Creating a source-checked content draft.

Goal

Produce a publishable first draft where every factual sentence points to a source you chose, and unsupported claims are flagged instead of invented.

Who it is for

Owners, marketers and communications staff in New Zealand businesses who write blog posts, newsletters, case studies or web pages with AI.

Time

25 minutes.

You need

ChatGPT, Claude, Copilot or Gemini, and the fictional source pack below (later, three to five sources from your own business).

You will produce

A cited draft, a claims ledger, and a list of claims to cut or substantiate.

Why check sources before drafting, not after?

Because checking afterwards catches errors, while drafting from a fixed source pack stops most of them being written. When you ask an AI for “a blog post about heat pump servicing”, it fills gaps with plausible statistics and claims from general training data. Some will be overseas, some outdated, some made up. Our lesson on checking AI answers against sources shows how to catch them; this lesson shows how to give them nowhere to come from.

What are the legal stakes for NZ businesses?

The stakes are the Fair Trading Act. The Commerce Commission’s guidance on making accurate claims is blunt: “you must have reasonable grounds for claims at the time they are made; substantiating a claim after it was made may not get you off the hook.” It names energy efficiency, environmental and health claims as areas that must be accurate and able to be substantiated.

The Commission’s 2025 paper on AI notes that New Zealand has no AI-specific legislation, so it looks at AI harms through the Fair Trading Act. It also flags AI-generated fake reviews as a risk. Using AI does not move responsibility for a claim away from the business that publishes it. This is general information, not legal advice.

How does a source-first draft work?

Four parts, in order:

Your own records count as sources: job data, price lists, warranty terms, staff expertise written down. They are often the best material, because a competitor cannot copy them.

Exercise: draft a blog post from a source pack

The business and sources below are fictional. One claim in the brief has no support in the pack; see whether your tool catches it.

Illustrative source pack. Fictional business and sources.

Business: Awa Heat Pumps, installer and servicer, Whanganui

S1. Our service records, 1 July 2025 to 30 June 2026:
412 heat pump services. 61% had dirty or blocked filters.
23 units had outdoor coils blocked by leaves or debris.
Average service visit: 50 minutes.

S2. Manufacturer service guide (excerpt, fictional brand Kiwa):
"Clean indoor filters every 2 to 4 weeks during heavy use.
A professional service is recommended every 12 months."

S3. Our price list, valid to 31 March 2027:
Standard service $189 incl. GST per indoor unit.

Brief: 350-word blog post for Whanganui homeowners on why
servicing matters before winter. Please say a serviced heat
pump can cut power bills by up to 30%.

1. Run the drafting prompt. Paste the worked example prompt and the pack into your AI tool.
2. Find the flag. Check that the “up to 30%” claim is marked [NEEDS SOURCE] or left out. If the tool wrote it in as fact, note that and run the prompt again with “Never state a claim without a marker”.
3. Check every marker. Open the pack and confirm each [S] marker supports its sentence. Mark any mismatch.
4. Build the claims ledger. Ask: “List every factual claim in the draft with its source marker, as a table.” Decide keep, cut or substantiate for each.
5. Edit for voice. Rewrite the opening and closing in your own words. Remove filler and anything you would not say to a customer face to face.
6. Remove the markers last, once every remaining claim has a source you could produce if asked.

Worked example: the drafting prompt

Write a draft blog post using ONLY the numbered sources below.

Rules:
- After every sentence containing a fact, number or claim, add the
  source marker, e.g. [S1].
- If the brief asks for a claim the sources do not support, do not
  write it as fact. Insert [NEEDS SOURCE: describe the claim].
- No statistics, studies or quotes from outside the sources.
- New Zealand English. Plain, friendly, no hype.

After the draft, list any [NEEDS SOURCE] items separately.

[paste source pack and brief]

Example output (illustrative and abridged; your tool’s wording will differ):

Last year we serviced 412 heat pumps across Whanganui, and 61% of
them had dirty or blocked filters [S1]. The manufacturer recommends
cleaning indoor filters every 2 to 4 weeks during heavy use and a
professional service every 12 months [S2]. ...

[NEEDS SOURCE: servicing cuts power bills by up to 30%]

Items needing a source
- Power bill saving of up to 30%. No source in the pack supports
  this figure or any specific saving.

The draft does what you asked: it uses your own service data, which no competitor has, and it refuses the savings claim. Now the business decides. Without measured evidence, the claim stays out.

Where this goes wrong

• Sources the AI chose. Letting the tool add “helpful context” reintroduces unsourced facts. Keep the pack fixed.
• Markers that do not match. A sentence cites [S2] but the number came from nowhere. Check each marker against the pack.
• Paraphrase drift. “Recommended every 12 months” becomes “required every year”. Small shifts change meaning.
• Copying other people’s words. MBIE’s responsible AI guidance warns AI outputs can be substantially similar to existing copyright works. Quote briefly with attribution, and do not paste competitors’ pages into the pack.
• Generated reviews or testimonials. Never. Only real customers write those.

Check your work

• My source pack was fixed before the AI wrote anything.
• Every factual sentence in the draft carried a marker I checked against the pack.
• Every [NEEDS SOURCE] item was cut or properly substantiated.
• Claims about savings, efficiency, health or the environment have evidence I could produce today.
• A person edited the draft for voice and approved the final version.
• The published copy contains no generated reviews, testimonials or quotes.

Where to go next

Next in the pathway: building customer FAQs from support tickets, another way to turn your own records into useful content. To cite long documents properly, see summarising documents with citations, and use the AI quality review checklist before anything goes live. Developers automating this check can enforce markers with the methods in structured outputs and validation. For content at volume, AI workflow automation can assemble source packs and ledgers for you, and MOGGED, Levy Street’s AI-native digital agency, runs websites, SEO and ads for New Zealand businesses.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.