We make facts easy for AI
to find, cite and compute.
An applied research lab for AI efficiency. We study how AI systems find, read and compute answers, publish what we learn, and build the tools that act on it: Agent C for AI visibility, Croutons for AI spend and serving.
Public signals only. No analytics access.
Trusted by innovative organizations
The thesis
Every AI answer has three costs
Finding the right facts, reading them, and computing the reply. Much of that is waste: models read whole documents to use a few sentences. We call the fix croutonization: reducing what a model has to find, read or compute to the self-contained facts an answer needs.
Find
Be found and cited
Entity salience, co-occurrence and ontology: how machines decide what a page is about, which entity it belongs to, and whether ChatGPT, AI Overviews, Perplexity or Copilot will cite it.
Read
Read less
How many tokens a model has to read to answer correctly, and how to cut them by giving it self-contained facts instead of whole documents.
Compute
Compute less
What it costs to serve each answer: prefill, cache hits and GPU time, measured from vLLM and DCGM metrics on your own stack. For Malaysia’s data centres, we propose reporting energy per answer next to PUE.
Research
What we publish
Method first, data public, and results reported even when they go against us.
Study
ChatGPT cites pages Google does not rank
We compared live ChatGPT answers with Google’s top 10 for the same questions. Only 5.3% of the pages ChatGPT cited ranked there.
Read the study →Studies
Who AI assistants recommend, city by city
Which businesses ChatGPT and Perplexity name for local questions, and what the named businesses have in common.
See the studies →Study
Can co-occurrence teach a model what you are?
Whether repeating a name next to its category, with no links, binds the two in machine memory within 21 days. Protocol published before data collection; zeros published too.
Read the protocol →Study
How many tokens does an answer need?
Full text, retrieval, summaries and croutons on the same questions. Four retrieved passages matched full-text accuracy with 89% fewer tokens; summaries failed.
Read the study →Tools
What we build
The research ships as tools. All of them run from one CLI, @croutons/cli. In development: the Marketing Learning System, which turns what changed on a site and what followed into reusable agent skills.
AI visibility
Agent C
Tracks whether AI assistants mention and cite your brand, and what to fix when they do not.
agentc.croutons.ai →AI spend and serving
Croutons
Breaks documents into self-contained facts and diagnoses token and GPU spend. One command: croutons scan.
croutons.ai →Machine-readable facts
INGEST
Publishes a site’s facts in formats AI systems can read and verify.
ingest.nrlc.ai →Who runs the lab
Joel Maldonado, founder
Joel leads SEO and AI search at a global domain registrar, where he runs the company’s visibility in Google’s AI features and Microsoft Copilot. He designs ontology architecture for AI retrieval, the same architecture behind the Marketing Learning System. NRLC.ai is where he tests what he sees in that work and publishes the methods.
Knowledge base
Notes on AI search
Failure modes, mechanics and measurement gaps in AI-mediated discovery.
Research
How AI systems retrieve and cite
How AI systems retrieve, score and cite content segments: mechanics and failure patterns.
Read section →Diagnostics
Indexed, but never in AI answers
Symptom-first troubleshooting for citation failures and retrieval suppression.
Read section →Measurement
Rankings stable, traffic gone
What can be measured in AI-mediated search, and what to expect.
Read section →Strategy
What still drives AI visibility
What still decides whether you are found, what changed, and how teams should adapt.
Read section →Risk
Brand risk in AI answers
Brand protection and governance when AI systems describe you.
Read section →Tools
What visibility tools can’t see
Where visibility tools and real outcomes disagree in AI-mediated discovery.
Read section →Field notes
Field notes
Observations of AI search behavior, with their limits stated.
Read section →Glossary
Glossary
Plain definitions for generative search and retrieval terms.
Read section →FAQ
Questions
What is NRLC.ai?
An applied research lab for AI efficiency. Every AI answer has three costs: finding the right facts, reading them, and computing the reply. We study those costs, publish our methods, and build tools that cut them: Agent C for AI visibility and Croutons for AI spend and serving. We also take on client work through our services.
What is croutonization?
Reducing what a model has to find, read or compute to the self-contained facts an answer needs. A crouton is one fact that makes sense on its own, with its subject named, so it can be retrieved and cited without the page around it.
Why doesn't AI search cite my content?
AI systems cite sources they can retrieve, extract and verify. Pages are more likely to be cited when the facts are stated plainly, the entity is clear, and other sources agree. AI Search Diagnostics documents common failure patterns.
Is ranking on Google enough for ChatGPT?
No. In our study of 100 ChatGPT answers, only 5.3% of inline citations were pages in Google’s top 10 for the same question.
Why did my traffic drop even though rankings stayed the same?
When AI answers the question on the results page, fewer people click, even at the same rank. AI Search Measurement explains what can and cannot be measured.
Bring us a question about your AI visibility or AI spend.
We take on a small number of client projects: AI search visibility, token spend, and serving diagnostics. Each one starts with a measurement.