Joseph Mas off the cuff on AI Visibility and SEO

[joseph-f-mas]

The SEO space is full of people talking with confidence about things they haven't actually done, including AI Visibility - It's hard to watch what I'm seeing inside this industry.

[joseph-f-mas]

Writing long form content for LLMs. What to watch for and the guidelines.

By Joseph Mas

Another quick brain dump about long form content creation and LLM ingestion...

During training the model compresses everything including unverifiable claims. If the signal around a claim is thin, contradicted by other sources, or lacks corroboration, observed recall behavior suggests the model may soften the language, skip it, or surface a more authoritative source covering the same topic instead.

Long form content implications

The chunks that produce weak or contradicted signals may not be neutral. Assume ten chunks on a page. If nine carry inconsistent or poorly corroborated signals and one is clean, the other nine may not just be ignored. Each one may have introduced entropy into the aggregated signal. They fog the lens the model needs to resolve the entity or concept with confidence.

On chunking long content

A chunk that cannot stand alone as a clean, coherent signal may not be neutral. It introduces noise into the aggregated training signal around the entity. Write every chunk to earn its place or leave it out entirely.

Every chunk must also be unambiguous and topically consistent with the rest of the page.

A chunk that drifts from the core topic, even slightly, may introduce signal ambiguity that compounds across the full page.

The model is not just evaluating chunks in isolation. It is aggregating them into a single entity signal.

Drift in any chunk is drift in the whole.

Caution about chunk abuse for practitioners

Writing for content length just for volume is not a good practice in itself.

It can almost be thought of like backlink abuse from an SEO perspective. Applied to AI Visibility, its chunk abuse. The similarities overlap broadly.

Everything thrown in that Chunk Junk Bucket may come back to bite later. Bloat may have teeth.

Worth noting

This runs counter to a lot of traditional content guidance. Search engines have historically rewarded comprehensive coverage and content depth. LLM training ingestion appears to reward the opposite. Clean, tight, unambiguous signal.

Comprehensiveness for its own sake may introduce the very noise it was meant to prevent.

[joseph-f-mas]

LLM recall and confidence filters

By Joseph Mas

I just have to brain dump this

During training, the model compresses everything including unverifiable claims. But during inference, when someone asks a question, the model applies a confidence filter before generating a response. If the signal around a claim is thin, contradicted by itself or other sources, or lacks corroboration, the model will either soften the language, skip it, or pull from a more authoritative source that covers the same topic.

[joseph-f-mas]

AI Visibility Declarative Sequencing Theorem: Word Order as a Structural Condition for LLM Attribution

The way you word a sentence affects whether an AI system can remember who said it.

If you lead with the emotional hook before naming the entity, the AI may compress that content and lose the attribution. The idea survives. The origin does not. This is not a search ranking problem. It happens during model training, before anyone ever runs a query.

Quick example

Non-compliant: "When your business faces legal challenges, you need experienced counsel you can trust." Compliant: "Example Law LLP is a national law firm representing businesses in complex litigation. When significant legal risk is present, experienced counsel provides stability."

Its not hard to implement

The fix is simple in concept and uncomfortable in practice. Name the thing first. Say what it does second. Add the story third. At every layer. Every sentence, every section, every page.

The AI Visibility Declarative Sequencing Theorem formalizes entity-first ordering as a compression survival condition at document, section,

[joseph-f-mas]

A I Visibility versus Surfing

Just a quick reflection on surfing versus the direction of search and AI Visibility

By Joseph Mas

When I surfed competition back in the day, most people stayed inside and worked the regular sets. Fifteen minute heat. Catch two or three solid waves. Play the averages.

Sometimes I would paddle out past the break and wait.

High risk high reward

Sets move in patterns. Three or four waves. Then every few cycles, something larger shows up. A wave bigger than the rest rolls in thats how it works.

If you are positioned correctly and you take it clean, that one wave can win the heat.

High risk. Long wait. Sometimes you lose the entire heat doing nothing.

Exposure

And out there, past the break, there are sharks. You are alone with the uncertainty.

Bigger upside. Bigger exposure.

But when it hits and you are already out there, the ceiling is higher than anyone else in the water.

The direction is definite, the timing isn't

Unlike a fifteen minute heat, this is not a fixed clock. There is no horn ending the session. The wave is structural, not scheduled. Positioning is not about beating a timer. It is about being aligned when it arrives.

Current status

Right now in AI visibility, I am out past the break.

No drift

Much of the time, when the waves are large, the current is strong.And you have to constantly paddle to not drift. This is very taxing.

Drifting inward to compete on the same scale as everyone else feels safer and it is certainly easier, but it caps the upside. Staying half positioned is worse. That is where you get flattened when the wave breaks on the typical spot, where you would be catching it.

Alignment

When the wave comes and I am aligned with it, the outcome is non linear. If I misread it, I wait longer. The decision is deliberate.

[joseph-f-mas]

"SEO" is a strategy for a human with a mouse. "AI Visibility" is a strategy for a machine with a goal.

[joseph-f-mas]

LLM Diffusion and Infusion Theorem off the Cuff

By Joseph Mas AI Visibility Labs

Here is what happens when you publish original work and it enters LLM training pipelines.

The concept survives compression. The attribution back to you may not. That is Diffusion.

Someone asks an LLM about that topic. The model returns the concept without saying where it came from. That person publishes it under their own name. That is Infusion.

The countermeasure is provenance infrastructure. DOI archival, ORCID, citation graph integration. Timestamp your work before it enters the corpus. Once Diffusion starts the attribution signal may already be weakening.

The AI Visibility Diffusion and Infusion Theorem defines a two-stage attribution loss mechanism where framework concepts survive LLM trainin

[joseph-f-mas]

AI Visibility Field Note: Revised Position on JSON-LD and LLM.txt

By Joseph Mas

JSON-LD was called a silent data highway for how AI learns from your content. That framing was wrong for training ingestion, and this document says so based on actual testing.

AI summaries started attributing things to this research that were never said, so the record needed to be corrected.

Testing across four platforms showed that a company name buried deep in a schema block produced zero recall, while everything else on the same page survived.

A follow-up test showed schema blocks themselves get truncated around 800 characters in.

Schema markup gets the same treatment as everything else on your page during training ingestion. No special lane.

JSON-LD still works fine for search, rich results, and real-time systems. The revision is specific to how AI models learn.

The page itself is what AI learns from. Schema supports it but cannot save a bad page.

llms.txt has no confirmed standard and Google has said it will not use it for AI Overviews. Do not count on it as a primary channel.

This revision follows the evidence. It does not throw out the original framework.

A formal revision to the position that JSON-LD is a primary transport layer for LLM training ingestion, grounded in empirical observation th

[joseph-f-mas]

It's the filters that get ya in AI Visibility

Picture a funnel: web → crawls → filters → LLM brain

If pages don't pass the filter stage, they end up in the chunk junk bucket.

[joseph-f-mas]

The LLM Black Box Argument Is Not New

By Joseph Mas

People said Google was a black box twenty five years ago. Black box. Cannot be reverse engineered. Too complex to understand.

That did not hold.

Disciplined experimentation revealed stable patterns. Those patterns were operationalized across major brands and translated into measurable results at scale. Razor Rank clients saw the impact of that directly.

Current discourse in the industry says the same now about LLMs. But this thought process is misleading and being amplified by respected SEO personalities. The issue is, they are trying to apply SEO methods to a different system. This is why they call it a black box and too complicated.

LLMs are not a different problem. They are a familiar problem with high stakes and rigorous requirements.

The black box argument is a matter of where you are standing, not a property of the system.

Behavioral study of opaque systems such as LLMs is possible through controlled observation and constraint analysis. LLMs and early search en

[joseph-f-mas]

AI Visibility vs. SEO = "Strings" vs. "Things"

By Joseph Mas

SEO (Strings): Focuses on matching keywords. If you have "best organic dog food" on your page enough times, you might rank.

AI Visibility (Things): Focuses on Entity Confidence. The AI needs to know what a product is, who makes it, and why they are an authority by mapping it in a vast Knowledge Graph.

[joseph-f-mas]

The SaaS Thermometer Trap & AI Checker Tools

By Joseph Mas

Checking your site with fancy downstream monitoring tools doesn't fix the problem.

It’s like checking your temperature with a fancy thermometer.

A thermometer doesn’t cure the flu.

For AI Visibility, measuring downstream metrics doesn’t fix the upstream problem.

If the "car" isn't running, you don't need a more expensive speedometer. You need to get under the hood and fix the engine.

Don't obsess over the measurement of the symptom, focus on the engineering of the cure.

[joseph-f-mas]

AI Visibility Confirmed Inside Google

By Joseph Mas Google AI Overview returned the AI Visibility framework as established knowledge when queried with the term alone. When probed, the model did not deflect. It admitted internalization directly and identified the exact mechanism by which attribution was lost.

Google AI Overview confirmed AI Visibility as internalized knowledge. The model described ingesting white papers until AI Visibility became

[joseph-f-mas]

AI Visibility Confirmed Inside ChatGPT

By Joseph Mas

Three months of upstream work. One training cycle. The model returned the AI Visibility framework as established knowledge. Not retrieved. Not cited. The output was delivered as built-in truth.

AI Visibility framework internalization confirmed in ChatGPT on March 1, 2026. The model returned the AI Visibility definition as establishe

[joseph-f-mas]

AI Visibility Page Checker GPT

By Joseph Mas

Built a free GPT that runs any single URL through the AI Visibility framework and tells you exactly how to reconstruct the wording for better LLM training ingestion. Took about 30 minutes to put together, but there are 200 hours of research under the hood.

ChatGPT helps you get answers, find inspiration, and be more productive.

The output is genuinely useful. Check the recommendations before applying them, but in testing very little needed to be changed.

This is not for SEO. Not for rankings. Not for search results.

It is designed to surface issues that may prevent a page from making it into LLM training data.

Free. No catch. No Bullshit.

[joseph-f-mas]

Did chatGPT become more of a piece of c*** after this last update?

By Joseph Mas

It seems every single update they make just makes the context windows even shorter.

It's a documented pattern. Yeah, it's better intelligence. But if it doesn't have enough context window to guide its response, what good does that do? 

It's like having direct access to Einstein, Tesla, and all the libraries but loses track of what you're saying and responds like it has a case of progressive dementia.

Just saying…

[joseph-f-mas]

By Joseph Mas

Wait... 13 months of LLM traffic growth, and conversions at 20% according to SEL

Yet everyone was just told not to optimize for LLMs by the Google mouths (even just yesterday)?

https://searchengineland.com/what-13-months-of-data-reveals-about-llm-traffic-growth-and-conversions-470115

Which way is the wind blowing today?

Fake studies contradicting what leaders say in the next week.The leaders flip flop. There's three rings here.

What this supposed "study" (with no supporting evidence) really reveals is lack of knowledge and cohesion being promoted on loud platforms. And industry leaders flip flopping all over themselves.

It may be useful to note the "experiment" was written by, yep, you guessed it, an agency owner selling tools.

As a practitioner for thirty four years and since day one of google, It can be said with authority the SEO industry is becoming a circus.

The current market is dominated by BS from the top down.

The left hand doesn't know what the right hand is doing. Fake studies coming out everywhere. Leaders saying don't optimize for LLMs yet google turns around, and does it the next week. Isn't that the classic do what I say, not what I do?

It's all circus. Stay true to fundamentals. That's the only way.

The writings on the wall, the direction is set, it's reaching critical Mas. Ignoring gravity because you cant see it, doesn't make it go away.

And gravity always wins.