August 17, 2026 · 7 min read
The Unobserved Intelligence of LLMs
Everybody watches the question and the answer. The intelligence is in between, and nobody is looking at it. Ten days at Hexa, a real market, and the layer of the stack nobody owns yet.
The Unobserved Intelligence of LLMs
Brands are being described to their customers by a machine, in a paragraph they never see written, built from sources they never chose. Most of them are now paying somebody to measure this.
Everybody is watching the question and the answer. The intelligence is in between, and nobody is looking at it.
The problem: nobody knows what happens in the middle
People used to type keywords into a box and get ten blue links. You could measure that. Rank 3 for "running shoes", rank 8 for "trail running shoes". Boring, but honest.
Now people type a sentence. A model reads it, decides what they actually meant, breaks it into pieces, goes looking for things, throws most of it away, and writes a paragraph. The brand either appears in that paragraph or it doesn't.
Everything that matters happens in the middle. And the middle is a black box.
Prompt
best CRM for a small team
Intent distilling
What do they actually want? A CRM recommendation for a team of a few people.
Decomposition
- best CRM small team
- what CRM should a 5 person startup use
- CRM price per seat
Retrieval
- review site · A
- forum thread · B
- vendor page · A
- listicle · B C
- news · C
- comparison · A B
Selection & ranking
- review site · A
- forum thread · B
- vendor page · A
- listicle · B C
- news · C
- comparison · A B
Synthesis
Writes one paragraph from what was kept. Brand B was in the sources. It is not in the paragraph.
Answer
For a small team, Brand A is a strong pick…
What happens between the prompt and the answer: intent distilling, decomposition, retrieval, selection and ranking, synthesis. Only the prompt and the answer are observable.
Every one of those arrows is a place where a brand gets in, or gets dropped. Nobody selling GEO today can tell you which arrow killed you. They can only tell you that you weren't in the final paragraph.
This is a statistics problem, not a ranking problem
Here's what breaks people's brains when they come from SEO.
Ask the same question twice, you get two different answers. Ask it in a slightly different way, you get a different set of sources. Ask it tomorrow, different again. There is no rank 3. There's a distribution.
So the work is sampling work:
- You don't run a prompt. You run it n times. One answer is an anecdote. A hundred answers is a measurement. Mention rate 34% ± 6 means something. "We appeared" means nothing.
- You cluster prompts by intent, not by string. "Best CRM for a small team" and "what CRM should a 5 person startup use" are the same question wearing different clothes. If you treat them as two keywords you'll double count and mislead yourself.
- You measure more than presence. Mention rate. Share of voice against named competitors. Position in the answer. Sentiment when mentioned. Which source got cited to justify the mention.
- You watch variance, not just the mean. A brand mentioned 90% of the time on one model and 10% on another does not have a 50% problem. It has two different problems.
- You need a control. Otherwise you can't tell if your content did something or the model just got updated on Tuesday.
That's the analytical spine of the whole category: sampling design, clustering, attribution, confidence intervals. It's closer to survey research than to SEO.
—
n = 0 · mentioned 0
Run the prompt. One answer is an anecdote.
The maths that actually earns its keep
Four tools do most of the work here.
Wilson score interval, for "are we sure?" Mention rate is a proportion, and the normal approximation falls apart at small n and at rates near 0 or 1, which is exactly where brands live. Wilson doesn't. And it tells you your sample size up front: to measure a mention rate to ±5 points at 95% confidence you need roughly 400 runs of that prompt cluster. At ±10 points, about 100. If your dashboard reports a number off 20 runs, it's reporting noise.
Beta-Binomial updating, for "did it move?" Model each prompt cluster's mention rate as a Beta posterior. Prior from last month, likelihood from this week's runs, posterior out the other end. It handles small samples honestly, it accumulates evidence over time instead of throwing last week away, and it gives you P(rate went up) directly instead of a p-value nobody can interpret.
Bayes, for "which lever moved it?" This is the one people skip. You observe a mention. What you want is P(our new page caused this | mention), and what you can measure is P(mention | page is cited). Bayes flips it, and the prior is the base rate of being mentioned without the page at all. Run it and most "our content worked" claims collapse. The base rate was already high and nobody checked. Same trick for the source chain: given a mention, which of your properties was cited, weighted by how often that property gets cited at all.
KL divergence, for "how far is our synthetic set from reality?" Cluster prompts by intent, get a distribution over clusters for your synthetic set and a distribution over clusters for real traffic, and measure the gap. KL (or Jensen–Shannon, which is symmetric and bounded, so easier to report) gives you one number for how wrong your simulated prompt set is. Without real prompts you can't compute it at all, which is the quiet problem with every synthetic-only platform. They cannot measure their own error.
Two more, briefly: difference-in-differences against a control set of competitor brands, so a model update on Tuesday doesn't get logged as your win; and a multiple-comparisons correction, because if you test 200 prompt clusters at p<0.05, you get ten "significant" results from pure chance.
What I found in the market
I used the ten days to actually talk to people. Companies with the problem, and companies selling the cure. Profound and the rest of that crowd.
The market is real. The pain is real. Brands are watching a channel they can't see, and it's growing.
But almost everyone is running on synthetic prompts. A model generates a list of questions customers "might" ask, and the platform measures those. It's clever. It works well enough to sell. I built with synthetic data too, and I got useful signal fast.
It's still a guess. A very well dressed guess about what humans actually type. And the second problem is worse: even with perfect prompts, you're only watching the input and the output. The middle is still dark. You don't get the sub-questions. You don't get the queries that fired. You don't get the sources that were considered and dropped.
Draw the loop and you can see where it snaps.
The working loop
The loop everyone runs
Two breaks, and they compound. Break 1 means you're optimising for questions nobody asked. Break 2 means that even when you win, you can't say which of your changes did it, so the next decision is a guess too. A loop with no clean feedback isn't a loop. It's a treadmill with a dashboard bolted on.
The prize is the raw data. Real prompts, real sessions, real intermediate steps. Whoever has that isn't guessing about the middle. They can see it.
Where GEO companies actually sit in the stack
Every data industry ends up with the same three layers, and it's worth knowing which one you're in.
Layer 1: collection. Owns the raw data at the source. Panels, crawlers, clickstream, receipts, telco logs. Slow to build, expensive, unglamorous. Best margins, hardest to copy.
Layer 2: analytics. Licenses or buys raw data, cleans it, normalises it, turns it into metrics and benchmarks, sells a dashboard. Most startups live here because it's the easiest place to start.
Layer 3: activation. Takes the metrics and does the work. Agencies, tools, content shops. Closest to the customer's budget, lowest margin.
There's usually a broker sitting between 1 and 2 as well, like LiveRamp, Snowflake Marketplace or the card-data resellers, whose only job is moving data from people who have it to people who want it.
Old SEO maps onto this cleanly. Collection was crawlers plus clickstream panels quietly sold by browser extensions and ISPs. Analytics was Semrush and Ahrefs, and note they built their own crawlers, so they own layers 1 and 2, which is exactly why they got big. Activation was every SEO agency on earth.
Two rules fall out of this, and they decide who survives:
- Layer 2 on its own gets commoditised. If ten companies license the same feed, they all report the same numbers and end up competing on UI and price.
- Layer 1 can walk up into 2 and 3 whenever it likes. It never works in reverse.
Now look at GEO today. Almost every player is a layer-2 business with no layer 1 underneath it. There is no feed to license, because the data lives inside somebody else's chat logs. So they manufacture the missing input, synthetic prompts, and sell analytics on top of data they invented.
Real prompts and real sessions are layer 1 of this industry, and nobody owns it yet.
- Layer 3 · Activationagencies, tools, content shops
- Layer 2 · Analyticsalmost every GEO player, on synthetic prompts
- ↕ brokerLiveRamp, Snowflake Marketplace, card-data resellers
- Layer 1 · Collectionreal prompts and real sessions: nobody owns it yet
So why would anybody give you their data?
They wouldn't. That's the whole game, and it's where most ideas in this space quietly die.
Nobody wakes up wanting to sell you their prompt history. There are only two honest ways in:
- Build something people want for a completely different reason. A service that stands on its own, that they'd use if you never touched the data. The data is a by-product of value you already delivered. This is how every data business that ever worked, worked.
- Pay for it. Straight up. Buy the data from people who own it, and be clear about what you're buying and why.
Everything else is either a privacy problem or a wish.
That constraint isn't a footnote on the business. It is the business. Any GEO company that gets to real behavioural data before its competitors doesn't win by 10%. It stops guessing while everyone else is still guessing.
Why I still think this is one of the best places to be
The agent already distils intent from your prompt. It already decides what to search. It already decides what to keep. That machinery is being built out right now, by everyone, very fast.
Understanding what happens between the prompt and the answer, which actions lead to which answer, and getting that from real data instead of simulated data, is going to matter to every company that sells anything.
And now the quiet part, because most people selling into this space will not say it. This is a good place to be for a limited time.
Two things close the window.
The first is law. Data rights are moving in every jurisdiction at once, and they are all moving the same way. Collecting behavioural data, using it, and reselling it each get harder, slower and more expensive. A business whose entire edge is holding real prompts and real sessions is a business whose entire edge is one regulation away from being a liability.
The second is that the data stops being special. What I said earlier, that synthetic prompts are a well dressed guess, is true today and less true every quarter. Look back at the anatomy of the middle. Intent distilling, decomposition, retrieval, selection. Every one of those is a step a model performs, which means it is a step a model can be asked to perform again, on demand. Once a model can reconstruct a realistic distribution of what people actually ask, the premium on having observed it collapses.
The cost of doing it yourself falls at the same time. Running a prompt cluster a few hundred times and doing the statistics on it is not an expensive computation, and it gets cheaper every year. Today you buy a platform for that. Soon it is a scheduled job inside your own stack, and nobody buys a dashboard for a scheduled job.
Then the last squeeze, and it is the one I would watch. The labs own the only genuinely unobservable part of this. As the model itself commoditises, and it is commoditising, they need revenue that is not tokens. Selling structured visibility into what happened between the prompt and the answer is the obvious diversification, and they are the only ones who can sell it without asking anybody's permission. On that day, every independent GEO company is reselling something its own supplier can offer directly.
So expect noise. Big rounds, big titles, a category that looks like the next Semrush for about three years. Some of that money is real. Most of the moats are not.
I still think it is one of the best places to be. I am less sure it is a good place to stay.
The part where I tell you this is real
I found all of this during ten days at the Hexa incubator. You show up, they hand you a problem, and your job is to prove their idea has merit. That's the deal.
It isn't a thought experiment. The idea came out of those ten days with enough behind it that Hexa will put around €750,000 into building it, and is carrying it forward with my partner.
I wasn't selected to keep going, but it was still one of the best ten days I've had in a long time. Real problem, real users on the phone, real market, no theatre. The work stood up.
I'd do it again tomorrow (but this time with YC 🤣).