How do you apply the book's 'always verify' rule to ChatGPT search?

A Beginner's Guide to ChatGPT closes its method with one non-negotiable step. After you set a role, give context, describe the task and format, and refine the answer over a few rounds, you check it. The book is blunt about why: ChatGPT can make mistakes and even invent information, so names, numbers, contact details and other facts need a human to confirm them. Its short version is a line worth keeping on a sticky note: "always verify the information's accuracy."

When the book was written, verification meant leaving ChatGPT and searching for the facts yourself. ChatGPT search, launched on October 31, 2024, changes the starting point. Answers now come with inline citations, small links attached to the statements they support, and a Sources button that lists every page used. The work of finding sources is partly done for you. The work of judging them is not.

What changes and what doesn't

Citations solve one problem and leave others in place. Here is the book's checklist, adapted:

The main shift is in the second row. Before search, the risk was a claim with no source at all. Now the risk is quieter: a claim placed next to a real link that doesn't fully support it. A summary can round a number, merge two pages into one sentence, or drop a condition such as "for new customers only."

A worked check of a cited answer

Suppose you are planning a class trip to a local museum and ask:

Search the web: what are the opening hours, group admission price and booking rules for the Riverside Heritage Museum this November? Cite a source for each point.

The museum is a made-up example, and the answer below is illustrative. Imagine ChatGPT replies with three claims, each with a citation:

1. Open Tuesday to Sunday, 10 a.m. to 5 p.m. (cites the museum's website)
2. Groups of 10 or more pay $8 per student. (cites a local news article)
3. Groups must book two weeks ahead. (cites a travel blog)

Now check claim by claim, using the method from the book: verify every fact you will act on.

Claim 1. Open the citation. It is the museum's own hours page, and it matches. One more look: does the page mention holiday closures in November? A summary rarely includes exceptions. Verified, with a note to check holidays.

Claim 2. The citation is a news article from two years ago announcing a new group rate. The page says what the answer says, but it may be out of date. Click the museum's own site and find the current group price. Not verified until the official page confirms it.

Claim 3. The travel blog says "we booked two weeks ahead," which is one visitor's experience, not a rule. Not verified. Call or email the museum.

Result: one of three claims is solid. The answer was useful for knowing what to ask, and the citations pointed to where to look, but two of the facts you would have relied on needed a better source.

Prompting for answers that are easier to verify

The book already advised asking ChatGPT for sources and references. With search, you can make that request more precise. Ask for "a source for each point" so every claim gets its own link. Ask for official or primary pages, such as the organization itself or a government site. Ask for the date of each page. And when a key claim has no citation, ask directly: "Which page says this?" If ChatGPT can't point to one, treat the claim as unverified.

You can also force a search by clicking the web search icon in the message box, which helps when you want current information rather than what the model remembers from its training.

Which checks still matter when answers come with links?

Reading the page still matters most. A citation proves that a page exists and is related to the topic; only opening it shows whether it supports the specific claim. On the desktop website you can hover over a citation to preview it, but for anything you will act on, click through and find the sentence.

Dates and authority still matter. A search returns whatever pages rank well today, which can include old news, an outdated price list or a personal blog. Prefer the organization that owns the fact: the museum for its hours, the government for its rules, the company for its prices. OpenAI's own description of the feature lists news partners and says any website can choose to appear, which means a wide range of sources, not a guarantee of quality.

And numbers, names and contact details still need an official confirmation, exactly as the book said. A phone number in a summary, a price in a table or a deadline in a list is the kind of detail an AI can get slightly wrong in a confident tone. Before you print, send or pay anything based on a search answer, confirm those details on the source that owns them.