— RESEARCH · AGNOSTIC FRAMEWORK
How to operate honestly on terrain nobody controls.
Frontier research: a reusable skeleton for working with opaque, probabilistic, shifting AI systems — without selling false certainties. The first instance is AEO, our core service.
The old disciplines assumed determinism. Here there is none.
Every domain where you work with an AI system has the same four traps. The old disciplines — SEO, classic media buying — assumed stable positions and causality. At the frontier, you have neither.
Opaque
You can't see why it does what it does.
Probabilistic
It doesn't repeat identically from one run to the next.
Shifting
It resets with every model update.
No causality
You can't prove what caused what.
The anti-theater rule: every abstract component must anchor to at least one concrete, repeatable, measurable artifact. If a component can't be anchored, it's smoke, and it gets cut.
14 questions with no clean answer — yet
The questions the framework was born from. They start in AEO but describe any opaque AI terrain. We don't solve them — we name them, so we don't sell certainties over them.
- 01
Does the recommendation come from memory or from retrieval?
When an AI recommends your brand, is it drawing on knowledge “frozen” in training (months or years old) or on live search? The output won't tell you — you optimize for one and move nothing on the other.
- 02
Does the AI “know” your brand, or rebuild it every time?
If “brand” isn't a stable node in the model but a statistical reassembly of everything that co-occurs with your name, then your identity is at the mercy of its semantic neighborhoods — controlled by what others write next to your name.
- 03
If a brand is co-occurrence, can it be poisoned?
If a brand's meaning is statistical, a competitor (or an attack) can shift your associations by flooding the corpus. And conversely: how long does it take to “teach” the model a new association before it sticks?
- 04
Is there a “rank 1”? Not the way SERPs have one.
There's a hidden ranking at the retrieval layer (candidate documents, ordered), but probabilistic generation sits on top of it. You optimize for a ranking you can't see, followed by a controlled lottery.
- 05
Non-determinism: what does “I'm recommended” mean if it isn't reproducible?
The same question, asked twice, can return different brands. “Visibility” becomes a probability distribution, not a rank — and most agencies treat it like a rank.
- 06
How often does it update — and why can't you audit it?
It depends on the engine and the layer. Training knowledge can be 1–2 years old; retrieval is instant. A rebrand or a scandal — when does each engine “find out”? There's no changelog.
- 07
There's no single AEO — there are N opaque, divergent judges.
Each engine has different training data + different retrieval + a different system prompt. You optimize simultaneously against several black boxes that don't agree with each other.
- 08
Attribution is a black box.
When the model says “Brand X is good for Y” — what caused it? A Reddit thread? A thousand reviews? Your site? There's no classic referrer, no Search Console. The feedback loop is broken.
- 09
“Page 2” disappears.
Search gave you 10 links and you could scroll. AI gives 1–3 recommendations. Winner-take-most, with no visible long tail. If you're not in the synthesis, you don't exist — and there's no “see more”.
- 10
The customer is no longer the human — it's the model.
The AI answer is mediated: the model reframes intent, decides what's relevant, picks the brands. You have to convince the model, not the human — and the human may never see your name.
- 11
Cited ≠ influential.
The visibly cited sources aren't necessarily the ones that shaped the recommendation. You can be influential without being cited, and cited without having mattered. Visible attribution lies.
- 12
Self-referential contamination.
As the web fills with AI-generated text, models train on and retrieve text also written by AI. Are you optimizing for reality, or for the echo of their own earlier hallucinations?
- 13
The hallucinated recommendation — with no right of appeal.
A model can recommend you for a use case you don't serve, or invent attributes for you. You can be “well optimized” into a position that misrepresents you — with no SLA, no recourse.
- 14
Personalization destroys the single “answer”.
When models personalize (memory, history), there's no longer a single answer to optimize. Your rank is a distribution across millions of contexts you can't observe.
The Websem Agnostic Framework: eight invariant components
Instead of inventing a new method for each AI domain, you instantiate the same skeleton. For any new domain, you fill in these eight — each with the question that instantiates it. The discipline stays; it just gets re-instantiated.
- 01
The honest premise
Why is this domain non-deterministic, and which old certainty falls away here?
You name the fog up front. That defuses false promises before you make them.
- 02
The system's layers
Into which differently-behaving layers does the system decompose?
No AI system is a single box. Decomposing it into layers tells you where you can act and where you can't.
- 03
The knowledge boundary
What do we know with reasonable certainty vs what is fundamentally opaque?
The line between the two is your intellectual capital. Honesty here becomes the brand.
- 04
The measurement boundary
What can we instrument repeatably vs what stays in the dark?
You can almost never measure why. You can measure how much, and how it evolves.
- 05
The levers
What can we actually move, mapped onto the layers?
The “to-do” part. Everything not here is only observed, not controlled.
- 06
The delivery stack
Foundation → Monitoring → Interpretation?
The three levels are invariant. Only their content changes per domain.
- 07
The promise boundary
What do we promise, and what do we explicitly refuse to promise?
The explicit refusal is part of the sale, not a weakness.
- 08
The positioning
How does honesty about limits become the commercial differentiator?
“We don't sell you false certainties; we sell you the only serious discipline + the measurement.”
The 3 layers every AI recommendation comes from
Any “knowledge” an AI has about a brand comes from three layers that behave completely differently. Most of the bad promises in the market conflate these layers.
- Parametric
Memory from training
Frozen, months or years old, slow to change — but always present and free. This is where the brand “lives” as a statistical association. You influence it through massive, coherent presence in the corpus.
- Retrieval
Live search, at the moment of the question
Fresh but volatile; it depends on what the engine indexes today (Perplexity, AI Overviews, ChatGPT search). You influence it through fresh, structured, citable content.
- Generation
The final synthesis
Probabilistic, non-deterministic, mediated by each engine's system prompt. This is where non-determinism is born: the same question, different answers.
The golden rule: any AEO intervention must declare which layer it works on. If someone can't say that, they're selling smoke.
Where what we know ends — and what we can measure
We know (with reasonable certainty)
- Co-occurrence defines the brand: what appears next to your name matters more than what you write on your own site.
- Factual, structured, easily extractable content is preferred at retrieval.
- Entity consistency (name, descriptions, schema) helps the model treat you as a stable node.
- Non-determinism is real: a single run is not evidence — you need a sample.
- The engines differ fundamentally: each is tested separately.
We don't know (fundamentally opaque)
- The exact weight of parametric vs retrieval in any given recommendation.
- Which specific source caused a recommendation (causal attribution).
- How long it takes for a change to “catch on” in each engine.
- Whether and how much each engine personalizes the answer.
- How the model behaves after the next update — everything can reset.
We can measure today
- Share of voice across prompts (many runs → a probability, not a rank).
- Citation tracking: which domains the retrieval engines cite.
- Attributes & sentiment: correct vs hallucinated attributes, tone, comparisons.
- Trend over time and differences between engines.
We can't measure
- How many real people saw the recommendation (there are no “impressions”).
- Why it appeared (causality).
- The counterfactual (what would have happened without the intervention).
- The exact attributable traffic (the AI referrer is partial or absent).
Three invariant levels: influence → measurement → report
- N1
Entity foundation
the only “to-do” part · parametric + retrievalStructured, citable content, coherent presence across trusted sources, structured data (schema), entity hygiene: the same name, coherent factual descriptions across every source.
- N2
AEO monitoring
repeatable measurementA fixed set of prompts, multi-engine, repeated sampling (20–50 runs per question), monthly. Output: share of voice + sentiment + citations. The result is a probability, not a rank.
- N3
Honest interpretation
reportingA report that explicitly separates what moved from what we can't attribute. Concrete recommendations, with no promises of position.
We promise
- Better, more coherent presence in the corpus.
- Content optimized for extraction.
- Repeatable, multi-engine monitoring.
- Clear visibility into the trend.
We do NOT promise
- Rank 1.
- Guaranteed results.
- Reproducibility at the individual-query level.
- Immunity to model updates.
AEO is the first instance. It's not the only one.
The same skeleton re-instantiates on any opaque AI terrain. Here are the next candidates from our business surface.
Ads & opaque bidding (Google, Meta)
The auction is a probabilistic black box with partial causality. Layers: audience / creative / bidding algorithm. Levers on the first two; an honest refusal to promise control over the algorithm.
Brand reputation in AI
How models describe you: correct vs hallucinated attributes, sentiment, comparisons with rivals. A sibling of AEO, but focused on how you're described, not whether you appear.
AI-generated content at scale
“AI content = automatically penalized” is false, but opaque. Layers: generation / human editing + voice / platform treatment. Levers on editing and structure; you measure performance, you don't mistake correlation for a penalty.
Competitive monitoring in AI
What models say about rivals — the same polling instrument, a different subject.
“We don't sell you rank 1 in ChatGPT — nobody can honestly guarantee it. We build you a solid entity foundation and give you a repeatable measurement system that tells you, month over month, where you stand in front of each AI engine. The other agencies promise certainties they can't deliver. We give you the truth plus the method.”
Honesty about limits isn't a sales weakness — it's the differentiator.
The framework, applied: our AEO service
Entity foundation + repeatable multi-engine monitoring + honest report. Exactly what the framework above describes, delivered as a service.