + 5{,}760 + 0 = 25{,}200$, subtract the cost, and you have the €16,200 that goes at the top of the brief. Doing nothing saves no commission, so the status quo counts as zero here. The worst case is minus €9,000, which the managing director can live with, and that sentence belongs in the brief too.

The €9.6 million, the 15% and the probabilities are invented. The structure is not.

Scoring five candidate insights

The same deck yielded five statements that people in the room wanted to call insights, scored here against the rubric.

Candidate	New	Non obvious	Explains why	Actionable (x2)	Prize	Score
Direct booking share fell three points	0	0	0	0	1	1
Mobile abandonment at the new date step explains the lost direct share at two hotels	2	2	2	3	2	14
Returning guests drift to platforms because the platform remembers them and the hotel does not	1	2	2	2	2	11
Weekend occupancy is higher than midweek	0	0	0	0	0	0
Guests who use the spa spend 40% more per stay	1	1	0	1	1	5

Take the second row as the check on the arithmetic: $S = 2 + 2 + 2 + 2 \times 3 + 2 = 14$.

The first row is the finding from the deck. It scores one point, for the prize, because there is money in it somewhere, and fails every other test: everyone knew it, it has no mechanism and it names no decision.

The second row is the insight that ends up in the brief. It scores 14 of 15. It loses a point on prize only because €43,200 a year is real money but not transformational at €9.6 million of room revenue.

The third row, the companion insight about returning guests, is slightly less new because the general manager had suspected it, and slightly less actionable because stored profiles are a bigger project than fixing one step. It still clears the bar and goes into the same brief as a second phase.

The fourth row is there because it appeared in the deck, in colour, on page 9, every quarter for three years.

The fifth row looks like an insight and is not. Spa users spend more. Of course they do: they bought a spa treatment. Without a mechanism the statement cannot tell you whether promoting the spa would raise spend or merely relabel the guests who were going to spend anyway. I wrote about this trap in Is your dashboard lying to you?. Score five, appendix, and a note that a proper test might promote it.

From data to decision

The path from a database to a changed decision has more gates than most teams draw, and the gates are where the value is created.

Read it from the top. Every finding meets two gates before it becomes an insight. Most belong in the appendix, and that is fine. When you cannot price a candidate, you go back and find the missing number. You do not write a vaguer brief. The last step closes the loop: the measured result is the data the next review starts from.

The second chart is the reason executives ignore analytics output, drawn as plainly as I can.

The horizontal axis is the quarter, the vertical axis a simple count. The upper line is the pages of reporting produced, the lower line the decisions the management team could trace back to a report. The figures are invented, but I have rarely seen the lower line above five.

Forty pages a quarter and, on a good quarter, three decisions. The gap is what happens when nobody is asked to take the fourth step. The fix is not fewer pages. It is a different first question.

The one page insight brief

Everything an executive needs from an insight fits on one page, and if it does not fit, the insight is not finished. Here is the template I use.

INSIGHT BRIEF                                   Date / Author / Version

DECISION      One sentence. Who does what, by when.
EVIDENCE      Three to five bullet points. Each one a number with its source.
PRIZE         Expected value per year, with the working. Best and worst case.
RISK          What would make us wrong, how we would notice, and how soon.
COST          Money and people time, including the people who will resist it.
OWNER         One name. Not a department.
CHECKPOINT    The date we look at the result, and the number we look at.

The decision comes first because the reader should know within five seconds what they are being asked to do. The risk sits in the middle rather than in a footnote, because an executive who has been given the risk in plain words will trust the prize, and one who has not will halve it anyway.

The habit that makes all of this work is starting with the decision. Not "what does the data say" but "what decisions are on the table in the next quarter, and which number would change each of them". If you start from the data instead, you will generate fifty findings and then hunt for a decision to attach to each one, which is how you end up with the spa row.

Worked example: the Ardcloon brief, filled in

• Decision: the web team fixes the mobile date and room step at Clifden and Roundstone by 31 March.
• Evidence: direct share at the two hotels fell from 44% to 41%. Mobile completion fell from 3.1% to 1.9% after the June switch. Desktop flat. The three hotels on the old engine flat. Returning guests' direct share fell from 70% to 61%.
• Prize: €25,200 a year expected, best case €43,200, worst case nothing.
• Risk: a mix effect. First time guest share is flat, which argues against it. Visible within one quarter.
• Cost: about €9,000, mostly the booking engine vendor's time.
• Owner: one named person on the web team.
• Checkpoint: 30 June, the mobile completion rate at the two hotels against 3.1%.

The hamper shop version

A second, smaller example. Priya runs Northgate Hampers, an invented online hamper shop in Cork at northgatehampers.example, and her corporate customers order in a tight window before Christmas.

Worked example: from finding to insight at Northgate Hampers

The finding, straight from the shop's admin: 40 corporate accounts ordered last year, for a total of €26,000 in the fourth quarter. Interesting, known, and it names no decision.

The analysis: of the corporate accounts from the year before, 24 of 40, or 60%, ordered again last year, nearly all between 28 November and 8 December. Any account that had not ordered by 10 December did not order at all. So 16 accounts, 40%, lapsed. If they spend like the rest, that is $0.40 \times 26{,}000 = 10{,}400$, about €10,400 of corporate revenue quietly walking away each year.

The insight: on 27 November, send every corporate account from last year a short personal email from Priya with last year's order prefilled and a one click reorder link. The example assumes the reminder wins back a quarter of the accounts that would otherwise lapse: $16 \times 0.25 = 4$ accounts, and $10{,}400 \times 0.25 = 2{,}600$, so around €2,600 of revenue for an afternoon's work. The downside is a few unsubscribes. Owner: Priya. Checkpoint: 12 December, compare the reorder rate of the emailed accounts with last year's 60%.

A decision, a prize with its arithmetic, a risk and a date, in one paragraph. The numbers are invented. The timing pattern is one I would bet on in any gifting business.

Running an insight review

The insight review is a thirty minute meeting once a month, and it replaces the reporting walkthrough.

Before the meeting, the management team lists the decisions that are actually due in the next ninety days. Pricing for next season. Whether to renew the platform contract. Five to ten items. For each one, someone writes down the number that would change it. This step alone usually kills half the standing reports.

The analysts then bring candidate insights, each scored on the rubric, each on one page in the brief format. Anything under 8 is not presented. It goes in a shared appendix that anyone may read and nobody must.

In the meeting, each brief gets five minutes: decision, prize, risk, questions. The outcome for each is one of three words: yes, no, or more, where "more" means a specific piece of analysis with a name and a date. Every yes gets an owner and a checkpoint before the meeting ends.

Worked example: the query behind the Ardcloon brief

The data work behind a brief is rarely exotic. For Ardcloon it was one query against the bookings table:

select h.name as hotel,
       b.device,
       case when g.stays_before > 0 then 'returning' else 'first time' end as guest_type,
       date_trunc('quarter', b.booked_at) as quarter,
       avg(case when b.channel = 'direct' then 1.0 else 0.0 end) as direct_share,
       count(*) as bookings
from bookings b
join hotels h on h.id = b.hotel_id
join guests g on g.id = b.guest_id
where b.booked_at >= date '2024-07-01'
group by 1, 2, 3, 4
order by 1, 4, 2, 3;

Each row is one hotel, device, guest type and quarter. The average of a column that is 1 for a direct booking and 0 otherwise is simply the direct share. The query starts in July 2024 so that Q3 of 2025 has last year's Q3 to compare with. The skill is not in the SQL but in knowing which cut to ask for, which comes from having a decision in mind first.

Four checks before writing the brief

Before I trust a candidate enough to write the brief, I run four checks. If one fails, the candidate goes back to "more".

1. Is the sample big enough that the shift is not noise? A share moving three points on a few thousand bookings is unlikely to be chance, a segment of two hundred bookings might well be.
2. Is it a mix effect? That means the population changed rather than the behaviour. A total can move, or even reverse, while every group inside it stays put, which statisticians know as Simpson's paradox.² That is why the first time versus returning split matters.
3. Does the mechanism show up somewhere else in the data? Here, in the funnel completion rates.
4. Would I find the reverse story equally convincing if the numbers had moved the other way? This is the check that catches most of my own wishful thinking.

Worked example: is three points noise?

Take the first check with numbers. Suppose each of the two Q3s holds 3,000 bookings at the two hotels, 44% direct last year and 41% this year. The pooled share is $p = 0.425$. The standard error of the difference is

$$\sqrt{0.425 \times 0.575 \times \left(\tfrac{1}{3{,}000} + \tfrac{1}{3{,}000}\right)} = \sqrt{0.2444 \times 0.000667} \approx 0.0128$$

and $z = 0.03 / 0.0128 \approx 2.35$, which gives a two sided p value of about 0.02. On 200 bookings a side the standard error is about 0.049 and $z \approx 0.61$: the same three points, and nothing you could act on. Foundations: hypothesis tests.

Six ways this goes wrong

Insight theatre. The word gets applied to everything. "Key insight: revenue grew 4%." Once every bullet point is an insight, the executive stops reading them all, including the one that mattered. Guard the word.

Correlation wearing a mechanism's coat. The spa row again. A plausible story that has not been tested is more dangerous than no story, because it stops people looking. If the "why" has not been checked against a second cut of the data, say so in the risk line.

A prize without a probability. The headline number in most decks is the best case, presented as the expectation. "This could be worth €43,200 a year." It could. It is expected to be worth about €25,200, and that gap decides whether the executive trusts your next brief. Always show the weighted column.

The action nobody owns. "We should improve the mobile booking experience." Who? By when? An insight that ends in the passive voice is a finding with ambitions. Put one name on it, and if nobody in the room can be that name, say whose decision it really is.

The decision that was already made. Sometimes the analysis is commissioned to justify a decision the board took last month. When that is happening, the honest answer is to say so and save everyone the fortnight. I have a longer piece on how this kills dashboard projects in particular: Why your dashboard project keeps dying on the shelf.

Waiting for certainty. The opposite failure. Because the probabilities are judgements, some teams refuse to write them down and wait for a number that never arrives. A written 0.45 that turns out wrong teaches you something. An unwritten "probably" teaches you nothing, and the season is over by the time you have argued about it.

How I do this for clients

I start with the free workshop, with the people who take decisions, not the people who make the slides. Together we list the decisions due in the next ninety days and, for each one, the number that would change it. Most clients leave that room with a shorter list of reports than they came in with.

Then I take two weeks of real work on me, before you commit to anything. I need read access to the systems the decisions depend on: the booking engine or shop database, the analytics account, the finance export. In those two weeks I produce the first three insight briefs in the format above, scored, priced, with the risk written down, and I run the first insight review with your team so the habit has been practised once before I step back.

What you get after that, if you continue, is not a dashboard. It is a monthly set of one page briefs, each tied to a decision on your list, each with an expected value and a checkpoint date, plus one dashboard if and only if it answers a question you ask every week. You own all of it: the queries, the briefs, the scoring sheet.

Where a brief needs a proper model behind it, that is the Data Science work covered in the rest of this series. Where the decision is a marketing one, Growth Plus pairs a monthly base fee with commission only on the additional sales I generate.

Costs are on the pricing page in plain terms: monthly terms, final prices, and the free workshop plus two weeks on me before you commit.

Questions to ask your own team

Take these to the next reporting meeting. If more than two get a blank look, the reports are not producing insights.

• Which decision, due in the next ninety days, does this report change, and which number in it would change that decision?
• For the last three things we called an insight, what did we do differently afterwards, and what did it turn out to be worth?
• What is the expected value of the recommendation on this page, and what probabilities did you assume to get there?
• What would have to be true for this recommendation to be wrong, and when would we notice?
• Who owns this action, by name, and what is the date we look at the result?
• If we stopped producing this report tomorrow, who would notice, and what would they do instead?

This series is inspired by Mike Grigsby's Marketing Analytics.³ The explanations, examples and numbers here are my own.

Footnotes

1. Gary Noakes, Which? says avoid OTAs when booking hotels, TTG Media, 3 March 2020, https://www.ttgmedia.com/news/which-says-avoid-otas-when-booking-hotels-21546

2. Jan Sprenger and Naftali Weinberger, Simpson's Paradox, Stanford Encyclopedia of Philosophy, 2021 (revised 2026), https://plato.stanford.edu/entries/paradox-simpson/

3. Mike Grigsby, Marketing Analytics: A Practical Guide to Improving Consumer Insights Using Data Techniques, Kogan Page, 2022, ISBN 9781398608191, https://www.netgalley.co.uk/catalog/book/267990