ArticleIssue #189 | What The LLM Doesn’t Know
I hope you all had a fantastic week! I spent most of mine in sunny Las Vegas at the Senior Care Marketing & Sales Summit (SMASH). This was my 7th year at the event, and each time, the team puts on a truly fantastic show. While there, I had a long conversation with two long-time readers of this newsletter, both of whom were curious about my thoughts on content marketing – where’s going, where it fits in the AI world + what agencies/brands should be doing.
Well, never let it be said that I don’t take suggestions/feedback from readers – because 72 short hours (and a very long plane ride) later, this week we have the first of two issues on the topic.
My general thought is best summarized as follows: it is the best of times to publish something original and the worst of times to publish anything else.
If we’re all being blunt, a competent 750-word AI-generated blog costs about as much as the coffee you drank while prompting it. End result? The vast majority of marketers, SEO agencies and in-house teams now generate 4-10 “articles” each month, do some light editing and record the difference between the ~$5 (AI cost) + the ~$300 (what they used to pay to a freelancer per piece) as the value they’ve created/”savings” realized.
We’ve seen the same thing as we’ve conducted several audits: content output targets have increased substantially, per-piece budgets are in freefall and KPIs have shifted to volume/output based metrics.
This is exactly what I predicted 3 years ago would happen: the average cost of content production would decline precipitously.
But – as longtime readers have probably read 100 times across these issues – averages hide insights. That average bundles two things: (1) the cost of producing a generic article + (2) the cost of producing something genuinely worth reading. The former is in free fall. The latter has increased markedly.
The kneejerk reaction to that claim is that it’s nostalgia – the “write for humans, the robots can’t do soul” argument that shows up every week on LinkedIn, usually from someone selling ghostwriting. That argument gets made as a matter of taste. And while I think taste is the last true bastion of artistic intent, taste alone loses every budget meeting it walks into. Selling a CFO on “taste” doesn’t work. Trust me, I’ve tried.
But just because CFOs aren’t buying doesn’t mean the argument isn’t valid; it simply means you need a more compelling, mathematically-literate + financially-robust argument.
One final disclaimer: this article began as the case for writing exceptional content; but in the process of making that argument, I found that the same principles gave rise to a much more interesting content playbook.
So, without further ado, let’s get to it.
Content Is Denominated In Units
We’ve done ~89 audits of senior living communities + operators over the last 4 years. Throughout those, we’ve reviewed hundreds of agency and freelancer agreements – everything from full-service engagements to SEO-only retainers to full-service, AOR-style deals. The size and scope of those varies tremendously, but almost all of them are measured using the same ruler: output. Articles per unit time (month, week, quarter, year, whatever).
Some tack on additional metrics, like number of pages indexed, pages ranked or number of queries where the brand appears on page #1. Every one of those is a unit count, and every one improves automatically when unit cost falls.
The easier it becomes to spam content, the easier it is to hit every one of those goals.
What I have yet to see is the SoW that measures the goal of the program. At its most fundamental level, a content program buys hits: pieces that earn a link, are cited in an AI answer, rank for a query that converts, are shared in a brand’s Slack channel or are forwarded to a colleague. Everything else is inventory.
So, the number that governs the economics is:
Cost per hit = cost per piece ÷ hit rate
This is the exact same structure as Creative Is Math Wearing a Costume.
If you reduce unit cost 90%, then you need to retain more than 10% of your old hit rate to come out ahead. Cut unit cost 98% and you need to keep 2%.
You might read this and think, “If that bar was any lower, it’d be on the floor. There’s NO WAY generated content fails to clear it.” And, on a topic nobody has ever written about, it probably does. In any category that has ever had even a small number of buyers, there are almost no such topics left (thanks, SEO people). Making matters worse, our data (and the data from 1,000s of websites) shows that as the number of articles on a specific topic (especially if said topic falls into a YMYL category) increases, the hit rate doesn’t decline gradually; it collapses faster than Anakin Skywalker’s moral compass.
That leaves us marketers in an interesting predicament: there’s relatively little net-new topics left to address. Solution = find new sources of originality.
The Corpus Is A Pool
Any article has three 3 sources of originality: (1) the topic itself; (2) the claims made; (3) the perspective from which the article is written.
At its most fundamental level, an article is a set of claims about a given topic from a particular perspective – each one a discrete assertion such as “replacement windows can reduce heating costs” or “check whether the warranty transfers before you sign.”
Any topic has a finite set of claims that can be made from public information. Call it P.
If writers draw from that pool without checking what’s already been said, the arithmetic is identical to the reach math from Issue #187 (Let’s Talk About Ad Fatigue) with claims in place of people.
After I claims have been published, the probability that the next one is new is: e^(−I/P):
A mature commercial query has thousands of ranking-eligible pages. It’s safe to assume that each one makes ~10+ claims, against a pool rarely more than a few hundred claims deep (note: the number of claims is not static; as brands introduce new products, new variants, new materials, new technologies, new research, etc., the pool expands as buyers are now curious about those new things). However, even with that caveat, claims published in virtually every category runs 10x to 100x the pool, which means e^(−10) is 0.005%.
Ergo: most of what is published on your category’s core questions today is a restatement of a restatement. It’s not even derivative at this point; it’s the equivalent of going from Codex Leicester to a serial killer letter.
And all of that assumes that the article (i.e. the set of claims around a particular topic) is generated by a real, informed person doing honest research, earnestly pursuing the goal of creating something worth ranking.
You might read that and think most writers are still doing honest research…and you’d be wrong. Spend 10 minutes on LI or X and it’s obvious most have handed that step to a model. Candidly, that’s worse. A LLM doesn’t sample claims evenly from the open web; it reproduces them (approximately) in proportion to how often they appeared in its training data. Kandpal et al. showed this directly at ICML in 2023: a model’s accuracy on a fact tracks how many pretraining documents contain it. It falls off a cliff for anything rare. Ask it for an article on choosing a senior living community and you get the claims that appear on the most pages, which by definition are ones with the least information left in them. Google, for its part, has patented the opposite preference: US 11,354,342 & US 12,013,887 both outline an approach for scoring a document (defined as either an entire article OR a specific claim) by how much net-new information it adds beyond what the reader has already seen.
The response to this objection – at least, what I’ve heard – is that models recombine – that they assemble content in novel ways and output sentences nobody has ever written. And, to be fair, that’s true. But a new sentence and a new claim are different things. Recombining the corpus still gives you the corpus (there’s some fascinating math behind this). A public LLM can’t tell you your community’s inquiry-to-move-in rate by lead source, because that number isn’t in its training data. Anything an LLM can tell you, it can – by definition, tell your competitor. The only exceptions are private/closed LLMs, which have access to both the training corpus AND your organization’s non-public data. That can create some novel content/claims relative to the public LLMs, but is still limited to what’s in the view.
Google holds a patent on scoring exactly this (granted way back in 2022) entitled Contextual Estimation of Link Information Gain. It provides a remarkable method (as in something I find quite impressive) that ranks a document by how much it adds beyond what the reader has already seen. Whether any system implements it as written is beside the point, because every retrieval system has the same problem: when 400 pages say the same thing, nothing distinguishes page #401. The only way to make P bigger is to bring an input the corpus doesn’t contain. If you remember from the Creative is Math Wearing A Costume article, this is identical reasoning to why a new execution of a tired angle is still shown to the same people.
The Bar Keeps Rising
Content hit rate falling would be expensive enough on its own; but what we’ve seen is that the threshold for what counts as a hit has also moved during the same period.
In a 2024 Science Advances study, the authors (Doshi + Hauser) found that writers given AI-generated ideas produced stories evaluators rated as more creative and better written, with most of the gain going to the least creative writers. The stories were also significantly more similar to one another.
Statistically: AI raised the mean while compressing the variance.
That matters because “remarkable” is relative, not an objective threshold. Attention in any category is roughly fixed (e.g., a handful of citations per answer, one page of results, the 3-5 sources of information your ICP trusts); earning it means landing in the top slice of everything published.
On its own, everything I’ve written to this point COULD be a relatively persuasive case for more human-driven writing: it makes logical sense (if I do say so myself), it seems to align with what we all know + it has the added benefit of making human involvement objectively beneficial. But here’s where things get fun – the math ALSO makes sense:
If we assume the pre-AI (i.e. beginning of written history to 2022) writer population as a standard distribution of quality – mean 0, standard deviation 1. Then, assume AI lifts the mean by half a standard deviation while compressing the spread to 0.6.
Under those assumptions (which roughly match the findings from Doshi + Hauser above), here’s the probability each writer clears a given attention threshold:
The two cross at the mean gain divided by the spread lost: 0.5 ÷ 0.4 = 1.25 standard deviations, which = ~89th percentile. Below that line, the AI-assisted writer is the better bet. Above it, the unassisted writer wins, by more the higher the bar sits.
Think about that for a minute. If you’re not a top-10% writer (in your specific category/area) mathematically, you’re better off using AI. If you ARE a top-10% writer, you should steer clear.
But quality is only half the equation; we also have to consider volume. If output in a category rises 10x and attention doesn’t, the slot that required top-1% work now requires top-0.1% work to achieve the same visibility (well, maybe – this is a bit of a simplification that assumes even distribution).
Put another way: the same AI tools that made content cheap ALSO flooded the marketplace with that cheap content. That has the side effect of pushing the “remarkable” threshold further into the right tail, which is precisely the range where the people using that AI tool are least likely to reach it. End result? The cost of an average piece falls and the cost of a remarkable one rises, all for the same reason. Elegant, isn’t it?
You might read this and think the assumptions I made above are what’s responsible for the conclusion. You’d be wrong. If you double the mean gain to a full standard deviation, then the crossover moves to the top 0.6% – still below where attention concentrates for any contested query. If you’re a math person, the crossover is Δμ ÷ (1 − σ), which is finite for any σ < 1.
A higher mean with a narrower spread always loses the right tail; the parameters only decide where the loss happens.
Exercising The Muscle
Everything above treats writers (and writer ability) as fixed. Any person who has taken or taught a writing (or any artistic) course knows that’s not true. People can improve. They can regress. Writing ability is not a static property of a person.
When you review your first draft of anything, you often see (or, at least, I do) where the outline was incomplete, where the argument was weak, how a critic could attack the claims I’m making, and (quite rarely, but it does happen) why my argument – as written – is wrong. But if you delegate the drafting process to an LLM, the critical thinking muscle that draws those conclusions atrophies. This is no different from the artist who uses photoshop in place of painting by hand. It took Da Vinci 31 years (~32, to be precise) of reps as a professional to gain the skill + eye to create the Mona Lisa (he painted his first major contribution at age 18 – painting the Angel in the Baptism of Christ; he painted the Mona Lisa when he was 50-51); I doubt he ever would have gotten to that level of proficiency if he hadn’t put in the work.
A 2025 MIT Media Lab study put EEG caps on people writing essays with ChatGPT, a search engine and nothing (i.e. completely unassisted). The ChatGPT group showed the weakest neural engagement. Most members of that segment could NOT accurately quote a single sentence from the essay they’d just turned in. Think about that for a minute – writers that can’t even recall their own writing.
This leads us to a bit of a quandary: at the same moment genuine skill is becoming more valuable, the muscles that lead to it are atrophying. That’s the inevitable result of the KPIs brands are using (namely, volume + cost per piece). The only way to achieve most client production goals is to use AI; the highest-probability path to create something remarkable is to not.
A few months ago, I made a variant on this argument – that the judgment of the few people in your organization who are actually good at this is the one input you can’t buy more of AND remarkable content is made almost entirely out of it. Most brands/companies/agencies are spending that finite resource on prompt-then-polish – and in so doing, allowing the capacity behind it to atrophy. When supply of an input declines while demand remains constant, the price goes up. That’s economics 101. Expect the writers who can still do this to get more expensive every year while the per-blog rate for everyone else races toward $0. That’s where we’re headed.
The Reader Who Has Read Everything
The usual version of “write for humans, not machines” treats this as a trade-off: content optimized for AI search on one side, content for “real people” on the other. The brand/agency/organization mandate is to pick one. Candidly, most of the market has chosen the machine – either out of perverse incentives (we need to make X pieces, AI is the only way) or out of a misunderstanding of what LLMs actually want (e.g. the theory that discovery is moving there and the machine doesn’t care about prose).
The more time I’ve spent with both AI systems AND the patents that the major AI labs have published, the more I reject binary framing.
Let’s start with what an answer engine actually is: a compression of a corpus. The size of that corpus can vary (and that’s why AI labs are scanning old books and trying to buy company data), but fundamentally, that’s what’s going on. More data = bigger corpus = more robust compression of said corpus.
Virtually every LLM has read every listicle, every “ultimate guide” + every restatement of the same ~100 claims in your category. That makes the LLM the one reader guaranteed to already know everything a generic article says. When it cites something, it needs at least one source (and I’d argue, many hallucinations occur as a result of the model having multiple sources with differing answers, leading the AI to “believe” that the answer is variable, not fixed). That source is the article/blog/guide/video where a claim originated. The 100th page to repeat a near-identical claim yields nothing new to cite, which means its value to the model is null. There’s literally no point in wasting compute on your generic-as-all-hell article.
But here’s where the binary rejection comes in: a human wants the same thing.
The adult child comparing 4 assisted living communities near them doesn’t need a 5th page explaining assisted living vs memory care. The B2B SaaS buyer doesn’t need another guide on attribution. The young couple trying to figure out which bank to open their new joint accounts doesn’t need ANOTHER article on savings accounts vs. CDs vs. money markets.
What each of those people DO need is something that addresses their real, often unsaid, concerns (not the ones the internet has already resolved; the ones that feel unique but probably aren’t) – what happens to Dad if his Parkinsons’ gets worse + he falls? If we put our nest egg in a CD, but a great house comes on the market, can we get it out? What happens if we launch a new product that has a very different customer journey? Will our attribution average it in, or can we segregate attribution by product?
The pendulum didn’t swing too far toward the AI stuff; it swung toward a version of AI that doesn’t actually exist. Writing for humans + writing for answer engines reduces to the same requirement: add something new to the conversation/corpus. Any brand treating those 2 things as competing goals is optimizing for something that doesn’t exist, then wondering why it isn’t working.
The Opportunity Is Wide Open (For a Narrow Reason)
When I say this on stage, most people (especially the AI maxi-bois™) assume that I’m an AI bear or (as I was recently told) that I’m in denial about what AI can do). Nothing could be further from the truth. I think AI today is roughly equivalent to the internet in 1996. The potential is incomprehensibly vast. It will throw the world as we know it into chaos, and from that chaos a new order will emerge.
But – as the late (and not-so-great) Petyr Baelish said, “Chaos is a ladder.”
SEMRush analyzed 1,094 categories in AI search. They found ~85% had either no consistent leader or only an emerging one. This is great news for every brand: most of the territory is unclaimed. The digital world is being remade anew, which makes everything a free-for-all.
As any good marketer/agency owner, I’ve spent a truly insane number of hours reading every playbook, reviewing the research/patents being published and demo-ing far too many AI tools. Its my job to understand the landscape so I can figure out how to win. The one constant I’ve found in all of this is that most of what’s out there misreads the evidence.
SE Ranking’s study of 129,000 domains found sites with 32,000 – 350,000 referring domains averaged 5.6 ChatGPT citations vs. 1.6 for sites with 1 – 200 referring domains. The standard conclusion from that data point (and the one pushed by link builders most heavily) is “build links.” Most people read that + wonder how any other conclusion is possible. For one, it misses the confounder: big, established brands have more of both, so as a prescription the finding mostly says “just get bigger.” Yeah, that’s great – except how?
But what if we run this same thing through the corpus model instead?
If you publish a finding that didn’t exist before, then every writer (especially the AI ones) who repeat it afterward (likely hundreds, in a mature category) has a reason to reference/credit the origin. We know from Google’s own patents that LLMs treat references + mentions similarly to traditional links. End result? The subsequent/derivative works add nothing to the corpus, but their references add credibility to yours.
Mentions, references + links are therefore an output of information gain in an AI-centric environment, not an input to it. The originator benefits disproportionately from every restatement; the 100+ restatements split nothing.
OK, So Now What?
This is all well and good (and I think, quite interesting) – but what you’re probably looking for is what do I do with this little bit of information? Good news: there’s a playbook you can run. It’s the same one I articulated years ago in the New Content Playbook, just updated what I’ve covered above:
1. Map what your category already says. Run the ~15 questions each segment of your ICP actually asks through ChatGPT, Claude + Google’s AI Mode several times. Read the cited pages. List every claim they make in a Google Sheet. That’ll give you ~90% of the covered pool. Nothing there is worth publishing again UNLESS there’s a material change in the underlying environment.
2. Inventory what only you know. Your CSMs, CRM + experts all hold claims the corpus doesn’t. For example, in senior living: what has families up at night, what questions adult daughters are asking on tours, what actual families/residents email the ED asking about in the months after a loved one moves in, what technologies/information families are asking for. In home services: what families are prioritizing; which improvements/updates realtors are prioritizing in resale; which warranties are actually being honored; what new approaches/designs/materials are starting to be asked about (and which ones are falling out of favor); why families aren’t moving forward with new projects (and why they are). That’s just a sampling; there are thousands of niche claims/questions that your organization knows about (in some sense) that aren’t being addressed in your content. Find them.
3. Publish 1 finding a stranger can safely repeat. One sentence with a sample, a method + a result (plus any the limitations stated next to it). The goal is to share information that isn’t found anywhere else.
4. Write everything around it for the person deciding. Build the supporting pages around the buyer’s blocking beliefs (I listed 8 for a window replacement in The Conviction Gap); answer the question in the first sentence of each.
5. Distribute where the answers already come from. Your citation map tells you which publications, creators + review sites inform your buyers’ answers. Send them the finding, the data and a chart they can embed. The pitch is “we used our data to measure something your audience already wants to know – here’s how + here’s what we found.”
6. Measure hits. Count the pieces that earned a citation, a link, a converting ranking or a forward, then compute cost per hit. You’ll probably be surprised at the result.
7. Make your best people write first drafts. Use the tools for transcription, formatting, metadata + schema. Keep the 1st draft of anything you want to be remarkable with a human who has to work out what they think while writing it. That’s the practice that builds the skills/expertise required to keep the tail of your distribution from closing.
There’s no denying that AI has made content cheap. But, by the same token, AI is the thing that has made novel information-adding content more expensive. The chasm between the two is only getting wider as more brands publish more middle-of-the-bellcurve stuff.
AI has already read every obvious, trite, done-by-someone else thing that most content writers/agencies want to publish. So, write the thing it hasn’t.
