In economics, there are many questions that sound obvious but turn out to be hard. One of them: what is the next unit worth?
The next dollar of ad spend. The next blog post. The next field in your onboarding form. The next meeting on your calendar.
Marginal analysis is the practice of evaluating decisions at the margin, comparing the incremental cost of doing one more of something against the incremental benefit.
The concept is not complicated.
It’s taught in the first week of any microeconomics course.
But it’s rarely applied to the decisions that growth teams, marketing leaders, and operators make every day.
We tend to think in totals and averages. How much did we spend? What’s our overall conversion rate? How many posts did we publish this quarter? These are useful numbers, but they don’t tell you whether the next unit of effort is worth the cost.
That principle applies beyond measurement. It applies to every activity. In this essay, we’ll be applying expected values and “next unit” thinking to AEO, content throughput, meetings, and even measurement itself.
The Measurement Curve
Let’s start with measurement itself (per my friend Matt Gershoff’s post). What’s the marginal utility of an additional measurement, analysis, or model?
Marginal analysis here means thinking in terms of the incremental value of doing one thing versus another, rather than asking whether a method is generically “good” or “rigorous.”
In analytics, you often have multiple possible approaches. Marginal analysis asks:
What additional information or decision quality do I get from using a more sophisticated method, relative to its additional cost/effort?
For example, suppose you’re trying to estimate AI search’s influence on pipeline. You could:
- Look at AI referral traffic in GA4.
- Add self-reported attribution.
- Manually analyze the self-reported responses.
- Build a multi-touch attribution model.
- Run an incrementality experiment.
- Build an MMM (media mix modeling) system.
Each step might theoretically improve your understanding. But the relevant question isn’t “What’s the best methodology?” It’s “Is the next level of rigor worth what it costs me?”
If self-reported attribution gets you from “we see almost no AI influence” to “holy shit, AI appears to influence a substantial portion of leads,” that’s potentially enormous marginal informational value for very little effort.
Going from that to a beautifully specified causal model might cost 50× as much and only make the resulting business decision 5% better. Low marginal value.
This is closely related to expected value and decision theory: choose the analytical method based on how much the additional precision could actually change the decision.
That’s probably why Matt says it’s especially underrated when selecting methods/approaches. Analysts can optimize for methodological sophistication when they should be optimizing for marginal decision value per unit of effort.
Again, the question isn’t “which is better” in the abstract. When running an A/B test, you can collect eternal samples, which may increase precision, but most analysts accept 2-4 weeks with a p value threshold of < .05 to be perfectly useful for business contexts.
Something to consider, without even getting into S curves and asymptotes, is asking “what would I do/change if I had this information?”
Douglas Hubbard put it well in How to Measure Anything:
“If a measurement matters at all, it is because it must have some conceivable effect on decisions and behaviour. If we can’t identify a decision that could be affected by a proposed measurement and how it could change those decisions, then the measurement simply has no value.”
The first measurement reduces the most uncertainty. Each subsequent layer of precision costs more and reduces less. At some point, you’re paying for confidence intervals that are narrower than the variance in your actual business decisions.
The Prompt Problem
This plays out at the micro level too. Take AI visibility measurement.
If you measure a single prompt for a single intent, say, “best SEO agencies,” you’re measuring one model’s response to one phrasing at one point in time. The variance is enormous. Run the same prompt tomorrow and you might get a different answer. The signal-to-noise ratio is poor.
Run five variations of the same intent, “best SEO agencies,” “recommend me top SEO agencies,” “what are the most respected SEO agencies,” and the variance drops substantially. Run each prompt a handful of times per day, and you’re also converging on a more stable probability of your brand’s appearance.
Now run a thousand. The variance drops further. But the marginal reduction from prompt 500 to prompt 501 is nearly zero, while the cost (compute, time, complexity of analysis) scales linearly. You’re buying precision that no longer changes what you know.
This is a useful example because the math is easy to see.
Plot the variance reduction against the number of prompts and you get a curve that flattens quickly. The first five prompts buy you most of the insight. The next 995 buy you decimal places. The practical question is where on that curve the cost of the next prompt exceeds the value of the information it provides.
Here, for example, is research from Profound that is pretty explicitly asking a marginal utility question: “Is once per day enough?”
They compare running prompts once per day vs. 10× per day across 753 prompts and seven platforms. Their central finding is that, for brand visibility, 10× the measurement effort only reduces day-to-day variability by about 11%. The aggregate visibility estimates were 78.7% versus 80.4%.
This is, by the way, effectively how the rule of five came to be in usability testing. NN/g found that with 5 users in a user test, you could identify 85% of issues and the curve flattens after that:
These are analyses that you can run, even with limited data. We, for example, are right in the middle of research on the number of intent variants needed to reduce variance on brand visibility (to be published soon).
The Gym and the Garden
Not all marginal curves decline smoothly from the first unit.
Some have a threshold effect, below which effort is effectively invisible.
Content publishing is one.
A company that publishes two blog posts per month is unlikely to see measurable organic results. Not because the content is bad, but because the volume is below the threshold where compounding effects become visible. It’s like going to the gym once a month. Arguably better than zero, but functionally there is little difference to show for it. If anything, you just get really sore once a month for no discernible outcome. The body doesn’t register it as a signal to change.
Increase to two or three times per week and the curve steepens. The soreness-to-results ratio shifts in your favor. Each additional session builds on the last. The body adapts.
In content terms, the indexing velocity increases, internal linking structures develop, topical clusters fill out, and the compounding effects that organic growth depends on start to engage.
Push it too far, and you shift into diminishing marginal utility, or worse, blowup risk (let’s call it the Icarus Effect, which is more generous than the colloquial Mount AI).
Quality suffers, you’re scraping the bottom of the barrel on topics, and in search, specifically, you’re basically begging for an algorithmic or manual penalty (depending on the level of your quality).
The gym equivalent is overtraining, right? Either you’re simply not growing much more per unit of effort, or you risk energy and burnout that takes you off the floor and into physical therapy.
The shape here is like an S Curve.
Flat at the bottom, steep in the middle, flat at the top, and potentially negative past a certain point. This is likely true for many marketing efforts, including outbound sales, advertising, content marketing, and PR campaigns. A lot of companies are stuck in the first flat. They ship or spend sporadically, see no results, and conclude that campaign doesn’t work for their business.
They may be right. But they may also just be below the threshold where results become visible.
The Expanding Circle
Paid media is the textbook case of marginal decline.
Your highest-intent audience converts first. Each expansion outward reaches people who are progressively less ready, less aware, and less aligned with what you sell. The cost per acquisition rises while the quality of the acquisition falls. Your marginal dollar, however, is still often worth it at declining efficiency due to defensibility, market share, and lifetime value. This is well understood in paid.
In organic, this may apply, but less directly. Certainly there are outer limits wherein you risk topic dilution outside of your core sphere of expertise. However, unlike paid, moving “upstream” towards lower intent funnel stages may actually support your salience and conversion at the decision stages.
In SEO, this was due to topical authority, domain authority, and a rising tide lifting all boats. In AI search, it may simply provide model context for recommendations downstream of a user solving a problem (in other words, if you can bridge the gap between “how to fix a stove” and “best stove cleaner,” you may be able to one shot a user conversation in AI search).
Or as I once told my research analyst, there’s no (realistic) upper limit for producing good research on AEO right now…
The Number One Problem
Which leads me to my next point: what game are you playing? Because sometimes the rational move is to keep spending past the point of diminishing returns.
Being first in a category accrues benefits that don’t appear on a marginal analysis chart (let’s call this The Matthew Effect). Being Michael Jordan in basketball. Being Nike for running shoes. Being the company everyone cites when they talk about a particular approach to marketing (you’ve seen it with Drift, Gong, HubSpot, Liquid Death, Red Bull, etc.).
These are winner-take-most positions, and the returns at the extreme aren’t diminishing. They’re compounding through a different mechanism: reputation, default status, brand gravity.
Essentially, these are the domains in which flywheel dynamics apply. I wrote about this with regards to proprietary data in The New Gold: how data aggregation and network effects produce increasing marginal returns because each additional data point makes the whole dataset more valuable. The same logic applies to category authority.
If you want to be the research-led growth company, there’s clear value in going from zero original research to one study per year, and more value per unit going from one per year to one per month. Past that, the channel metrics flatten. Each additional study drives roughly the same incremental traffic as the last.
But being THE research organization, the one people cite not just for the individual research, but for the meta concept of being the research company, well that may well be worth spending past the point of rational nominal spend. The second-order effects operate on a different curve: citations, backlinks, inbound opportunities, brand association. The first-order metric says stop. The second-order position says keep going.
Going back to our gym analogy, imagine you are not you, and you are Arnold Schwarzenegger in his early days. Of course, the same mechanics apply to lifting weights 1 versus 2 versus 3 times per day. The third workout will inevitably produce less marginal muscle for the effort required.
However, because he is competing to be number one in the universe, which accrues its own second-order benefits and prizes (such as becoming a mega wealthy actor and a state governor), it is worth it to eke out the last increment of potential investment.
Races are won by milliseconds. If the benefits of being the absolute best are astronomical compared to number two, then the investment is warranted regardless of the first order marginal analysis.
The Labyrinth (or Why You’re So Busy)
Now flip the lens inward.
Every process step, data collection point, meeting, and internal memo has a marginal cost. Almost nobody asks about the marginal value, nor do they delete steps in the process that have overstayed their welcome.
You add a field to your onboarding form to collect industry data. That data is supposed to personalize the product experience. Sounds good in theory, as literally all of this stuff does (…that’s the problem).
But has anyone measured whether the personalization moves activation, retention, or satisfaction?
One of my favorite essays I’ve ever written is on this topic. It covers the flip side to personalization, which is complexity, management costs, and marginal analysis.
Personalization is one of those things that few will argue against. A personalized email must, in a sense, be better than a generic one. But is it? And at what cost?
Personalization on a website incurs not only marginal analyses on its impact, but also the cost of managing 100, 1000, unlimited dynamic personalization arms. This has downstream impacts on customer support, sales, and marketing that often supersede the benefits of running it.
I am not arguing here that personalization is useless. Depending on the unit, it could be wildly effective. Remember the first example I gave with regards to AI search measurement? If you were running ABM campaigns, and were previously personalizing nothing, and then you went to building custom landing pages for your top 100 accounts, then that surely could prove net positive.
Where it gets wily is when we think of it as an unbridled good and hit the long tail of absurdity.
Which, returning to your internal organization practices, is something to consider.
Think about meetings, or more broadly, artifacts and rituals that are about the work but not the work itself. The weekly sync that made sense when the team was five people and is now a standing meeting for twenty. The quarterly business review deck that takes a week to prepare and changes no decisions. The approval workflow that was added after a single incident and now applies to every project regardless of risk.
Each of these was added for a reason. Few were evaluated on whether the next unit of process was worth the next unit of cost. And unlike external activities, internal processes almost never get cut, they only accumulate. The labyrinth grows.
Hubbard’s question again: what decision does this meeting inform? What behavior does this data collection change? If the answer is unclear, the marginal value of the activity is unclear. Which usually means it’s close to zero.
As Benyamin Elias wrote, “tear down every process that isn’t based on a strong understanding of what makes the work successful.”
Running Your Own Marginal Analysis
Marginal analysis can be very fancy (see prompt variance research above, or advertising spend modeling), but it can also be a simple question with an a priori estimate. To me, it’s more of a mental model, which sometimes warrants a true “analysis” and sometimes just a thought experiment.
Here’s the version I’d suggest for anyone running a growth team, a marketing program, or an internal operation.
For any activity you’re doing, ask three questions:
- What decision does this inform or what outcome does this drive? If you can’t name one, the activity likely has no measurable value. It might have value you haven’t articulated, in which case, there’s a lot of value in articulating it. But if you genuinely can’t connect the activity to a decision or outcome, it’s a candidate for removal.
- What is the shape of the curve? Is this an activity where the first unit is the most valuable and each subsequent unit is worth less (measurement, ad spend expansion)? An activity with a threshold below which results are invisible (content publishing, fitness)? Or an activity with increasing returns where dominance changes the game (category authority, network effects, data aggregation)?
- Where are you on the curve right now? If you’re below the threshold, the answer isn’t to optimize, it’s to increase volume until you cross it. If you’re in the zone of diminishing returns, the answer is either to hold (if you’re playing the channel game) or to push through (if you’re playing the category game). If you’re past the point of negative returns, the answer is to pull back and consider what sunk costs you’ve fallen prey to.
Most growth teams default to one of two modes: always more, or always optimize.
The useful move is to know which mode matches the shape of the curve for each specific activity.
Some things need more volume. Some things need less. Some things need to be cut entirely. And some things need to be pushed past the point where the spreadsheet says to stop, because the spreadsheet is only measuring the first-order effect.
Want more insights like this? Subscribe to Field Notes
