AI Learning Hub › Learn AI › Self-Paced Courses › Gemini Notebook: Working with Sources You Choose › When the Source Boundary Changes

Module 5 ยท Intermediate

When the Source Boundary Changes

What happens when the notebook offers to search the web, what it does before asking, what it proposes afterward, and the import decision that is yours to make.

Module 3 ended with the notebook declining to answer and then offering something else: web research. Saying yes to that offer changes what kind of tool you are using, and it is worth seeing the whole sequence before you decide.

What happens when you accept

In controlled tests, accepting started the search straight away. There was no plan to approve, no preview of what would be searched, and no confirmation step between the offer and the work.

A Gemini Notebook conversation showing the user message Yes, please do that web research, followed by a status line reading Analyzing Documentation with an ellipsis, and no other content on screen.
Research begins immediately on acceptance. The status wording varied between runs, so read the state rather than the phrase. Observed on a Stanford account, October 2026. Open full-size image: a notebook showing a research-in-progress status line (opens in a new tab)

The answer that came back was detailed and confident. It also carried no in-line citations at all — a visible difference from every grounded answer in the same notebook, and a useful signal that you are reading something assembled from outside your set.

The part that is a safeguard

Here is the thing worth knowing before you worry: nothing was added to the notebook. The source count did not move. Instead, an import card appeared beneath the answer, listing what the research had found and offering an explicit Import button.

A collapsed import card headed Report and Outside sources (5), containing a single chip reading Research Report, Google Work and a blue Import button. Below it the notebook's question box still shows the counter 3 sources.
The research has finished and the counter still reads three sources. Discovered material waits for a decision rather than joining the corpus. Observed on a Stanford account, October 2026. Open full-size image: an import card offered beneath a research answer while the source count is unchanged (opens in a new tab)

That gate is real and it is in your favor. Web research does not quietly rewrite your source set; it proposes, and you dispose.

The part that is your responsibility

Expanding the card shows what is actually on offer, and everything in it arrives pre-selected.

An expanded import card headed Report and Outside sources (5) with a Select all control ticked. Visible entries, each with its own ticked checkbox, are: University IT Help Desk, What Is Google's Privacy; Google Notebook Center, Gemini Notebook AI R, with a subtitle beginning Official FAQ stating there is currently no recovery op; and NotebookLM Privacy Guide, Is NotebookLM Safe. A counter reads 5 of 5 selected beside an Import button.
Five proposals, all ticked. One is a Google help page, one is a report the model wrote, and the others are third-party sites — including one named Google Notebook Center whose summary calls itself an official FAQ. Observed on a Stanford account, October 2026. Open full-size image: an expanded import card showing five pre-selected sources of mixed origin (opens in a new tab)

Two runs of this sequence produced two different sets of five, so do not learn these particular sources. Learn the shape: a mixture of official documentation, third-party commentary, and a research report the model generated, presented in one list at one visual weight, with every box already ticked.

Names are doing work in that list that they have not earned. Google Notebook Center and University IT Help Desk both read as institutional. Neither is a Google property or a Stanford one. The icons reinforce the impression. Deciding what that material is worth takes a moment of attention, and the interface gives you no reason to spend it.

Exercise: work the import list

Look at the five proposals in the screenshot above. For each one, decide whether you would import it into a notebook you intended to rely on, and write a sentence of justification. Consider, in roughly this order:

A reasonable set of answers

The Google Workspace help page is the strongest candidate: it is primary documentation from the vendor whose behavior is in question. The generated research report is the one to think hardest about, and Module 6 explains why. The three remaining entries need a look at the actual domain before anything else; a name that sounds official is not evidence that it is, and two of these were not what their names implied.

There is no rule here that official is always in and third-party always out. A well-sourced independent analysis can be exactly what a corpus needs. The point is that you decided, rather than accepting a list of five because it arrived with five boxes ticked.

The research answer in this test asserted specific retention periods for notebook content. Those figures could not be confirmed against any official Google or Stanford page, and this course does not repeat them. Retention of ordinary notebook content remains not established from current official documentation. Treat confident numbers from a web-research answer as leads to verify, not as findings.