NotebookLM vs Perplexity: after Gemini Notebook

The NotebookLM vs Perplexity comparison must now take into account an important change: Google now presents NotebookLM under the name Gemini Notebook. The product remains organised around notebooks and curated sources, but it can also discover sources on the web and launch in-depth research. Perplexity remains centred on web search and the display of citations. Neither of the two removes the need to open the document that supports a decision.

Ask yourself: do you first want to build a controlled corpus, or obtain a web summary directly? The functions now overlap; the useful difference lies mainly in the flow of selecting, validating and keeping sources.

Contents

We distinguish the two starting points, the preparation of a corpus, web discovery, a two-step research chain, confidentiality and the counter-cases.

Two centres of gravity, despite overlapping functions

Gemini Notebook favours work within a notebook of curated sources, whereas Perplexity favours direct web search. Gemini Notebook, however, now offers source discovery, Fast Research and Deep Research: the « closed corpus versus web » opposition is therefore no longer absolute. The first risk remains inheriting an incomplete file; the second, selecting an inadequate or dated source.

The Gemini Notebook documentation describes notebooks, source-grounded answers and citations; its page on adding and discovering sources also documents web search. The Perplexity help centre presents its search modes and its citations. The functions and limits vary with the plans; check the pages applicable to the account tested.

Prepare a usable corpus in Gemini Notebook

A notebook is reliable only within the limits of the sources loaded. A missing document cannot be cleanly guessed, and two contradictory versions must be flagged rather than silently merged.

Clean the list of documents

Name the files, add the date and remove duplicates. For an architecture project, separate the brief, the regulations, the minutes and the specification note. State which version is in force. A file named « final_v2_bis » already prevents good traceability before any AI intervention.

Interrogate the absence

Ask the tool to list the sources used and the points not covered. A useful question is: « which information needed for this conclusion appears in no source? ». Then check the citation in the original document. A reference to the right file may still point to the wrong passage.

Our article on the confidentiality of AI tools provides a classification method applicable to document corpora.

Use Perplexity to discover, not to delegate the proof

Perplexity speeds up the search for pages, but the list of citations remains a starting point. A secondary source can summarise a rule that has changed. An excerpt can lose a condition. A well-ranked page may not be competent on the subject.

Define the admissible sources before the query: institution, publisher, official text, scientific publication or professional body. Open each critical link and look for the exact passage. Note the date, the jurisdiction and the scope. If two sources contradict each other, the contradiction must appear in the note.

The Educasium comparison Perplexity or ChatGPT for professional research details this discovery and validation protocol.

Build a two-step chain

The tools can be complementary if the provenance stays visible. Web search serves to find documents; the notebook then serves to work on a selected set. Mixing results without a label destroys this traceability.

Hypothetical implementation. A team could use Perplexity to draw up an initial list of public documents, verify the sources manually, then import only the retained versions into a NotebookLM notebook. The final summary would cite the original document and its date. This scenario illustrates a possible method; Educasium does not present it as a client deployment.

The manual step between the two stages is a strength when it forces a selection. Automating it without control can import a wrong page faster. Keep a simple table with URL, author, date, reason for selection and verification status.

Confidentiality: the corpus changes the level of risk

A public search and the analysis of an internal file do not present the same risk. Before loading a file, check the account, the processing conditions, the retention and the access rights. Remove the unnecessary information.

The CNIL recalls the principles applicable when personal data is processed. The Educasium guide on the confidentiality of AI tools proposes a practical classification. The fact that a tool is convenient for PDFs does not authorise sending a client file.

Delimited experience: learning to contradict the summary

Educasium trains in the critical reading of outputs; we do not publish a proprietary accuracy rate for these two tools. Our view is that a training exercise must contain an outdated source, a missing piece of information and a contradiction. The learner must detect them.

A demonstration where all the documents are consistent teaches how to produce a summary, not how to secure a decision. The skill sought is the ability to say « the corpus does not allow a conclusion » and to ask for the missing document.

Limits and counter-cases

Gemini Notebook is not, on its own, a governed document base. Its web search functions do not replace the rules of selection, version, access and archiving. Perplexity is not an editorial authority: it aggregates and presents results whose quality varies.

For legal, medical or contractual monitoring, use these tools to explore and organise, then return to the source in force and to the competent person. If the cost of an error is high, the fluency of the summary must never shorten the control.

Frequently asked questions

Can NotebookLM now search the web?

Yes. Under its current name Gemini Notebook, the tool offers source discovery on the web, Fast Research and Deep Research. The results must nonetheless be examined and then selected before import into the notebook. Information found on the web does not automatically acquire the quality of a primary source. Also check the functions available in the plan used.

Is Perplexity reliable because it displays citations?

Citations ease the checking, but do not guarantee that the passage supports exactly the sentence. You must open the page, identify the author and check the date. A correct source can be used out of context. Reliability comes from the full protocol, not from the presence of a link.

Can client documents be loaded into NotebookLM?

Not without prior analysis. Check the plan, the conditions, the settings and the contractual rules of the file. Minimise the data and remove what is not necessary. Ask for internal validation when the documents are confidential or contain personal data.

Which tool suits an architect?

NotebookLM can help interrogate a corpus of authorised documents; Perplexity can help find public sources. In both cases, the architect remains responsible for the reading and the consistency of the file. Test on public or synthetic documents before any project use.

Training a team in verifiable research

Training eligible for OPCO/FIF PL funding subject to conditions. Qualiopi-certified organisation. Discover the introduction to AI and prompting training or contact Educasium to build a research protocol suited to your sources.