Product marketing corpus

​

Overview

The product marketing corpus is where you hand the Merchandising Agent your own marketing material — not a form you fill out, but the actual documents and pages you already use to sell: your media-kit site, template proposal decks, one-sheets, package and sponsorship menus, case studies, audience cards, spec sheets, and seasonal calendars. This is where the agent learns how you merchandise — how you name and package inventory, the story you tell about it, who you say it reaches, and what you point to as proof. From each piece of material come merchandising claims: discrete statements about how you sell, each one recorded with where it came from — lifted from a document you upload, or captured from what you tell Murph about a page you point it at. Confirmed claims can then feed into how the agent composes a proposal, and the corpus checks each one against your actual live inventory so a promise your marketing makes is never pitched to a buyer your catalog can’t back up.

​

What to give it

Anything you already use to market your inventory to buyers counts:

• Your media-kit website
• Template proposal decks and upfront presentations
• One-sheets and sell sheets
• Package or sponsorship menus
• Case studies
• Audience or first-party-data cards
• Spec sheets
• Seasonal or tentpole calendars

Give it both a site and documents if you have both — a media-kit page and a sponsorship one-sheet aren’t competing sources, they’re two pieces of the same corpus.

​

Adding material

There are two ways to add a source, and both go through Murph:

• An upload. Drop a deck, PDF, spreadsheet, or one-sheet into the conversation. Murph parses the file itself, so every claim it proposes comes back with the verbatim sentence and where in the document it sits.
• A URL. Talk your media-kit page through with Murph and record what it says. We don’t fetch that page today — nothing on our side reads it — so what gets stored is your own account of it, which you approve before it is kept. It still needs to be a public https address with no sign-in credentials in the link (https://name:password@… is refused), because the URL is stored as the source of that entry and shown to anyone on your team who reads it. If you want quote-level provenance for that material today, upload the document version of it.

Either way, Murph proposes the merchandising claims and you review them before anything is kept: nothing writes to your corpus on its own. The confirmation shows you every claim in full — for an upload, with the exact quote and location it came from, or the stated reason it has none; for a page, as the words you gave Murph — and you can keep fewer than were proposed, but what gets kept is always the claim exactly as it was shown to you. Nothing can be reworded or re-attributed on its way into your corpus, including by Murph. Only after you confirm does the material and its claims join your corpus, through confirm_product_marketing_source. Add one document at a time. Claims are read for the upload as a whole, so if several documents arrive together nothing can say which one a given claim came from — rather than guess, we refuse the confirmation and ask you to add them one by one. Confirming the same site or document again — say, after you update your deck — doesn’t duplicate your corpus. It lands as a new version of that entry, so your corpus always shows the material you’re currently standing behind while keeping the history of what changed. A site is matched by its URL. An upload is matched by the name you give the material, because every upload lands at a new location: re-upload your refreshed rate card under the same name and it replaces the old one — the previous version’s claims stop feeding proposals immediately. Give a genuinely different document a different name.

​

Where a claim’s provenance comes from

The two ways of adding material can support different kinds of proof, and your corpus records which one each claim has. Your corpus records three kinds, and keeps them distinct:

• Quoted from your document. We parsed the file, so the claim carries the exact sentence and where it sits.
• From your document, no quote available. Also from a file we parsed, but that part of it had nothing quotable — an infographic, a chart — so the claim carries the stated reason instead of a sentence.
• Stated by you. Your own account of a page we did not fetch. That is real provenance — you reviewed and approved it — but it is not a quote anyone read, so it is never dressed up as one, and these claims carry no quote or location at all.

Everywhere a claim appears — your corpus, the gap list, the ingredients that fed a proposal, and the instructions your agent composes from — all three are labelled distinctly, so nobody downstream has to guess which kind of fact they’re reading or treat them as equally proven.

​

What a claim is

A merchandising claim is one thing your own material says about how you package, position, or prove your inventory — kept in your own words, and always recorded with the provenance it actually has: the exact quote and location (a slide number, a page) when it came from a document we parsed, the stated reason when that part of the document had nothing quotable, or your own approved account when it came from a page we did not fetch. See Where a claim’s provenance comes from. So a claim is never a paraphrase you’d have to take on faith, and never presented as better evidenced than it is. Every claim has a kind:

​

Corpus vs. reality

Some claims make a checkable promise — “we run CTV sponsorships,” “our audience card includes an in-market auto signal” — naming a channel, a creative format, or an audience signal by name. For those, the corpus compares the claim against your storefront’s actual live inventory:

• Backed — your catalog currently supports what the claim promises.
• Gap — your catalog does not (yet) support it, named specifically: which channel, format, or signal is missing.
• Not checked — the claim doesn’t name anything checkable (a proof point like “the most trusted name in local news” isn’t a testable promise), or your live inventory couldn’t be read at the time. This is silence, not a finding — it is never shown as if the claim had been cleared.

A gap is a decision, not a verdict against you — your marketing may be ahead of your catalog, or it may be describing something you no longer sell. There are two honest ways to close it:

1. Add the inventory. Bring your catalog up to what your marketing already promises.
2. Retract the claim. If the promise is outdated, withdraw it so it stops being told to buyers.

A claim with a proven gap is automatically held back from shaping proposals until you close it one of those two ways — the agent never repeats a promise your live inventory can’t keep.

​

Undoing a claim

If a claim no longer reflects how you sell — the promise was retired, or you’d rather fix it than keep it — ask Murph to retract it. Withdrawing a claim:

• Stops it from shaping any new proposal immediately.
• Keeps the source material and the claim’s history in your corpus — nothing is deleted, and any past proposal that already used the claim keeps an honest record of that.

You can always re-confirm the same or updated material later to teach the claim again as a new version.

​

Where you’ll see it work

Once you’ve confirmed material, an eligible claim can shape how the Merchandising Agent describes a product to a buyer — never prices, floors, eligibility, or availability, which have their own owners (Playbook and Business Rules). When a proposal used your product marketing, its fed-by detail names “Your marketing” alongside the exact claim that shaped it, so you can always trace a description back to the material that taught it.

​

Related