Giant Context builds your marketing site with a model, and it writes without being asked. That makes where the content comes from the whole question.

A model handed nothing but a request invents. It writes a warm testimonial no customer gave, a user count from nowhere, a feature the product does not have. That leaves you with a site full of claims that were never true.

You bring your own material, your documents and specs and pricing and the record of the business. The site is built from that, and every claim on the page traces back to something you provided.

The fix is to give the model something real to build from, and that material is your own files. Your business has a corpus of them, product docs, pricing sheets, case studies, the team page, the specs, the record of what it does and who it is. Once you upload that corpus it becomes the ground the site is built on. Before the model plans a section, the grounding pipeline reads the files and does two things with them. Its own description states both:

The first conditions the writing. It pulls the file content most relevant to the section at hand and puts it in front of the model as real project data, so the model writes from your material rather than from nowhere. The second verifies. Every concrete fact the model proposes, a testimonial, a statistic, a named team member, has to anchor to file content above a similarity threshold. Anything that does not is dropped. A claim reaches the page only if it came from your files.

A law firm uploads three things, a page of practice areas, two closed-case write-ups and a rate sheet. The site that comes back lists those practice areas, quotes the two cases as proof of work, and shows the rates from the sheet. It invents no fourth practice area to balance the grid and manufactures no client quote to fill the testimonial row, because neither is anywhere in the files. The page is shorter than a fabricated one would be.

Not everything in your files is a fact to state. You upload your own pricing. You also upload a comparison against a competitor. The first is yours to claim. The second is background, useful for reading the market and not a price you charge.

So every file is tagged when you upload it, fact or context. The verification pass will not let a context chunk anchor a claim, so the competitor's prices never surface as yours. You set the tag at the moment you hand over the file. The model never guesses which prices are yours.