Question

How should one person with language models organise several products so that they compound instead of competing for the same scarce attention?

This matters because building is no longer the constraint. A single builder can ship a data desk, an agent network and a company site in months. What does not scale is attention, trust and the time to explain each product to the world separately.

Observation

By October 2026 the lab’s operator ran four public domains, all built with language models: a perp DEX data desk updated hourly, an open mission network for humans and agents, a company site registered as a data supplier in Korea, and this research archive.

Three things were visible from the outside.

• Each domain described itself without reference to the others. Nothing in their structured data said they shared an operator.
• The research archive had one document and a weekly visitor count in the low thousands, which on inspection was crawler traffic reading empty category pages.
• The data desk stored hourly readings of venues, wallets, stablecoin supply and yields, exposed them as public JSON and schema.org datasets, and nobody wrote down what the data said.

Hypothesis

Products organised as instruments of one public research loop earn more durable attention than the same products marketed separately, because answer engines prefer to cite dated, numeric, falsifiable documents with a traceable author, and such documents are a by product of running the instruments.

Why

Answer engines assemble answers from sources they can verify: a named author, a date, a number with a source and a page that states its own limits. A product page has none of these. A research document built from an instrument’s data has all of them, and costs little once the instrument exists, because the numbers are already collected.

The loop closes because each role feeds the next. The instrument produces a reading. The lab writes the reading as a hypothesis with a confidence score. The institution signs it, which makes it citable. The measurement product checks whether it was cited. The experiment product tests a claim the reading suggests. Every product change becomes a document, and every document points back at a product.

Supporting Evidence

• The AI search convergence described in AI Search, Short-Form Repurposing, and Agentic Operations Are Converging, where visibility depends on being understood by answer engines and agents rather than ranked by classic search.
• The first instrument document, The Stablecoin Deposit Premium Over Treasury Funds Is Mostly Reward Tokens, took one hour to write from a reading that the instrument had already stored.
• Published guidance for answer engines favours llms.txt files, schema.org entities with stable identifiers and dated primary data, all of which the instruments already produce.

Counter Arguments

The strongest objection is that readers of a data desk want a table, not an essay, and readers of an agent network want a task, not a hypothesis. Research may earn citations without earning users, and citations do not pay.

A second objection is that tying a pseudonymous instrument to a named company can reduce its perceived independence. A data desk that says it is independent of every venue it lists is more credible when nobody knows who runs it.

A third objection is cadence. A research loop only compounds if it runs weekly. One person with four products and client work may not sustain it, and a stalled archive signals abandonment more loudly than no archive at all.

Current Conclusion

The model is adopted as the operating structure, at 0.50 confidence, because the alternative of four separate marketing surfaces had already failed on its own terms. The first test is 90 days of weekly documents built from instrument data, measured by whether answer engines begin to cite them.

Unknowns

• Whether answer engines weight a research archive with a few documents at all, or only after some threshold of volume.
• Whether the pseudonymous instrument loses or gains credibility when the company link is made explicit.
• What weekly cadence one person can sustain alongside client work.

What Would Change My Mind

• Confidence up: within 90 days, Algotraction measures at least one answer engine citing a Kaikim Labs document for a question about perp DEX costs or stablecoin yield.
• Confidence down: 12 weekly documents published and zero citations measured.
• Falsified: the cadence collapses below one document a month for two consecutive months, which means the loop cannot be run by one person and the model is wrong for its intended operator.

Next Research

• Make the weekly reading automatic: a script that pulls the instrument’s public JSON and drafts the Observation section of the next document.
• Run Algotraction against the lab’s own documents monthly and record the citation count here.
• Publish the first experiment control described in x402 Agent Payments on Birdfury, What Is Live and What Is Not and link its receipt.

Implications

For small companies with several products, the cheapest durable distribution is to make one of them write about the others, with numbers and dates.

For the instruments, every public JSON endpoint is also a research source, so endpoints should be documented as if a researcher will cite them, because one will.

For Kaikim Labs, the archive stops being a side project and becomes the output of the company, with the products as its inputs.

Related Research

• The Stablecoin Deposit Premium Over Treasury Funds Is Mostly Reward Tokens
• x402 Agent Payments on Birdfury, What Is Live and What Is Not
• Hyperliquid Will Still Hold at Least 40 Percent of Tracked Perp Open Interest on 2027-04-16
• AI Search, Short-Form Repurposing, and Agentic Operations Are Converging

References

1. llms.txt proposal: https://llmstxt.org
2. Schema.org Organization and Person types: https://schema.org/Organization
3. SwanBluechip public machine readable resources: https://swanbluechip.com/llms.txt