How AI actually answers questions about your brand
When someone asks ChatGPT, Claude, or Google’s AI a question like “what standing desk has the best warranty?”, the AI usually doesn’t answer from memory. It looks things up first, using a technique the industry calls RAG — retrieval-augmented generation. The name is jargon; the idea is not.
Picture a librarian. Long before anyone asks a question, the librarian takes every page you’ve published — your website, your press releases, your product pages — and cuts it into index cards. Each card is filed by topic in an enormous card catalog.
When the question arrives, the librarian doesn’t reread your whole site. She pulls the three or four cards that seem most relevant and hands only those to the AI. The AI writes its answer from whatever happens to be on those cards — nothing more.
The AI never sees your brochure. It sees a handful of index cards cut from it. The question is: what got cut in half?
Chunking: the cutting of the cards
Chunking is the industry’s word for how your content gets sliced into those cards. And here’s the part every writer should know: the slicing is done by software with a ruler, not by an editor with judgment.
Some systems cut every 500 words or so, wherever that lands — even mid-sentence. Some cut at paragraph breaks. The more careful ones try to cut where the topic shifts. You don’t get to choose which system reads your copy; every AI pipeline picks its own knife.
If your key claim — “lifetime warranty on the frame” — happens to straddle one of those cuts, it ends up half on one card and half on another. The librarian retrieves one card, the AI reads “lifetime warranty on the—” and the answer your prospect gets is vague, wrong, or credits your competitor instead.
Why this is a marketing problem, not an IT problem
A growing share of buying research now happens inside AI assistants instead of ten blue links. When the AI answers, it quotes whichever chunks it retrieved. Copy that chunks cleanly gets three concrete advantages:
- ✓You get quoted accurately. A self-contained claim on a single card arrives intact — numbers, product name, and proof attached.
- ✓You get found more often. A card that carries one complete idea matches a question better than a card holding two half ideas — so it’s retrieved more often.
- ✓Your differentiators survive. The claims that set you apart are exactly the ones you can’t afford to have severed at a chunk boundary.
For PR teams the stakes compound: a press release is written once but ingested by dozens of AI systems, each with its own knife. One badly-positioned sentence can be mangled everywhere at once.
How to write copy that survives the knife
Five habits, no tools required:
- 1.One idea per paragraph. Paragraph breaks are the most respected cut points. If a paragraph carries one complete idea, it becomes one complete card.
- 2.Make key claims self-contained. State the claim, the name, and the proof in a single sentence: “Meridian desks carry a lifetime frame warranty.” Never split it across two sentences that a cut could separate.
- 3.Name names — sparingly with pronouns. A card that says “it also includes…” is an orphan. A card that says “The Meridian desk also includes…” still works when read alone.
- 4.Keep paragraphs short. Under roughly 150–200 words, a paragraph fits inside almost any chunk size, whole.
- 5.Front-load what matters. Cuts hit the middles and ends of long passages far more often than the openings.
Then verify it
See exactly where the machines will cut your copy.
Paste your draft into the bench, and ChunkTool maps every boundary across six real chunking strategies — and flags any key phrase that doesn’t survive.
Open the bench