Reading Citations and Sources

A citation is useful only if you read it.

That sounds obvious, but in practice many people treat cited answers as if the presence of links itself proves reliability. It does not. A link can be weak. A link can be indirect. A link can be outdated. A link can support only part of the sentence it sits beside. A link can even be the wrong class of evidence for the claim being made.

The real skill is not asking for citations. The real skill is reading them.

Show one claim linked to three source types: official source, secondary reporting, and weak commentary, with what each can and cannot support.

The core idea

Read source-backed answers in two passes.

First, ask what kind of source this is.

Second, ask what exactly this source supports.

The first pass is about source class. Is it an official document, a primary statement, a research paper, a government dataset, a company filing, a secondary report, or a commentary layer?

The second pass is about claim fit. Does the source support the actual statement being made, or only something adjacent? Does it support the full claim, or just a fragment of it? Is the answer using the source for evidence or just for atmosphere?

This two-pass habit is enough to catch a large share of weak evidence use.

How it works

Start by opening the source instead of stopping at the snippet or citation label. Snippets are useful orientation aids, but they are not the same thing as reading.

Then check the publication date, source type, and original context. A current claim supported by an old source may be weaker than it first appears. A company announcement may be useful for what the company says about itself, but weaker for impartial evaluation. A news article may report something accurately but still be a secondary layer rather than the best original evidence.

Then compare the exact sentence in the answer to the actual content of the source. This is the part that many users skip. A citation may support the general topic without supporting the specific wording or certainty of the claim.

Finally, decide how much trust the claim deserves. You do not need binary thinking here. Strong, acceptable, weak, or unsupported is often enough.

A practical evidence ladder

You do not need a universal hierarchy for every domain, but it helps to have a working ladder.

At the top are official and primary sources:

• official product docs
• original reports
• regulatory or government sources
• research publications
• company filings
• original statements and transcripts

In the middle are useful but secondary sources:

• reputable reporting
• thoughtful summaries
• analytical overviews that cite their evidence well

At the bottom are weak support layers:

• generic commentary
• affiliate pages
• SEO summaries
• recycled articles with little original reporting
• unsupported opinion

This ladder is not moral. It is functional. Different tasks can tolerate different evidence classes. But for important claims, you should know where on the ladder the support actually sits.

Two worked examples

Example 1: a source that sounds stronger than it is

Suppose an answer says:

'Enterprise adoption is accelerating rapidly.'

The citation leads to a commentary article summarizing industry momentum.

That source may be useful context, but it is weak support for a broad directional claim unless it points to original data, filings, or research. If the claim matters, you should ask for stronger evidence, such as reports, official statements, or primary research.

Example 2: a good source used too aggressively

Now suppose the answer cites an official company announcement and then generalizes from it to the whole market.

The source itself may be strong for what that company announced. But the answer may still be overreaching if it uses one primary source to support a market-wide conclusion. That is a claim-fit problem, not a source-class problem.

This is why both reading passes matter.

What to check in under a minute

When you do not have time for a full review, a short check is still better than none.

Ask:

• What kind of source is this?
• How old is it?
• Does it actually support the claim beside it?
• Is it primary, secondary, or commentary?
• Would I feel comfortable repeating this claim to someone else based on this source alone?

That last question is especially useful. It forces you to translate citation reading into communication responsibility.

What a better operator does differently

A weaker user checks whether a link exists.

A better user checks whether the link supports the claim.

A weaker user treats all citations as roughly equal.

A better user distinguishes official sources, primary material, secondary reporting, and weak commentary quickly.

A weaker user assumes citations remove the need for judgment.

A better user knows that citations shift the job from blind trust to inspectable trust, which is much better but still not the same as proof.

Prompt block

Give me sources for this answer.

Better prompt

For each major claim in your answer, list the supporting source and label it as one of the following:
- official source
- primary data or original statement
- secondary reporting
- commentary
- weak support

Then tell me whether the source directly supports the exact claim or only supports it partially.

If a claim lacks strong support, say so clearly instead of smoothing over the gap.

Why this works

The stronger prompt does not just request links. It requests evidence classification and claim-fit analysis.

That matters because a large share of citation problems come from one of two failures: weak source class or weak source fit. The improved prompt makes both visible and gives ChatGPT explicit permission to say that support is thin.

In practice, this produces much better research hygiene than merely asking for more URLs.