Custom GPT Trained on My Documents vs an AI Agent That Interviews Me: Which Actually Writes Posts That Sound Like My Company?

The interview wins on specificity. The custom GPT wins on factual recall. Neither one fixes voice.

That is the honest answer, and it frustrates people, because "it doesn't sound like us" is almost always two complaints wearing one coat. One is specificity: the post is true but generic, with nothing in it only your company could have said. The other is tone: the rhythm, the sentence length, the words you actually use. Documents and interviews both feed specificity. Tone comes from a third input entirely, your already-published writing plus an explicit, written description of how you sound. Pick the input that matches the problem you actually have.

What does a custom GPT trained on your documents actually do well?

It recalls facts. Upload your case studies, your pricing rationale, your product one-pagers, and retrieval will pull the right numbers into the right paragraph. That is real value and it is not nothing.

Where it falls down is mechanical. Knowledge files get chunked before they are searched, so the model sees fragments, not documents. Voice is a property of a whole document: how you open, how you turn a corner, how long you let a sentence run. Chunk it and that connective tissue is gone. Worse, semantic search retrieves topically, not stylistically. Ask for a post about your pricing philosophy and it pulls pricing content, not your best-written content. There is no "retrieve by voice similarity" switch, because that is not what retrieval does.

Then there is silent failure. When nothing in your corpus matches well, the model answers from its own general knowledge and does not tell you. The paragraph reverts to default assistant register, and you usually cannot tell which parts were grounded and which were invented filler.

One more trap: most teams upload too much. Twenty files written by three agencies across two repositionings produce an averaged voice, which is mush. Four excellent, current, cleanly formatted documents beat twenty mediocre PDFs with page numbers and broken tables baked into the extracted text.

What does an AI agent that interviews you do differently?

It gets at what was never written down. Documents contain what you already published. An interview gets the positioning nuance you hold in your head, the objection you answer on every sales call, the reason you built it this way.

Clark, our agent, runs an AI Business Interview inside the Knowledge center: one adaptive question at a time about customer stories, differentiation, objections, methods, proof, and point of view. It is resumable, skippable, and repeatable. Each answer saves the moment you advance, reopening drops you back on the exact unanswered question, skipping asks a different angle, and a finished interview reopens as an update round with genuinely new questions informed by what you already said.

The output is not a transcript sitting in a folder. It becomes a short internal knowledge summary that grounds drafts in first-party detail. The cost is your attention, in one sitting or in pieces, and that is the only reason people skip it.

Document upload vs interview, criterion by criterion

• Factual recall: Documents win. Exact figures, dates, and named frameworks live in files, not in conversation.
• Unpublished knowledge: Interview wins outright. Retrieval cannot find what was never written.
• Point of view: Interview wins. A document states what you do; an interview question like "what do founders get wrong before they find you" extracts an argument.
• Setup effort: Documents win. Drag and drop beats sitting still and answering questions.
• Staying current: Documents go stale quietly. An interview that reopens as an update round gets refreshed on purpose.
• Tone and cadence: Neither one. This is the part nobody wants to hear.

So how do you actually fix the tone problem?

Separate it from knowledge and treat it as its own input. We do. Clarity extracts a voice baseline from up to 10 of your already-published posts, and the blog agent takes a free-text brand voice description on top of that. The interview teaches Clark what you know. The voice baseline and the description teach it how you sound. Those are two different mechanisms, and anyone who tells you an interview builds your brand-voice fingerprint is selling you something.

Write that description as constraints, not adjectives. "Punchy and conversational" is not executable. "Never open with a windup, ban the phrase 'in today's landscape', keep paragraphs under four lines, use 'we' for us and 'you' for the reader" is. Negative constraints outperform positive adjectives because a model can check them.

For retrieval to be worth anything, it also has to be real retrieval. Clark merges full-text search with semantic vector search using reciprocal rank fusion, then builds a deduplicated index of your named frameworks, stories, positions, stats, and coined phrases, where the same framework appearing in two books and four keynotes collapses into one asset with multiple sources. A usage ledger records what each post used and marks it as recent, so the next draft rotates to different material instead of retelling the same customer story in a new headline. That last part matters more than people expect: generic is often just repetition.

Which should you pick?

Pick document upload if your material is already written down, current, and consistent, and your complaint is that the AI gets your facts wrong. Curate five to ten canonical files in clean formats and stop there.

Pick the interview if your complaint is that the output is true but forgettable. That is a specificity problem, and specificity lives in your head, not your file folder.

Do both if you want posts a competitor cannot republish with their logo swapped in, which is the real bar. Facts from files, argument from the interview, cadence from your published work. Then check the structure, because a post that sounds like you and still buries the answer will not get quoted by an answer engine either.

And sounding like you is the input, not the result. The result is whether ChatGPT, Claude, Gemini, or Perplexity starts naming you on the question you wrote the post to win. When Clark takes on a content gap it reuses that buyer question word for word, so monitoring measures the exact change the post produced. Voice gets you read. Measurement tells you it paid.

If you are weighing this against hiring someone to write it, we worked through that comparison separately. And if you want to see how both inputs feed the same writer, the overview walks through what Clark does with them once they are in.