The Cheap Always Floods: Jevons’ Paradox, Coca-Cola’s AI Christmas, and Why Being Remembered Is the Last Scarce Asset in Marketing
In 1865, a young economist warned Britain that making steam engines more efficient would not save coal. It would burn more of it.
19 Oct 2026
The economist William Stanley Jevons, said efficiency would burn more coal in 1865, but how does this relate to modern marketing?
In 1865, William Stanley Jevons, then twenty-nine, published a book with an unexciting title and an unsettling argument. The Coal Question asked how long Britain's industrial supremacy could last on a fuel that was finite and getting harder to dig.[1]
The comfortable answer of the day was engineering. Engines were getting better. Each generation of steam engine, from Newcomen's to Watt's and beyond, needed less coal to do the same work. Surely, the reasoning went, efficiency would stretch the supply.
Jevons said the opposite. Every improvement that made coal more productive made coal more worth using. Cheaper steam power did not mean the same mills burning less; it meant more mills, more pumps, more engines, more locomotives, in more places, doing work that had never been worth mechanising before. Efficiency did not conserve the resource. It unlocked demand for it.[2]
He was right. Britain consumed roughly 93 million tons of coal in the year his book appeared.[3] Output kept climbing for almost half a century, peaking in 1913 at around 292 million tonnes, roughly three times the level that had frightened him.[4] The engines had never been more efficient. The country had never burned more coal.
"Efficiency did not conserve the resource. It unlocked demand for it."
The principle now carries his name, and it has rarely been more relevant than in the budget meetings marketing directors are sitting in this autumn.
Why cheaper efficient content means more content, not less spend
Jevons' paradox does not hold everywhere. Economists are precise about when it bites. Two conditions have to be met.[2] First, the efficiency gain must lower the effective price of what the resource produces. Second, demand for that output must be elastic: when it gets cheaper, people must want a great deal more of it. Where demand is fixed, efficiency really does save. Farming is the classic counterexample; better yields made food cheaper, but nobody ate three times as much bread.
Content production meets both conditions more completely than coal ever did.
The price drop is close to total. A blog post, a product description, a social carousel, a campaign variant, a thirty-second video: generative AI has moved the marginal cost of each towards zero. And demand for content inside organisations is about as elastic as demand gets. There is always another channel, another segment, another A/B variant, another post the algorithm supposedly rewards. Give a marketing team ten times the output capacity and almost none of them will produce the same volume for a tenth of the cost. They will produce ten times the volume.
Multiply that across every competitor at once and the result is the content equivalent of 1913: more engines than ever, running harder than ever.
But there is a difference between coal and content that Jevons never had to model. Coal was consumed by machines. Content is consumed by people, and people bring a fixed resource to the transaction. Herbert Simon put it most clearly in 1971: a wealth of information creates a poverty of attention.[5] The supply of content can grow without limit. The hours a buyer spends looking cannot.
"The supply of content can grow without limit. The hours a buyer spends looking cannot." - Herbert Simon
That asymmetry decides who wins. When one input becomes free and the other stays fixed, value does not disappear. It migrates, entirely, to whatever controls access to the fixed one. In content markets, that is attention. And the asset that commands attention without having to buy it fresh each time has a name.
The flood is already measurable
This is not a forecast. The first half of Jevons' prediction has already played out in public data.
In late 2025, the SEO firm Graphite analysed 65,000 English-language articles drawn from Common Crawl, published between January 2020 and May 2025, and ran them through an AI detector.[6] In 2020, around 5% were classified as machine-written, a figure close to the detector's own error rate. Within roughly a year of ChatGPT's public release, the share was approaching 40%. In November 2024, AI-written articles briefly outnumbered human-written ones. By May 2025 the two were roughly level, at about 48% machine to 52% human, and the curve had flattened.[6][7]
In under three years, roughly half of everything newly published in written form on the open web became machine-made. That is the volume half of the paradox, arriving faster than coal ever did.
The second finding is the one marketing directors should pin above their desks. The same research found that most of those AI-written articles do not appear in Google's search results or in ChatGPT's answers. Both systems still overwhelmingly surface human-made work.[7]
"Half the new content on the web is now machine-made. Most of it is never seen."
So the volume exploded and the visibility did not follow. Production multiplied; the attention available to it did not move. Each additional unit of cheap content is competing for the same fixed seam, and most of it is mined at a loss. This is the flood. It is not a threat that might arrive. It is the market your next campaign launches into.
The saving your CFO has already read about
The other half of the story is arriving in boardrooms as a number.
In May 2024, Klarna told investors that generative AI had helped cut its sales and marketing spend by 11% in the first quarter. Spending on external marketing agencies fell by 25%. The company attributed roughly $10m of annualised savings to AI, and said it was running more campaigns, not fewer, with in-house tools producing imagery and copy that agencies and photographers had previously supplied.[8]
To be clear about what this publication is and is not saying: Klarna's numbers are a genuine efficiency, credibly reported, and most marketing teams will find savings of a similar shape. The point is not that the saving is false. It is that the saving is the part of Jevons' story that shows up on a spreadsheet, and the rest does not.
Read the announcement again with Jevons in mind. A lower cost per asset, used to produce more assets. That is the paradox operating at the level of a single firm, in real time, reported as a success. Repeat it across every competitor reading the same headlines and the market-level result is the flood described above.
For marketing directors, the consequence is political before it is strategic. Once a saving like Klarna's is in the public record, it becomes the benchmark. Finance teams begin to ask the obvious question: if content now costs a fraction of what it did, why does the marketing budget not? It is a reasonable question. The honest answer, which Jevons supplied a century and a half ago, is that the cost of producing content was never the constraint that mattered. The constraint was always being noticed, and remembered, by people with a fixed amount of attention to give.
"The saving is the part of Jevons' story that shows up on a spreadsheet. The rest does not."
The marketing director who walks into that conversation with only a production budget to defend will lose it. The one who can show where the value has moved will not. The clearest evidence of where it has moved came, of all places, from a Christmas advert.
Case study: Coca-Cola's AI Christmas and the score that did not move
In 1995, Coca-Cola released "Holidays Are Coming": a convoy of illuminated red trucks rolling through a snowy landscape, lights coming on as they pass, set to a jingle that many viewers now associate with the start of Christmas itself.
In November 2024, the company remade it using generative AI.[9] The reaction from the creative industry was immediate and hostile. Critics said the machines had drained the warmth from a holiday classic and treated it as a threat to everyone who makes a living in advertising. Comments on the company's YouTube post ran so negative that Coca-Cola restricted the page.[10]
Then the ad was tested on people who do not work in advertising.
The research firm System1 ran the full spot past ordinary viewers, without telling them how it had been made. It scored 5.9 out of a possible 6, a rating System1 describes as exceptional. The most common emotional responses were excitement and feeling uplifted. Only 3% reported contempt. And the score was identical to the one System1 had given the original 1995 film.[10]
Figure 1 — What changed in Coca-Cola's Christmas ad, and what did not
1995 original | 2024 AI remake | |
|---|---|---|
How it was made | Conventional film production | Generative AI |
Typical production lead time | Lead times of up to a year, per Coca-Cola's CMO[11] | Around a month |
Industry reaction | Became a seasonal institution | Widespread criticism from creatives[9] |
System1 consumer score | 5.9 / 6 | 5.9 / 6[10] |
Distinctive assets on screen | Red trucks, lights, snow, jingle | Red trucks, lights, snow, jingle |
Coca-Cola drew its own conclusion. In November 2025 it released a second AI-made Christmas film, again a homage to the 1995 trucks. The industry criticism returned, this time aimed at visible flaws such as trucks whose size and wheel count shift between shots. A Coca-Cola global vice president said most of the complaints came from inside the advertising business, and System1 again assessed the work as strong for both short- and long-term market share.[11]
It is worth being precise about what this case does and does not prove. It does not prove that AI makes good advertising. Both films were near-copies of a thirty-year-old original, and System1 itself noted that nostalgia may have done much of the work.[10]
That caveat is the finding.
"Coca-Cola changed everything about how the ad was made and nothing about what it scored. The value was never in the production."
The score did not hold because AI is good. It held because the assets were. The trucks, the lights and the music had been built into the memories of millions of people over three decades. AI did not create that value. It rented it, cheaply, from an asset Coca-Cola already owned. The production method was swapped out entirely, and the effect did not move, which tells you exactly where the effect was stored.
Efficiency multiplies what you own and divides what you do not
Put the three pieces of evidence side by side and Jevons' logic completes itself. Production became cheap (Klarna). Volume exploded, and visibility did not follow (Graphite). And the one test case where AI-made work performed as well as anything human-made did so because it was running on thirty years of stored memory (Coca-Cola).
The marketing science has a precise name for that stored memory. Byron Sharp's work at the Ehrenberg-Bass Institute frames brand growth around mental availability: the likelihood that a brand comes to mind, quickly and easily, in the moments when people are ready to buy.[12] Jenni Romaniuk's research on distinctive brand assets explains how that availability is built and triggered: through colours, shapes, characters, sounds and symbols that people reliably link to one brand and no other, built through consistent use over years.[13]
Red trucks. A string of lights. A jingle. Those are distinctive assets in the textbook sense. They are why a remake made in a month could score like a film made in 1995.
This is where Jevons' paradox stops being a story about marketing costs and becomes a story about enterprise value. When the cost of production collapses, it collapses for everyone at once. The ability to make content stops being a source of advantage, because every competitor has the same tools at the same price. What remains scarce is the thing the tools cannot manufacture on demand: memory structures that already exist in buyers' heads.
So the same efficiency gain does opposite things to different businesses.
Figure 2 — One efficiency gain, two different outcomes
Business with distinctive assets | Business without them | |
|---|---|---|
What cheap production does | Puts its assets in front of buyers more often, at lower cost | Adds more anonymous content to the flood |
Effect per piece of content | Holds, because each piece triggers existing memory | Falls, because each piece starts from zero |
Where the saving goes | Into reach and consistency of the assets | Into more volume, competing for the same attention |
What compounds | Memory, which is to say brand equity | Output, which depreciates on publication |
In the coal economy, cheaper engines raised the value of whoever owned the seams. In the content economy, cheaper production raises the value of whoever owns the memories. AI is an extraordinary engine. Its fuel is brand.
"AI is an extraordinary engine. But it runs on memory, and memory is the one thing it cannot dig for you."
This publication argued in July that the measure always wins: the most measurable function in a business tends to end up setting the strategy.[14] The AI content boom is that problem with the throttle open. Output is easy to count; a dashboard will show a team producing five times the assets for half the cost. Memory is hard to count, and no dashboard will show it being diluted by those same assets if they carry nothing distinctive. A marketing function optimising the visible number will produce the flood with great efficiency, and report it as success.
Where the argument is weaker
An argument this tidy deserves resistance, and there are four honest objections.
The paradox needs elastic demand, and a firm can choose not to supply it. Jevons describes what markets do, not what any single business must do. A marketing director who holds output flat and banks the saving escapes the paradox at firm level. Klarna's own figures show a real reduction in spend.[8] The flood is a market outcome; participation in it is a choice.
Coca-Cola is the extreme case, not the typical one. Few brands own assets with three decades of seasonal repetition behind them, and both AI films were close copies of an original. Nostalgia flatters the test. A challenger brand will not get a 5.9 by remaking anything.
The visibility data may not last. Graphite's finding that AI-written content rarely surfaces in search and AI answers describes the systems as they stood in 2025, and its detector carries its own error rate.[7] Search and recommendation systems change. A world in which machine-made content is surfaced as readily as human work would weaken the visibility half of this argument, though it would deepen the flood.
AI can help build distinctive assets, not only dilute them. Consistency is the hardest discipline in brand building, and much of it is execution: the same colours, characters and cues, applied correctly across thousands of touchpoints. Tools that make consistent execution cheap could help smaller businesses build memory structures faster than was ever affordable before.
The last objection is the most important, because it is not really an objection. It is the strategy. The paradox does not say AI is bad for brands. It says that the gains flow to whoever points the new capacity at memory rather than at volume.
What a marketing director should do with the saving
Every marketing leader will spend the next two budget cycles answering some version of the same question: AI made content cheaper, so where is the money? Jevons gives a better answer than defending the old budget or surrendering it. Four moves follow from the argument.
- Decide where the saving goes before finance decides for you. An efficiency left unallocated defaults to one of two places: back to the centre as a cut, or into more volume. Both are Jevons outcomes. The third option, reinvesting it in putting your distinctive assets in front of more buyers more consistently, is the only one that compounds, and it has to be argued for explicitly.
- Know what you own before you scale anything. Which colours, shapes, characters, sounds or phrases do buyers already link to you, and to no one else? Most businesses have never tested this. Romaniuk's research gives a method for doing so.[13] Scaling production before answering it is building engines with no idea where the seam is.
- Make the distinctive asset a condition of shipping. If cheap production is going to multiply your output, every piece of it should carry something that triggers your brand in memory. A piece of AI content that could have come from any competitor is not neutral. It is a small contribution to the flood your next campaign will have to cut through.
- Report memory alongside output. The efficiency number will be the one finance sees first. Put a memory number beside it: recognition, brand linkage, share of mind in your category. If you only report what is easy to count, you will be managed on it.
For finance directors and chief executives reading over the marketing director's shoulder, the argument reduces to a single line. The cost of producing marketing has fallen. The cost of being remembered has not, and in a flooded market it is rising.
"Content has never been cheaper to make. Being remembered has never cost more."
Connor Whiting
Jevons was writing about engines, and the locomotives and mills they drove. His point was that the efficiency everyone celebrated did not save the thing they thought it would save. Its corollary is the one that matters now: when the engine gets cheap, value moves to whoever controls the fuel. A hundred and sixty years later, every business has been handed the same extraordinary engine at the same moment, at almost no cost.
The engine is no longer the advantage. Everyone has the engine now.
The only question left is whether you own anything worth burning in it.
Written by
Notes and references
Original framing. The application of Jevons' paradox to AI-produced marketing content, the reading of the Coca-Cola System1 result as evidence that effect is stored in distinctive assets rather than production, and Figure 2 are this publication's own analysis. Jevons' paradox has been applied to AI before, notably by Microsoft's chief executive, but usually to argue that AI will increase total demand for computing or labour, not to locate where marketing value migrates.[2] Coal figures for 1865 and 1913 come from different sources and units (tons and tonnes); "roughly three times" reflects that. Graphite's AI-detection figures carry a stated error rate and should be read as estimates. This piece deliberately leaves aside the team-capability questions covered in our May piece, Augment, Don't Replace: The AI Integration Test.
- William Stanley Jevons, The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines, Macmillan, 1865. Overview: The Coal Question. Accessed October 2026.
- Jevons paradox, overview of the conditions for the paradox (lower effective price, elastic demand, rebound above 100%) and modern applications including AI. Accessed October 2026.
- Resilience, Jevons' coal question: why the UK coal peak wasn't as bad as expected, August 2011. UK consumption of about 93 million tons in 1865. Accessed October 2026.
- Our World in Data, The death of UK coal in five charts. Production peak of 292 million tonnes in 1913. Accessed October 2026.
- Herbert A. Simon, "Designing Organizations for an Information-Rich World", in M. Greenberger (ed.), Computers, Communications, and the Public Interest, Johns Hopkins Press, 1971.
- Axios, AI writing hasn't overwhelmed the web yet, 14 October 2025, reporting Graphite's analysis of 65,000 Common Crawl URLs. Accessed October 2026.
- TechRadar, The internet is now mostly written by machines, study finds, October 2025. Includes Graphite's finding on search and AI-answer visibility and its stated detector error rates. Accessed October 2026.
- Payments Dive, Klarna uses gen AI to cut marketing spend, 29 May 2024; see also PetaPixel, Klarna CEO says he is spending less on photographers thanks to AI, June 2024. Accessed October 2026.
- Adweek, How Coca-Cola's AI holiday ad went from praise to rage, 20 November 2024. Accessed October 2026.
- MediaPost, Not everyone hated that AI-generated Coca-Cola holiday ad, November 2024. System1 score of 5.9/6, matching the 1995 original; 3% contempt. Accessed October 2026.
- Kyunghyang Shinmun, Coca-Cola's second AI Christmas ad, November 2025. Production lead times and System1 assessment of the 2024 and 2025 films. Accessed October 2026.
- Byron Sharp, How Brands Grow: What Marketers Don't Know, Oxford University Press, 2010.
- Jenni Romaniuk, Building Distinctive Brand Assets, Oxford University Press, 2018.
- The Brand Ledger, The Measure Always Wins: Goodhart's Law, Nike's $28 Billion Day, and How the Dashboard Took Over Business Strategy, July 2026.