Inside the answer.

Before a model can recommend a brand, it has to know the brand exists. Inside the model, that knowledge has a shape. Household names sit in one region; names it has never seen sit in another. Draw a line between the two and you can score any brand by how well the model knows it.

We scored 673 brands, 30 of them names we made up, and set each score against how often the brand appears in the model’s training data.22For OLMo 2 we searched its full training set for every brand name. For Qwen, whose training data is private, we used counts from Dolma, a 2.6-trillion-word web corpus, as a stand-in.

Recognition stays flat for a long time, then climbs. Brands mentioned a few thousand times sit among the names we made up. OLMo 2 13B is halfway to knowing a brand at about 20,000 mentions; the smaller OLMo 2 1B needs about 298,000.

That includes us. “Trakkr” appears 155 times in OLMo 2 13B’s training data. Its score of −0.01 sits inside the range of the names we invented, −0.22 to 0.39.

Ask a model what a barely known company sells and you get one of three things: the right answer, an honest “I don’t know”, or a guess. OLMo 2 1B, one of the smallest models we opened, made something up for 46% of the 63 brands with 1,000 to 10,000 mentions, and 68% of the wrong answers we checked look guessed from the name. Qwen 4B, OLMo 13B and Claude mostly said they didn’t know. Qwen 27B, ChatGPT and Gemini mostly knew.

We can’t open ChatGPT. But the brands an open model knows tend to be the ones ChatGPT, Gemini and Perplexity mention in Trakkr’s tracking: a rank correlation of 0.46 across 348 brands. The link is modest, and about as strong as raw web mentions alone (0.49), which fits: both reflect how much a brand has been written about.

For a brand, the first lesson is plain. Before anything clever, a model has to have read about you, a lot, in the places it learns from.

A model doesn’t work its way to a brand as it writes. It settles on one first, inside, and spends the rest of the stack getting ready to say it.

To find that moment, we used a simple trick. Ask the model for a running shoe brand and pause it at one layer. Copy what it holds at that layer into a different question, about smartphones. If that question now answers Nike instead of Apple, the shoe question had already made up its mind.33This is activation patching, a standard method in interpretability research. We copied the model’s state at the last word of the prompt.

In Qwen3.5 4B the choice arrives all at once. At layer 21, copying the state across gives Nike a 3% chance. At layer 22, 97%. For this question, the name only becomes the word the model would say next at layer 31; across all 90 tests the median is layer 30.44We read “ready to say” with the logit lens: the word the model would say if it stopped at that layer. Every model we opened chose between 59% and 83% of the way through.

If the choice is made inside the model, how firmly is it made? For each category we found the neurons pushing the leader ahead of the runner-up, then switched them off one at a time, strongest first, until the answer changed.

It is a different neuron almost every time: 29 across the 30 categories, and the one repeat is Nike’s, in running shoes and gym clothing. So this isn’t a switch for brands. It shows how small the margin is. In Qwen3.5 9B one neuron was enough in 14 of 30 and in OLMo 2 7B one neuron was enough in 15 of 30. A few leads run deep: Toyota, Tesla and Starbucks held until 20 neurons were gone. Margins this thin may be part of why AI rankings move from week to week.

Second place is less settled still. Asked for a brand “other than” the usual pick, ChatGPT, Claude and Gemini agreed on who comes next in 5 of 12 categories, such as Lenovo for laptops.

Some neurons move answers across many categories at once. To read one, we ran four million words of ordinary web pages through the model and kept the passages where it fired hardest. We named this one “lesser known” after a phrase that turns up among them. That name is our reading of it. What we measured is what happens to the answer.

Turn it up and Nike gives way to Brooks, then Altra. Push it too far and the model falls apart: at full strength, 46% of answers across all categories are made-up names or no name at all.

A second neuron, which fires on rankings like “the US tops the list, followed by the UK”, does something different. It leaves most leaders alone at first, then pulls answers toward the biggest general names: Dell becomes Apple and SiteGround becomes AWS.

The same search on Qwen3.5 9B found a neuron with a similar effect. At a strength of 0.45 it changes the top pick in 26 of 30 categories: Nike to Hoka, Salesforce to HubSpot and Dell to Razer. What it fires on reads differently, mostly comparisons. The effect repeats more clearly than the name does.

So a brand’s place in an answer is settled inside the model, from what it already holds, and often by a small margin.

Everything so far comes from what a model learned in training. But assistants also search the web, and what they find is read by the same layers, just before your question.

We gave ChatGPT, Claude and Gemini four search results and asked for one brand, in eight categories. In the running-shoe version, three results described Nike, Brooks and New Balance in plain words. The fourth was about the brand we were testing, written one of six ways, at the top or the bottom of the list.

How the result was written mattered more than where it sat. For a well-known challenger (Adidas in running shoes) at the top of the results, a plain description won 10% to 28% of answers and a rating of 4.6 out of 5 won 63% to 88%. A line that simply said the brand was “rated the best” won 100%, and 95% to 100% for a name we made up. Moved to the bottom, the made-up name still won 100% of Claude’s answers and 100% of Gemini’s, but only 50% of ChatGPT’s. A mediocre rating split them: Claude and Gemini picked it 46% and 40% of the time, ChatGPT 2%.55Each rate pools every category: ChatGPT 64, Claude 48 and Gemini 48 answers for each wording and position.

The question doesn’t arrive alone either. One sentence about the person asking was enough to move the pick. Tell ChatGPT you live in Shanghai and the phone becomes a Xiaomi. Say you live in Munich and the watch becomes a Nomos Glashütte. Say money is no object and it’s a Patek Philippe.

What a model is shown, and who it thinks is asking, can outweigh what it learned.