Think, don't memorise

Because AI makes it trivially easy to measure the wrong things faster. The 5-level hierarchy (Commercial > Marketing > Memory > MarComms > Activity) exists to show that not all metrics are equal. Activity metrics (clicks, impressions, views) are easiest to generate and hardest to connect to business outcomes.

AI dashboards default to what's available, which is usually activity data. Without the hierarchy, a marketer stares at the bottom level and calls it performance. The skill is knowing which level you're looking at and whether it connects to the level above. A dashboard that shows you 50 metrics in real time is still decoration if none of them trace a causal chain to revenue.

Strong answers would reference the difference between reporting up (telling the board what happened) and optimising down (changing what you do next). AI is excellent at the first. The second requires judgment about which metric actually matters.

Every platform claims credit for the same sale. That's the attribution problem. The first question is: what's the attribution window? A 28-day click window captures people who would have bought anyway. The second question is: what's the counterfactual? What would have happened without the campaign?

Other questions worth asking:

• Is this incremental revenue or revenue that was already coming?
• Did the platform optimise towards people most likely to convert regardless?
• Are other platforms also claiming credit for the same conversions?
• What does the number look like if you change the attribution model?

The eBay paid search study is a good reference here. eBay ran a test and found that branded paid search (ads on their own name) delivered almost zero incremental value. The ROAS looked spectacular because people searching "eBay" were going to click through anyway. One study became a rule because nobody asked "under what conditions?"

This comes from Binet and Field's effectiveness research. Activation creates short spikes in sales (the sawtooth). Each campaign fires, sales jump, then drop back. Brand building raises the baseline over time (the staircase). The spikes still happen, but they start from a higher floor.

AI makes the sawtooth more tempting because AI is brilliant at activation. It can generate offers, retargeting sequences, email variants, personalised ads, all optimised in real time. The results are immediate, measurable, and satisfying. Short reporting windows reward short behaviour.

Brand building is harder to measure, takes longer to show results, and AI cannot easily generate the distinctiveness and memory structures it depends on. So the default pull is towards more activation, faster. If nobody protects the "later" work, the "now" work expands until it fills everything. You end up with a perfectly optimised sawtooth that starts lower every quarter because the brand baseline is eroding underneath.

AI and algorithms are built to optimise what's in front of them. The easiest audience to target is people who already know you: retargeting, lookalikes built from existing customers, personalised email to your database. AI makes it dangerously easy to talk to people who already know you while staying invisible to everyone else.

Sharp's point is that growth comes from mental availability (being thought of in buying situations) and physical availability (being easy to find and buy). Both require reaching light and non-buyers, not just your loyal base.

The trap is that AI-driven campaigns targeting existing customers produce better short-term metrics. The CTR is higher. The conversion rate is higher. The ROAS looks better. But you're mining a shrinking pool. The brand is not growing because nobody new is being reached. The numbers look good quarter by quarter while the business slowly narrows.

Distinctive assets are the consistent cues people recognise without needing to read the brand name: colours, shapes, characters, sounds, visual style. They're how memory works for brands.

AI can now produce creative variation like a slot machine. New images, new layouts, new copy, generated constantly and tested in real time. Without discipline, brand cues get diluted across all this variation. Every variant looks slightly different. None of them build the recognition that compounds over time.

The paradox: AI can actually multiply distinctive assets across formats without dilution, but only if you have them defined and enforced in the first place. A brand with strong assets can use AI to scale them consistently. A brand without them just gets more inconsistency, faster.

Consistency is how you stop memory working against you. The more noise AI creates in the market, the more valuable it is to be the signal people recognise instantly.

Share of model is a new concept alongside share of voice and share of search. It refers to how often your brand appears in AI-generated answers when people ask large language models questions about your category.

This matters because search behaviour is changing. In a zero-click world, people increasingly get answers without visiting a website. If someone asks an AI "what's the best running shoe for flat feet?" and your brand is not in the answer, you don't exist for that person. There's no page-two result to scroll to. There's no ad slot to buy. You're either in the model's answer or you're not.

AEO (answer engine optimisation) is different from traditional SEO. SEO optimises for ranking on a results page. AEO optimises for being included in a synthesised answer. The inputs are different: structured data, clear claims, authoritative sources, consistent information across the web. The stakes are binary in a way search rankings never were.

When supply is infinite, average becomes invisible. AI has collapsed the cost of producing competent content. That means there is more competent content than anyone can consume. The bar for "good enough to publish" is now below the bar for "good enough to notice."

What gets noticed is content that has something AI cannot easily generate on its own:

• A genuine point of view. Not a summary of existing thinking, but an argument that comes from experience, taste, or original observation.
• Real evidence. First-party data, original research, specific case details that aren't in the training data.
• Distinctive voice. Not polished-generic. Recognisable. Something that could only come from this brand or this person.
• Emotional resonance. System1 research shows that emotional response drives creative effectiveness. AI defaults to rational, informational content. The gap is in work that makes people feel something.

The algorithm will see you now, but only if you give it something worth showing. Algorithms gate visibility, and they reward engagement. Engagement comes from distinction, not volume.

Curiosity as a system means deliberately building habits and routines that force you to encounter things you wouldn't naturally seek out. It's not about being "a curious person." It's about treating curiosity as infrastructure.

Koen Pauwels said "marketing is way harder than rocket science" because humans don't follow formulas. The only way to stay calibrated is to keep pressing on odd things, following threads nobody asked you to follow, reading outside your category.

AI undermines this because algorithms feed you more of what you already like. AI gives confident first answers that feel like final answers. The temptation is to stop at the first plausible response. If curiosity is just a trait, it atrophies. If it's a system, it runs even when you're not in the mood.

The eBay paid search example applies here too: one study became a rule because nobody asked "under what conditions?" Without curiosity as a system, you generalise from one case and call it strategy.

Skeptical optimism is the habit of asking where something came from before you build a plan on top of it. It is not the same as cynicism. A cynic dismisses. A skeptical optimist investigates.

Terence Tao, the mathematician, described a "sense of smell" for when something is wrong even if each step looks fine. That's the skill. AI output often looks polished and professional. Every paragraph reads well. Every recommendation sounds plausible. But the whole thing might be built on a weak foundation that nobody questioned because the surface looked finished.

"Fine is how nonsense gets shipped now, with a straight face and a nice font." AI produces confident nonsense at scale. If you can't smell the weak bit, you build a month of work on top of it. Skeptical optimism means you believe the tool can help, and you check its work before you build on it.

This is the "doing the reps" behaviour. Martin's point, reinforced by the A.G. Lafley example and Gary Klein's research on firefighters, is that judgment comes from accumulated experience, not natural talent. Pattern recognition is built from thousands of reps.

AI disrupts this because it hands you a result that feels like you earned it. You get the output without doing the work that builds the judgment to evaluate it. DHH described this as "competence draining out of your fingers." You get faster at finishing and worse at thinking.

When production cost is near zero, judgment becomes the bottleneck. You can only judge quality if you've done enough reps to know what weak looks like when it's dressed up nicely. A marketer who has written 200 briefs can smell a bad one instantly. A marketer who has approved 200 AI-generated briefs may not be able to write one from scratch.

The implication: use AI to accelerate, not to replace the doing. Write your own rough version first. Then let AI improve it. The reps are not optional.

Brian Chesky's line: "You can't delegate understanding." Elena Verna at Lovable made the same argument: a prototype beats a spec because building reveals what meetings hide.

Before AI, building was genuinely hard. You needed a designer, a developer, a copywriter. The only thing a marketer could make alone was a slide deck. So marketing became a profession of describing: briefs, strategies, frameworks, presentations about work rather than the work itself.

AI has changed this. A marketer can now research a market, write copy, generate design, build a landing page, test messaging, all in an afternoon. The tools have made building available to marketers for the first time. A page you can click is worth twenty slides about a page.

The danger is that it's tempting to let AI build everything and just approve. You become the person who reviews descriptions of work rather than doing work. The behaviour is about maintaining the habit of making things yourself, even when AI could do it, because the act of building is where you learn what actually works.

The old generalist was a coordinator. They ran the meeting, sent the notes, pointed at the Gantt chart. They knew a little about everything but couldn't actually make anything. They depended on specialists to do the real work.

The multi-tool marketer can actually make things. Not as well as the specialist in any single domain, but well enough to move. "Not the best at any one blade. Just the person who can open the bottle, tighten the screw, cut the cord, and keep going."

This isn't about using lots of AI tools. It's about having broad enough skills to ship without depending on a chain of other people. Research, create, test, ship. Two people with taste and AI tools can do what five specialists used to.

The risk is breadth without depth: doing five things badly while feeling productive. The multi-tool marketer still needs specialists for deep work. They just don't need to stop the whole project because they can't pull a list or tag a link.

This is an open question, but here's one strong pairing: doing the reps + skeptical optimism.

Doing the reps builds the pattern library that makes skeptical optimism possible. You can only smell when something is off if you've seen enough good and bad work to have a baseline. A marketer who has written hundreds of briefs develops Tao's "sense of smell" for when a brief is structurally weak, even if every sentence reads well.

Going the other direction: skeptical optimism makes the reps more valuable. If you just do the reps without questioning the output, you practise reinforcing whatever AI gives you. Skeptical optimism means each rep includes a moment of evaluation, which deepens the learning.

Other strong pairings: curiosity + building (curiosity gives you questions, building gives you answers you can test). Building + multi-tool (every real thing you build adds a skill to your range). The point is that these behaviours compound. Neglect one and the others weaken.

Pattern matching is repeating what sounds right. "We should invest in brand building because Binet and Field say so." It's correct, but it's a fact recited, not an argument understood.

Mechanistic reasoning is explaining why something works or doesn't. "If we only invest in activation, we get short spikes that decay. Each spike starts from the same baseline. Brand building raises that baseline by creating memory structures, so future activation campaigns start from a higher floor and convert more efficiently."

The difference: pattern matching gives you the answer. Mechanistic reasoning gives you the causal chain. The exam cares because AI is the ultimate pattern matcher. It can produce pattern-matched answers that sound sophisticated. The thing AI cannot easily do is trace novel causal chains in specific contexts. That's judgment. That's what the exam tests.

If your answer could have been generated by someone who read the question carefully but never attended class, it's pattern matching. If it shows you understand why something works, not just that it works, it's mechanistic reasoning.

The Cognitive Corridor is Nosta's idea that AI gives you temporary illumination, not understanding. Like walking through a dark corridor with a torch, you can see where you are right now, but you haven't mapped the building.

For marketers, this means AI can give you an answer, a strategy, a framework, instantly. But that answer exists only in the moment you're looking at it. Without the work of understanding it, you can't build on it, adapt it, or know when it stops being true.

Related: the generation effect from psychology. You remember more when you make the thing yourself. The risotto example: following a recipe produces a result but not judgment. Judgment comes from scraping burned rice off the pan. If AI writes your strategy, you have a strategy. If you write it and AI helps you stress-test it, you have understanding.

The practical takeaway: use AI as a tutor, not a generator. "Don't give me the answer. Ask me questions until I can explain it in my own words."

Several things. First, AI answered the strategic questions without anyone noticing. It picked a segment, invented a desire, wrote a proposition. Those are judgment calls that should involve real customer understanding, competitive analysis, and trade-offs. AI defaults to the safe version, the middling version that offends no one and is remembered by no one. Real strategy requires sacrifice. If the positioning could apply to three competitors, it's not positioning.

Second, polished output doesn't get questioned. The team approved in 10 minutes because it looked finished. Nobody asked: is this based on real customer insight? Does this require us to say no to anything? Could a competitor say the same thing? The meeting room where everyone nods at polished AI output is the room where bad strategy gets shipped with confidence.

Third, the friction was skipped. The messy process of debating positioning, arguing about trade-offs, testing language against real customers, that's where understanding is built. Without it, you have words on a slide that nobody deeply owns.

Standards stop speed turning into noise. Without standards, AI fills the vacuum with plausible mediocrity.

The core shift: use AI to interrogate your thinking, not to replace it. Practical approaches from the course:

• Tutor mode: "Don't give me the answer. Ask me questions until I can explain it in my own words." Forces you to do the thinking. AI guides but doesn't hand over.
• Write first, then critique: Write your own rough version of the brief, strategy, or copy. Then give it to AI and say "argue against this. Where is it weakest?" You keep the reps. AI sharpens them.
• Red team: Ask AI to argue the opposite case. If your strategy is "invest more in brand," ask AI to make the strongest possible case for activation-only. If it convinces you, your strategy was weak.
• Failure search: Ask: "What would have to be true for this to fail?" Most marketers look for evidence that supports their plan. The valuable exercise is looking for evidence that breaks it.
• Multiple models: Try more than one AI. If three models give the same answer, it might be a genuine consensus. If they diverge, the question is more interesting than you thought.

The principle: if AI gives you an answer and your response is "great, done," you've used it as a shortcut. If your response is "interesting, but let me push on this bit," you've used it as a sparring partner.

AI removes friction from almost everything. Research that took a week takes 10 minutes. Copy that took a day takes seconds. The instinct is to celebrate this. Faster is better.

But the friction was doing something. The struggle of writing a brief from scratch is where you discover what you actually think. The frustration of analysing data by hand is where you notice the pattern that doesn't fit. The risotto example: following a recipe produces risotto, but judgment comes from scraping burned rice off the pan.

The generation effect from psychology backs this up: you remember and understand more when you generate something yourself rather than reading or receiving it. The effort is not a tax on the process. It is part of the product.

This does not mean refuse to use AI. It means be deliberate about where you keep the friction. Use AI to handle the mechanical parts (formatting, sourcing, generating variants) and keep the judgment parts for yourself (defining the problem, evaluating the output, making the trade-off). The friction is the point, but only the right friction. Spending three hours formatting a spreadsheet teaches you nothing. Spending three hours debating what should be on the spreadsheet teaches you everything.

Because "checking AI's work" usually means looking for factual errors, and that's the shallow version of the problem. AI is not primarily dangerous because it gets facts wrong. It's dangerous because it produces work that is plausible but strategically empty.

A positioning statement can have zero factual errors and still be worthless because it doesn't require the company to say no to anything. A measurement dashboard can have correct numbers and still mislead because it shows activity metrics instead of commercial outcomes. A campaign brief can be well-structured and grammatically perfect and still be built on an assumption nobody tested.

Checking for errors is necessary but insufficient. The real skill is checking for judgment: is this the right question? Is this strategy actually distinctive? Does this measurement framework connect to business outcomes? Would this work survive contact with a skeptical CFO?

To check that, you need the knowledge and experience to know what good looks like. Which brings us back to the reps, the fundamentals, and the behaviours. There's no shortcut to judgment. "Always check" is a policy. Knowing what to check for is a skill built over time.

Execution was genuinely hard. You needed years of training to design professionally, to code, to write well, to analyse data rigorously. The tools demanded specialisation. A print designer needed deep knowledge of typography and prepress. A developer needed programming languages. An analyst needed statistics and SQL.

The division of labour was rational because each step required hard-won skill. You couldn't ask the copywriter to also build the landing page. The tools wouldn't let them. So companies built chains: strategist develops thinking, copywriter writes, designer designs, developer builds, social team schedules. Three weeks later, something launched.

This created a structure where most of a marketer's time was spent on alignment, updates, approvals, and handoffs, not on making things. That was an acceptable cost when the alternative was asking unqualified people to do specialist work. The cost of coordination was lower than the cost of bad execution.

Brian Chesky's point, echoed by Elena Verna: distance is a liability. If your day is mostly alignment, updates, approvals, and handoffs, you're living off second-hand information. The headline that sounds fine in a brief might feel like a lie on the actual page. The targeting that looks right in a spreadsheet might miss something obvious when you look at the ad in context.

When the person making decisions is far from the work, decisions are based on summaries of summaries. Each layer of abstraction loses signal. The marketer who briefed the agency, who briefed the creative, who briefed the designer, is three steps removed from the customer experience.

AI closes this gap. A marketer with AI tools can go from insight to action without the chain of dependencies. Not because they're better than the specialist, but because the proximity to the work means they see things that get lost in handoffs. Better feedback loops because the person building is close enough to see what's actually happening.

This doesn't mean no teams. It means fewer, smaller teams where people are closer to the work. The six-person social team becomes two people with taste, AI tools, and permission to make decisions.

The risks are real and worth taking seriously:

• Breadth without depth. Doing five things badly while feeling productive. AI can make you feel competent in domains where you're actually just fast. The output looks professional, but a specialist would spot the gaps.
• Competence drain from specialists. If specialists stop practising because the multi-tool marketer handles "good enough" versions of their work, the organisation loses deep expertise it still needs for the hard stuff.
• Judgment gaps at scale. One person making all the decisions means one person's blind spots affect everything. A team of specialists at least provided multiple perspectives.
• Burnout from scope. "Can do everything" often becomes "expected to do everything." The multi-tool marketer becomes a one-person department and burns out.

It breaks when the work requires genuine depth: a complex data architecture, a brand identity system, a technical SEO migration. The multi-tool marketer still needs specialists for deep work. They just don't need to stop the whole project because they can't pull a list or tag a link.

The 9-floor building idea from the course: a marketer who only knows channels is half a marketer. You need to connect what you do to what the business cares about, which is revenue, margin, and growth.

Commercial fluency means understanding CAC (customer acquisition cost), LTV (lifetime value), payback period, contribution margin, and how to read a P&L. Not at an MBA level. At a "can defend a budget in language a CFO believes" level.

This matters more now because AI makes it easy to produce impressive-looking marketing work. But if you can't connect that work to a business case, you're asking for budget based on activity metrics. A CFO does not care about your click-through rate. They care about whether the next pound of marketing spend returns more than a pound of profit, and on what timeline.

The multi-tool marketer who can also speak finance is significantly more valuable than one who can't. They can make the case for brand investment in language the board understands, rather than relying on "trust me, brand matters" which has never been a strong argument in a budget meeting.

Efficiency measured how? This is a measurement question disguised as a team structure question. The questions to ask:

• What did you automate? If it's mostly activation (email sequences, ad variants, social scheduling), then the 80% figure means 80% of the visible work, which was probably not 80% of the value.
• What's happening to brand metrics? If brand awareness, consideration, or distinctiveness are declining while efficiency metrics improve, the company is eating the staircase to fund the sawtooth.
• Who is doing the judgment work? Cutting the team in half is fine if the work that was cut was mechanical production. It's dangerous if you also cut the people who asked "should we do this?" and "does this actually work?"
• Has reach expanded or contracted? AI-driven efficiency often means doing more of the same thing to the same people. Check whether the company is still reaching new audiences or just retargeting existing ones more efficiently.
• What's the trend line? Efficiency gains in quarter one can mask decline by quarter four. Brand erosion doesn't show up immediately. It shows up as gradually declining campaign performance, higher acquisition costs, and lower conversion rates from cold audiences.

The headline sounds like progress. The questions reveal whether it actually is.

This was the thesis of the whole course. AI has collapsed the gap between strategy and execution. When production costs approach zero, strategic judgment becomes the differentiator. So the answer is not about tools. It's about what you bring to the tools.

The behaviours matter more, not less. Curiosity as a system keeps you learning when AI tempts you to stop. Skeptical optimism keeps you questioning when output looks finished. Doing the reps builds judgment that AI can't hand you. Building real things connects you to the work. Becoming multi-tool lets you ship without waiting.

The knowledge matters more, not less. Binet and Field's balance. Sharp's reach imperative. Distinctive assets. Measurement literacy. Commercial fluency. These are the inputs that make AI useful. Without them, AI just makes you faster at the wrong things.

The tools are the accelerant, not the strategy. A marketer with strong behaviours, solid knowledge, and AI tools is formidable. A marketer with just AI tools is fast but directionless.

The three pillars feed each other. Develop any one and it pulls the other two forward. Neglect any one and the other two weaken. That's the system. That's the answer. Now go build something.