What happens when you
ask an LLM a question.

Not everything in the system is a black box. Some parts are private to the AI company, some are shared across all models, and some are things you can directly influence.

Private to the LLM
Universal across AIs
User customization
Business customization
You Ask a question
Shaped by
Platform ChatGPT, Claude, Perplexity, Gemini — each has different defaults baked in by the AI company
User preferences Web access on/off, connected tools, "always search" settings, conversation memory
System prompt Hidden instructions from the platform. You can't see or change these — they're proprietary.
Query type Creative or general → stays internal. Factual, recent, or local → triggers external lookup. Same logic across models.
The model decides Where to get the answer
Training data Proprietary. Scraped from the web, books, code — but what's included, how it's weighted, and when it's updated is controlled entirely by the AI company.
Web search The model searches the web and reads pages — your pages. How your site is structured, what it says, and how clearly it says it directly affects what the model finds.
Retrieval (RAG) You choose what goes in the knowledge base. Uploaded docs, company data, internal files — all under your control.
Tools Code execution, calculators, image generation, APIs. Standard capabilities shared across platforms.
Output Response to you

Specifics

Private

Training data is frozen and proprietary

Models have a knowledge cutoff. What's in the training data, how it's weighted, and when it's refreshed is decided entirely by the AI company. You can't see it or change it.

Private

You can't see the system prompt

Every platform wraps your query in hidden instructions before the model sees it. These control tool use, tone, safety behaviour, and grounding rules. Proprietary to each platform.

Universal

Web search works the same way

When any model "searches," it queries a search engine, gets a handful of results, and reads those pages. The mechanics are the same across ChatGPT, Gemini, Perplexity, and Claude.

Universal

The same model can take different paths

Identical model weights, different infrastructure. GPT-4o in ChatGPT can search. GPT-4o via the API with no tools can't. The model is one piece — the platform decides what it can do.

User

Your settings shape the response

Web access toggles, memory preferences, connected tools, custom instructions — these are levers any user can pull to change how the model behaves before it even reads your query.

Business

How the AI reads your website

When a model searches and lands on your site, it reads your content in real time. Clear structure, specific facts, semantic HTML, schema markup, llms.txt — all things you control that directly affect what the AI extracts.

Business

What goes into RAG

Retrieval-augmented generation pulls from a knowledge base you build. The documents, their structure, and their accuracy are entirely under your control. This is the most direct lever.

In practice

Missed The plumber who doesn't get found
User asks "Find me a plumber near Bondi who does emergency callouts"
Model searches The LLM generates its own search query: "emergency plumber Bondi 24 hour callout." These aren't the terms the business optimized for.
What happens The plumber's site says "residential plumbing services — Eastern Suburbs." No mention of "emergency," "Bondi," or "callout." The site doesn't rank for the LLM's query. It's never even loaded.
Result The business doesn't exist in the answer. Not rejected — just invisible. The model recommends competitors whose pages matched the search terms it chose.
Found The plumber who gets recommended
User asks "Find me a plumber near Bondi who does emergency callouts"
Model decides Factual + local → web search
What it finds Service page lists "24/7 emergency plumbing — Bondi, Bondi Junction, Bronte." Schema markup confirms LocalBusiness type, service area, and hours. Clear, specific, machine-readable.
Result Business is recommended by name, with services and availability cited directly from the page.
Lost The plumber who loses the comparison
User asks "Compare AquaFix Plumbing and DrainPro for a bathroom reno in Bondi — who's better?"
Model searches both The LLM loads both websites to pull pricing, services, reviews, and specialties for a side-by-side comparison.
What happens DrainPro's site is a single page with a phone number and "we do all plumbing." No pricing, no project examples, no reviews. AquaFix lists bathroom renovation packages, past work in Bondi, and has structured review data.
Result The model can only build a case for AquaFix. DrainPro gets a line saying "limited information available" — which reads as a recommendation against them.
← Back to Shelf Engineering