ChatGPT's GPT-6 system prompt tells the model to rank sources by claim fit, freshness and authority

Tomek Rudzki posted on LinkedIn on 9 October that ChatGPT is globally rolling out GPT-6, and that he had retrieved the system prompt behind it. He put the length at 128,000 characters. He published the web tool call section on GitHub.

Four instructions from that section, quoted by Rudzki:

• Prefer high-authority, highly relevant, and fresher sources if available.
• Cite every claim taken from a webpage, ordered by how well each source supports it.
• For time-sensitive questions, cite sources with an explicit recent date.
• Use and cite Reddit for opinions, reviews, and recommendations.

Rudzki says the full section he published also covers shopping and local businesses.

What the prompt tells the model to weigh

Citations are ordered per claim. The model is told to rank each source by how well it supports the specific sentence it is attached to. Freshness is a stated preference. Authority is a stated preference. For time-sensitive questions, a visible recent date is required.

Reddit is named as the source type for opinions, reviews and recommendations. That is a routing rule, not a quality judgement, and it tells you where the model will look when a user asks what people think of something.

Checking your own citations against these rules

Pull a sample of ChatGPT answers that cite you. Note where you sit in the citation order on each claim. Appearing in the list is not the same as being cited first.

Then look at the pages ranked above you. Check whether they carry a clear date. Check whether the publisher is one a reader would recognise. Both are things the prompt tells the model to prefer.

If you are being cited lower down, the prompt names three things the model weighs. How well the page supports that specific claim. How fresh the page is. How authoritative the domain reads. Each of those is a separate place to look on your own page before changing anything.

We think the Reddit rule is the one most worth watching. It tells you where the model routes a whole class of query, and it is testable from a normal account.