You ask ChatGPT which companies are worth looking at in your category, and your own company isn’t in the list.
Then you type the name in directly — tell me about this company — and out comes a tidy paragraph about what you do and what you’re good at. Accurate, even flattering.
So it knows. It just doesn’t bring you up when it’s the one choosing.
That gap is a genuinely weird feeling, and I’ve had it too. It’s like having a friend at a party who never once introduces you to anyone.
But there’s a decent explanation for it hiding in the model’s internals, and Anthropic published something in July 2026 that gets at it.
The Little Shared Workbench Inside The Model
What the researchers found is a small shared region inside Claude, which they call (the writeup is here).
The idea behind it comes from , which is a way of describing the human brain. Global workspace theory says a huge amount of processing runs in parallel down there, and only a tiny slice of it ever reaches consciousness.
That slice lands on a shared workbench, and from the workbench it gets handed out to everything else. Claude, it turns out, has something that works a lot like that.
Here’s the part that matters for us: J-space isn’t the AI’s visible “thinking out loud”. It’s the quiet thinking the model holds internally without writing down.
How Do You Read A Thought The Model Never Writes Down?
Which raises the obvious problem. If the thinking is silent, how do you watch it from outside?
The team built a new analysis method for exactly that, called (J-lens).
J-lens traces which internal patterns are pushing up the odds of which words, across the model’s layers. Which, for you, means the concepts active in the model at a given moment can be read out as a list of words.
So you can photograph the workbench mid-job and see what’s sitting on it. That’s the trick.
Take The Workbench Away And Only “Thinking” Stops
And what they found is that the workbench is small. Smaller than I’d have guessed:
- Concepts held at once: a few dozen
- Share of the model’s total activity: under 10%
- In some regions, the connection density of J-space patterns runs about 100x normal
So it’s under a tenth of the model, wired up unusually tightly. Then the researchers did the interesting thing and removed it:
- Multi-step reasoning (chaining facts together to reach an answer) dropped to roughly zero
- Fluent writing and simple fact recall came through basically intact
So the model takes different routes depending on the job. Talking smoothly and actually working something out run on separate tracks.
”Knows It” And “Puts It On The Bench” Are Different Things
Translate that into marketing and it starts to bite. What the model has stored somewhere and what it loads onto the bench while assembling an answer aren’t the same set.
Being describable on request and being named during a decision are two different states. I don’t think most of us have been budgeting for that as two things.
There’s a related piece of this in Say The Same Fact Lots Of Ways And ChatGPT Can Recall It, where how a fact is written in training data changes how retrievable it is. Today’s finding sits one step downstream: which of the stored stuff makes it into the answer being built right now.
And in When The AI Explains Why It Picked You, It Might Be Making That Up we looked at models narrating reasons that weren’t the actual reasons. A separate channel of silent thinking gives that one a mechanism.
The Caveats, And What Survives Them
Time to put the fence up, because this is easy to over-read.
This is Anthropic studying Anthropic’s own model. Whether other models have the same structure is outside what this work shows — it’s simply not in the publication’s scope.
Knowing how the workbench operates is also a long way from knowing how to get your information onto it. The work doesn’t hand anyone that lever.
Still, I think the finding earns its keep. Once “has it” and “uses it” come apart, the question of where to measure your visibility gets a lot less fuzzy.
Stop Asking If The AI Knows You, Start Counting How Often It Says You
Whether the models recommend you isn’t settled by whether they know you. It’s settled by whether you make it onto that narrow bench while the answer gets built.
So the thing to check isn’t “does ChatGPT know who we are” — it obviously does, and that tells you almost nothing. It’s “when someone asks for the best options in our category, what fraction of the time does our name come out?”
Track the visible output: mention rate, citation counts. Not the model’s stated reasons for recommending you, which aren’t a record of the decision anyway.
The one thing that plausibly helps in the meantime is boring: keep the fact that you’re a company in this category stated the same consistent way everywhere. Tidy, repeated, easy to pick up — that’s the sort of thing a small bench can hold.
Count the mentions first. Once your measure is a rate, everything you try afterward comes back as a number instead of a hunch.
Sources
- Anthropic, “A global workspace in language models”, July 6 2026, anthropic.com (Reports a small shared workspace inside Claude, “J-space”, analogous to the global workspace described in human cognition. Using a new method, the Jacobian lens (J-lens), which traces across layers which internal activity patterns raise the output probability of which tokens, the concepts active at a given moment can be read out as a word list. J-space holds on the order of a few dozen concepts at once and accounts for less than 10% of the model’s total activity, with connection density in some regions around 100x normal. Ablating J-space collapses multi-step reasoning to near zero while fluent language generation and simple factual recall are largely preserved. J-space corresponds to “silent” thinking distinct from written chain-of-thought. Caveat: this is model-developer research on the developer’s own model, and whether the same structure holds in other companies’ models is outside the scope of this publication.)
