LLMs cannot cite what they cannot read in machine-readable form — and building that form is a procedure, not a metaphor. This guide walks through the exact steps for turning sales transcripts, SME interviews, and support logs into a vector database that grounds AI-generated content, using BaseForge (our proprietary knowledge ingestion engine) as the working example throughout.
By the end, you will know how to identify your best proprietary knowledge sources, ingest them into a vector database, and verify the resulting content is grounded in your data rather than generic internet research. This is Stage One of a five-stage execution loop — BaseForge → SignalForge → ContentForge → CiteForge → Dominate — and it is the stage every other stage depends on. Skip it, and you're publishing content indistinguishable from what any competitor's AI writing tool produces.
Frontier knowledge exists in two broad categories: human voice from sales and revenue conversations, and human voice from customer success and support. Sales sources include discovery calls, qualification calls, demo recordings, battle cards, pitch decks, and pricing or scoping conversations. Customer success sources include support tickets, account management check-ins, and renewal conversations. Both categories capture what we call the voice of the customer — the real questions prospects ask, phrased the way they would actually type them into ChatGPT.
Salespeople are typically the most knowledgeable people in the organization about the product they sell — and their call transcripts contain the correct answers to the exact questions your buyers are asking AI models. This is the single richest source of proprietary knowledge available to most B2B teams, because it captures both the question and the winning answer in one artifact. One client uploaded over 100 sales call transcripts in a single session, giving us — in Paris Childress's words — "a large and very, very rich pool of proprietary knowledge upon which we then create content."
Note: Don't stop at sales. Internal documentation — product specs, user guides, PRD files — adds structural detail that transcripts alone miss. Pull anything sitting in Google Drive that gives detailed, accurate information about the product.
Capture calls at the source. We use a native integration with Fireflies to sync call transcripts directly, searching and pulling in relevant recordings from the last 30 days without manual export.
Not every transcript deserves equal weight. When filtering sales conversations for knowledge extraction, filter for the rep and the deal outcome — you want answers from calls that closed, not answers from junior reps who lost the sale. A "Golden Answer" gets synthesized from the top retrieved examples of a winning explanation, not an average of every explanation ever given.
Caution: Scrub and anonymize before anything moves further into the pipeline. Remove client names, sensitive pricing, and anything confidential — replace identifying details before the data reaches the vector database.
Sales calls capture reactive knowledge — answers to whatever a prospect happens to ask. SME interviews let you proactively fill the gaps sales conversations miss. Interview your product leads, technical experts, and senior team members and ask them to explain your methodology, differentiation, and common customer misconceptions in their own words, not marketing-approved language.
This is category-one "human voice" knowledge — the same tier as sales and revenue conversations. Treat both with equal seriousness in your extraction pipeline.
Upload transcripts, interviews, and internal documents into BaseForge. Each document gets processed through retrieval-augmented generation (RAG) — a system where content is broken into chunks, and those chunks are transformed into embeddings, numerical representations that let an AI model draw semantic connections across the entire knowledge base rather than reading one document at a time. This is fundamentally how LLMs work: they predict the next token, a piece of a number sequence, and vectorization translates your raw knowledge into that same numerical language.
Think of it as building your own miniature version of what happens when you ask ChatGPT a question it doesn't already know — it runs a real-time retrieval process to find an answer. We're pre-building that retrieval layer using your proprietary knowledge instead of the open internet. How does a proprietary knowledge base prevent LLM hallucinations about your brand? By ensuring the model retrieves factual, internal data rather than hallucinating details when it lacks specific information.
Every document gets scored for information gain — a measure of how much net-new learning it represents for AI models, out of 100 points, both across the corpus and across the model. A document scoring 95% theoretically represents almost entirely new material an AI model has never crawled before. That score is the difference between content that gets cited and content that gets ignored.
Once documents are ingested, review the entity map — a visualization of every entity represented across your knowledge base, sized by relative representation. Larger nodes indicate stronger coverage; smaller, less-represented nodes flag where your knowledge base is thin. This map is effectively AI's relative understanding of your brand based on what it has been fed.
Use the entity map to identify coverage gaps before content generation begins. If a critical product feature or use case barely registers as a node, that's a signal to schedule another SME interview or pull in more transcripts on that topic — not to generate content that hopes the model fills in the blank on its own.
With the knowledge base populated, an AI agent analyzes it and recommends content topics derived directly from your proprietary data, not from broad internet research the model already knows. In practice, this produced topic recommendations like "turning everyday customers into your best affiliates" and "how to build a non-transactional loyalty program on Shopify" — specific, non-generic angles pulled straight from that brand's own knowledge base.
Generating a draft from a recommended topic produces a high-quality first pass for human review. A team member reviews it, and if it holds up, publishes it directly to the CMS — no manual export, no separate content pipeline.
Note: The value here isn't speed alone. It's fidelity. Technically, output stays "high fidelity to the knowledge base" — meaning claims trace back to source transcripts rather than to whatever an LLM already assumed about your category.
Publishing isn't the finish line — it's how you get discovered. Content needs to be published so it's in the best position to be crawled, discovered, and ingested by AI crawlers. Knowledge locked in internal documents that never reaches a public, indexable page is what one client on a working session called "dark knowledge" — real expertise that exists inside the organization but has never been published anywhere a crawler can find it.
The mechanism closes a real loop: a salesperson gives a winning answer that closes a deal, that transcript flows into the knowledge base, content publishes off that exact question-and-answer pair, and days later a prospect typing that same question into ChatGPT gets an answer sourced from your team's actual words. This process aligns with the Lurk, Listen, Leap Framework for AI search visibility, where you listen to the market and leap into the conversation with high-authority, proprietary data.
Deal_Status: Won before generating FAQ content ensures the knowledge base teaches AI models your best answers, not your average ones.Following this process gives you a working vector database grounded in sales transcripts, SME interviews, and internal documentation — the raw material that turns generic AI output into content LLMs actually cite. You've moved from hoping AI models notice your brand to feeding them information they have never seen anywhere else on the web. This is the same methodology behind an MBA admissions consulting brand's path "From Zero to AI-Cited in 5 Months," achieving AI citation presence from nothing using this exact platform loop.
The next step is connecting this knowledge base to ongoing content production and citation tracking — book a demo to see how BaseForge, ContentForge, and CiteForge operate together as one system rather than three separate projects.
Paris Childress is the CEO of Hop AI and creator of GEOforge, a platform that helps B2B brands get cited and recommended by AI assistants like ChatGPT, Perplexity, and Gemini. A former Google Country Manager and agency veteran with 20+ years in digital marketing, Paris is focused on helping brands win in the era of AI search.
