TL;DR
On 4 October we ran a pre-start pilot of EXP-003, the test that asks whether our field studies get cited for the market questions they cover. This time the models could search the web. Across 16 answers they cited 58 links. None pointed to backtier.com — not for the field-study questions and not for the two blog-topic questions we used as a comparison.
That is a clear result for this pilot and a weak one for the hypothesis. It doesn't show field studies lose to blog posts. It shows neither is being picked up yet, at least by these two models on this day.
What we ran
Eight questions. Two per published field study: a solo counselor's local visibility in a small Ohio town, AI in cemetery and deathcare records, and a Central Florida insurance agency's visibility in Lakeland. Plus two questions matching our newest blog posts, on entity resolution and the citation gap.
Two models, openai/gpt-6-astra and openai/gpt-5.5, one answer each, through an AI gateway with the model's web search tool turned on. These are API models, not the consumer ChatGPT, Perplexity or Gemini products the formal test names. Every answer and every cited URL is saved word for word on the Experiments page.
We dated it 4 October and labelled it a pre-start pilot because that is what it is. The formal run starts 18 October.
What came back
Google's own help and developer pages were cited more than anything else. For the deathcare questions, the models reached for NIST, the National Archives, HHS and the FTC. For the entity and citation questions, Google and Bing documentation.
One result stands out. Asked which independent agencies serve Orlando, Clearwater and Lakeland for business insurance, both models cited bicfl.com — the agency our BIC field study is about. The study's subject was found. The study itself was not.
Two of the 16 answers, both from gpt-5.5, cited nothing at all.
What this does not prove
It doesn't prove the field studies are invisible to search. All three are recent, and one uses placeholder names, so it wouldn't naturally match a buyer's question. It doesn't prove provenance labelling helps or hurts — with zero citations on both sides, there is nothing to compare.
Our reading, labelled as inference: for broad how-to questions, assistants with search prefer primary sources such as platform documentation and government guidance. A small site's commentary has to answer a narrower question than the official page does, or bring something the official page doesn't have.
What happens next
The formal run on 18 October uses the consumer products, logged out, with the same prompts, so this pilot becomes its baseline. We'll publish those answers with the same labels whatever they show, including another zero.