Diagnosis

A strategy is a diagnosis before it’s a plan. Most bad strategies skip straight to a list of things to do. A good one first does the harder work: an honest, evidenced read of what’s actually going on, and only then decides where to push.

You’ve named the north star and its inputs. Diagnosis is the step where you figure out why the number is where it is, and where the real leverage sits, before anyone argues about what to build. Done well, it’s the difference between a plan people can poke holes in and one they can trust. This page is the method; the Storefront example at the bottom works it end to end.

Lead with the answer#

Write the diagnosis the way a consultant presents it: the answer on top, the support underneath. Open with a single governing thought: one sentence that states where the leverage is and why. Everything after it exists to defend that sentence.

Governing thought (Storefront): We can add ~350 orders/week (+18%) this half with zero new traffic, by closing two quantified leaks in a funnel we already own. Checkout friction alone is more than half the prize, and it’s entirely in our control.

If a reader has to reach the last paragraph to learn the point, the diagnosis is written backwards. The governing thought is what a busy exec should be able to repeat back after reading one line, and it’s the thing every Big Rock downstream will trace to.

A governing thought earns its place only if it’s specific (names the leverage, not a vibe), evidenced (a number stands behind it), and falsifiable (you could be proven wrong). “We need to improve the customer experience” fails all three.

Decompose the outcome (MECE)#

You can’t find the leak by staring at the outcome. Break the north star into the chain of steps that produce it, so that the steps are mutually exclusive (no overlap) and collectively exhaustive (they add up to the whole). That’s a value tree: the outcome at the top, its drivers beneath, each driver a number you can inspect.

For a funnel business the tree is the funnel itself (sessions → carts → checkouts → completed orders) and each step has a conversion rate you can measure. For a subscription business it might be new × retained × expanded. The shape depends on how value is actually created; the discipline is that the branches multiply back up to the outcome, with nothing missing and nothing double-counted.

Decomposing does two things at once: it shows you where the number is lost, and it forces the vague problem (“orders are flat”) into a specific one (“we lose 20% of everyone who reaches checkout”). You can’t act on the first. You can act on the second.

Size the prize#

A leak you haven’t sized is a hunch. For every step where value is lost, put a number on how much of the outcome you’d recover by closing it, in the currency of the north star, not in percentages floating free of it.

This is what turns a list of problems into a ranked one. “Checkout completion is 80%” is a fact. “Lifting checkout completion from 80% to 88% is worth +190 orders/week” is a decision input: now it can be compared against every other lever on the same axis. Sizing is also the honesty check on your own governing thought: if the leak you’re excited about is worth 30 orders a week and a boring one is worth 200, the diagnosis just told you that you were about to work on the wrong thing.

Size roughly and transparently: an order-of-magnitude estimate with its assumptions visible beats a precise number nobody can question. The goal is to rank the leverage, not to forecast to the decimal.

Weigh the options on explicit criteria#

Diagnosis narrows the field, but there’s almost always more than one way to attack the gap. Lay the real options side by side and score them on criteria you name before you look at the answer; otherwise the option you already liked wins on criteria invented to justify it. Useful criteria are usually some mix of:

Prize: how much of the sized opportunity it captures.
Control: how much of the outcome is in your hands vs. a vendor’s, the market’s, or luck’s. Leverage you own compounds; leverage you rent doesn’t.
Time to value: how fast you’d know it’s working.
Cost / risk: capacity spent, and what breaks if it goes wrong.

The output is a small table, not a single blended score: forcing a rock and a bug-fix onto one number produces false precision. The table’s job is to make the choice legible: this is why we’re doing X, and just as importantly why we’re not doing Y. A diagnosis that funds every option has diagnosed nothing.

Build targets from evidence, not round numbers#

By the end of diagnosis you’ll set targets for each input (Strategic Planning turns them into rock-level goals). A target picked because it’s round (“let’s get to 90%”) has no theory behind it and no credibility. Build each one from something real:

A prior result. “The saved-address A/B lifted completion 3 points before a bug pulled it” → 88% is that lift, compounded, held conservatively.
A benchmark. “Best-in-class one-tap checkout runs 90%+” → 88% sits just under the frontier, not past it.
A driver model. “Our P75 pages already load in 1.5s once lazy-loaded” → 1.5s is the median if we ship what the fast pages already do.

If you can’t say where a target came from, you guessed, and everyone downstream inherits the guess.

The Storefront diagnosis#

Here is the whole method on the running online-store example. Start from the north star (completed orders/week ≈ 1,900) and decompose it.

One week, four numbers. Sessions are healthy and still growing, so the problem is not traffic: the value disappears at the two steps we fully control, marked in blue. Slow product pages cost roughly 160 orders a week and re-typing address details at payment costs roughly 190, which is about +350 orders a week, or +18%, with no additional traffic at all.

The tree makes the diagnosis undeniable: sessions are healthy and still growing, so the problem isn’t traffic; it’s that we lose value at two steps we fully control. Sizing ranks them: the checkout leak (leak ②) is the single biggest recoverable chunk, page speed (leak ①) is second, and both beat anything at the top of the funnel.

A third lever sits beside the funnel rather than inside it: add-on attach (18% → 30%) doesn’t create orders, it deepens the ones we already win, worth roughly +$1.9k/week in margin. It’s in the diagnosis because it’s cheap and adjacent, not because it moves the north star.

Now weigh the options: the honest field of ways to grow orders, scored before picking:

Option	Prize / wk	Control	Time to value	Verdict
Buy more traffic	high, but CAC rising	Low (paid, competitive)	Slow	✗ pouring water into a leaky funnel
Fix conversion (checkout + pages)	+350 orders	High, our own code	Fast	✓ cheap, ours, compounds on every future order
Loyalty / retention program	?	Medium	Slow	✗ the pilot moved nothing: no new hypothesis
New region / new SKUs	?	Low	Slow	✗ defer: would eat the whole expansion budget

Targets, built from evidence: completion 80% → 88% (the saved-address A/B lift compounded, held below best-in-class ~90%+) · page load 3.5s → 1.5s (the P75 of pages already lazy-loaded) · attach 18% → 30% (one-tap removes the re-entry step that suppresses today’s rate).

That’s a diagnosis: a governing thought on top, a value tree that proves it, a sized and ranked prize, an options table that makes the choice legible, and targets you can trace to evidence. What comes next (turning it into bets and a committed plan) inherits real numbers instead of asserted ones.

Where diagnoses go wrong#

Asserted, not decomposed. “Checkout is our problem” with no funnel behind it. If there’s no tree and no number, it’s a hunch wearing a suit.
Boiling the ocean. A list of twelve problems, none sized. Sizing is what forces the list down to the two or three that matter.
Round-number targets. 90% because it’s tidy. A target with no source is a guess everyone downstream inherits.
Options theatre. One real option and three strawmen built to lose. If the alternatives were never live, the table is decoration, not a decision.
The unfalsifiable governing thought. “Improve the experience.” You can’t be wrong, so you can’t be right, and nothing downstream can trace to it.