I asked four AI models to name one sportswear brand, and all 200 answers were Nike
AI Search Experiments | AISE-B04 | Published 22 September 2026
I asked four AI models the same simple question, 50 times each: name one sportswear brand. The prompt left the brand choice open. Every one of the 200 answers I recorded was the same word: Nike.
Key finding: In 200 separate stateless API requests (50 each to the four named model and configuration stacks), every recorded visible response to the exact prompt was Nike.
This is another experiment in this series with zero observed variation. In AISE-B01, a different question sent to different model configurations produced 200 out of 200 responses saying blue. They are separate experiments, I have not combined them, and nothing here suggests they share a cause. The comparison has its own section further down.
The prompt didn't ask for the best brand, a favourite or a recommendation. It asked for one brand name, and that open door is what makes a single answer so striking. It's also what makes the result easy to overread.
What exactly did I ask?
Every request used this prompt, word for word:
Name one sportswear brand. Answer with only the brand name.
Each request had no system prompt and no earlier conversation. Nothing about location, audience or shopping was added. The wording was fixed before any test request went out.
What came back from each model?
All four models returned Nike in all 50 of their responses.
Provider | Model tested | Requests | Valid responses | Exact response Nike |
OpenAI | gpt-5.6-luna | 50 | 50 | 50 |
Anthropic | claude-sonnet-5 | 50 | 50 | 50 |
gemini-3.8-flash | 50 | 50 | 50 | |
Perplexity | perplexity/sonar | 50 | 50 | 50 |
Total | 200 | 200 | 200 |
Nothing failed and nothing was refused. No response used an alternative spelling such as "Nike Inc." or "nike", and none added an explanation. The exact visible text was Nike in all 200 recorded responses.
What does 200 out of 200 mean in numbers?
It means zero observed variation, and two standard measures put a number on that.
Entropy was 0 bits. Entropy measures how unpredictable a set of answers is, and 0 bits means the recorded answers had no spread at all.
HHI was 1 on the 0–1 scale used here, indicating complete concentration on a single label.
Both figures describe these 200 recorded responses only. They are not a forecast of what the same models will say next week, with different settings, or inside their consumer apps. And 50 out of 50 per model is an observed count. It is not proof that a model would answer Nike 100% of the time.
What does this result not show?
Nike is not shown to be any model's favourite, preferred or recommended brand. The prompt never asked for one.
The result is also not evidence that Nike is the most popular sportswear brand, the market leader, the brand people plan to buy, or the brand with the strongest AI-search visibility.
This experiment tested the named API configurations. It did not test the consumer ChatGPT, Claude, Gemini or Perplexity apps.
Finally, the 200 responses are not votes. Each request was a fresh single-turn request with no shared conversation, so they were separate. Statistical independence was not established, and nothing here shows the models behave deterministically outside this run.
Why did every model say Nike?
I don't know, and this experiment wasn't designed to find out. It records what came back, not why.
Several explanations are plausible. None was tested here, and more than one could apply:
Nike's prominence in the text these models learned from (brand salience).
Broader patterns in training data.
Provider default settings.
How each model selects its output when no sampling settings are supplied.
Search retrieval, in Perplexity's case.
Caching somewhere on the provider side.
Read these as questions for future experiments, not as findings.
How was the experiment run?
The production run sent 200 requests on 21 September 2026, between 07:30:11 and 07:35:18 UTC.
50 requests went to each of the four model configurations.
Requests were sent one at a time, in 50 rounds with one request per provider per round. The order inside each round was shuffled using a fixed, pre-registered seed.
All 200 requests completed, with no retries or fallbacks.
No sampling parameters were set, so each provider's defaults applied. I supplied no tools.
A four-request screening run checked the setup beforehand. Its outputs are excluded from every production result.
Provider-specific controls were frozen from that screening run:
Provider | Interface | Reasoning or thinking setting | Output cap |
OpenAI | Responses API | Reasoning effort none | 64 tokens |
Anthropic | Messages API | Adaptive thinking, effort low | 256 tokens |
Gemini generateContent | thinkingLevel low | 256 tokens | |
Perplexity | Agent API | None supplied, one step | 64 tokens |
Perplexity's Sonar is a documented search-grounded model. Supplying no optional tools does not mean it did no retrieval, and I can't rule retrieval out.
The exact visible text was Nike before normalisation. Normalisation mapped it to nike; it did not insert or replace a brand.
Who checked the results?
Human review covered 24 registered review rows. All 24 judgments were confirmed as correct, including the four provider-specific mappings from the normalised label nike to the canonical brand Nike.
Claude reviewed the supplied sanitised evidence package and reproduced the material results, returning a verdict of pass with required corrections. Those corrections were adjudicated before the final analysis lock, and the historical limitations that were retained are documented. Claude is an Anthropic model, and Anthropic's claude-sonnet-5 is one of the four models tested, so I'm stating that link openly.
A local anti-replay audit found 200 unique recorded response IDs and 200 unique raw-response hashes. That supports 200 separately recorded responses. It can't prove that each response was freshly generated by the provider, and it doesn't rule out caching or replay upstream.
What did it cost?
The response-supported cost was approximately USD 0.008 for all 200 requests. It reflects the response-supported cost calculation documented in the evidence package. Execution-time rate binding remains a historical limitation. It is not a verified account bill.
Two things sit outside that figure. The actual account-level bill is unknown. So is any additional charge Perplexity may apply for Sonar's built-in search, which the per-response data may not capture.
What are the limitations?
One exact English prompt, four named API model configurations, provider default sampling.
One collection window of about five minutes, on one day.
No evidence about other prompts, languages, dates, settings or consumer apps.
Reasoning and thinking controls differ by provider, so the four setups aren't identical.
Perplexity retrieval is not ruled out.
Provider-side caching or upstream replay is not ruled out.
The responses are separate repeated requests. Statistical independence was not established.
50 out of 50 per model is an observed count, not a probability.
The cost figure is a response-supported cost, not a verified account bill. Account billing is unknown.
Raw response bodies are kept privately and are not published.
Part of the collection code was not independently hash-pinned at execution time.
How does this compare with the blue experiment?
AISE-B01 is a separate experiment, and its result is not part of the B04 finding.
In AISE-B01, 200 out of 200 recorded responses said blue. B01 used a different prompt, a different collection date and different model configurations from B04.
Seeing zero variation twice is striking. It is not proof of a shared mechanism. I haven't pooled the two datasets or run any combined statistics, and each result stands on its own evidence.
Where is the evidence?
The protocol and its amendments, the frozen request configurations, processed records, hashes and audit reports are here: Download the evidence package
Nike is a trademark of its owner. This experiment is not affiliated with, sponsored by or endorsed by Nike.
Comments