Choosing the Right AI Tool · Article Two
Suppose you want to investigate why Australian adults return to learning the piano. You may want evidence about the size of the market, motivations for study, lesson formats and the reasons people stop. Give the question to a deep-research tool and, within minutes or perhaps a little longer, a substantial report with many citations appears.
The experience encourages an understandable illusion: research that once required several days has been completed. A more accurate description is that a large portion of searching, selecting and preliminary synthesis has been delegated. The report may reveal material that would otherwise have been missed. It may also place an industry forecast and an official statistic in the same paragraph without adequately distinguishing their evidentiary status. Numerous citations do not mean that the research question has received a dependable answer.
A useful comparison of deep research in ChatGPT, Gemini and Perplexity therefore cannot rest on report length or citation count. We need to see whether the user can define the source boundary before research begins, steer the work while it runs, inspect the basis of the result and move the report into the next stage of work.
Similar assignments, different starting points
All three products now go well beyond one search followed by a summary of a few pages. They can break a question into parts, make successive searches, follow new leads and compile a longer report. The visible similarities are substantial. The differences lie mainly in how sources enter and where the result goes afterwards.
| Tool | Main controls before research | Visibility while it runs | Where the report goes next |
|---|---|---|---|
| ChatGPT Deep Research | Public web, uploaded files and connected apps; nominated sites can be required or prioritised, and the plan can be edited | Progress is visible; the user can interrupt and adjust the focus or sources | A cited report with activity history, downloadable as Markdown, Word or PDF |
| Gemini Deep Research | Google Search by default; Gmail, Drive, uploaded files and Gemini Notebook can be added or used instead; the plan can be edited | The completed research is available through the conversation and Canvas | Content can be copied, shared or exported to Google Docs, with audio and visual derivatives available in supported circumstances |
| Perplexity Research | Centred on the public web and able to process uploaded documents; the Research mode selects its model automatically | Research progress and interim findings are shown, and further questions can be added during the run | Reports can be edited and shared, then exported as a document, PDF or Perplexity Page |
The table describes documented product design. It does not demonstrate that one system finds better sources for every subject. Results also depend on the question, the accessibility of relevant pages, language, region, specialist depth and the models available at that moment.
When you already know where the evidence ought to be
Some research is not an open search for the largest possible collection of webpages. Law, public policy, product specifications and institutional history often have a recognisable source hierarchy. A study of Australian unemployment definitions should begin with Australian Bureau of Statistics methodology, not allow the number of explanatory blogs to determine the answer. Official documentation and release notes are usually closer to the facts of a software feature than a comparison site.
ChatGPT Deep Research currently allows the user to nominate websites, either restricting the task to those sites or prioritising them while retaining broader web search. Its research plan can be edited before commencement, and the running task can be interrupted and redirected. This design suits a common situation: the question is complex, but the researcher already has an initial view of the evidence hierarchy. The system is being asked to read more widely without being allowed to redefine the source boundary by itself.
That is not a claim that ChatGPT is inherently more accurate. Source restrictions introduce their own blind spots. If the selected institutions omit an important controversy, a restricted search will reproduce the user’s selection bias efficiently. A safer design is often a two-stage search. The first pass stays within primary sources and establishes institutional facts. A second, broader pass looks for criticism, counterexamples and research that the official record may not include. The product supplies a control; the researcher remains responsible for using it intelligently.
Completed reports can be downloaded as Markdown, Word or PDF. For people whose next stage takes place in local files, version control or formal editing, those formats remove some handling. They do not make a report publishable, but they create a clearer object for later verification.
When the material already lives in Google’s world
Gemini Deep Research uses Google Search by default. A user can also include personal Gmail, Drive, uploaded files and Gemini Notebook, or deselect Google Search when public web material is not wanted. Gemini proposes a research plan that can be revised before the run begins.
The value of this design easily disappears in a general leaderboard. Imagine preparing a project retrospective whose evidence is dispersed across Drive documents, an email history and an existing notebook. The difficulty is not discovering more public pages. It is placing an internal chronology beside changes in the outside world. A tool that can reach those sources within the user’s permissions may eliminate repeated downloading, uploading and explanations of context.
The report can be exported to Google Docs. That may look like an ordinary export button. For a team already using Docs to comment and revise, it determines whether the report becomes a working document or remains a chat output. Audio overviews and visualisations can help with learning and presentation, but they do not replace source checking. The easier a chart is to generate, the more carefully its units and data definitions need to be examined.
Ecosystem convenience also has boundaries. Personal Google AI plans and organisational Workspace arrangements do not always provide the same entitlements. Source connections, administration and data conditions may depend on the account. The practical question is not whether a marketing page displays a feature, but whether a particular account, region and administrator configuration actually permits it.
When research begins as rapid exploration of the public web
Perplexity has long placed search at the centre of its product. Its Research mode performs repeated searches, reads a large set of sources and assembles a report. The 2026 update called Advanced Deep Research added document processing, a code sandbox, a progress display, questions during the research process and editable reports. In Research mode the system selects the models it needs; the user does not nominate a particular model.
When the starting point is “what reliable material is available on the web now?” and the question is likely to change quickly, Perplexity’s product structure is a natural fit. Ordinary search, follow-up questions, Research and Spaces occupy neighbouring parts of the experience. A user can begin with lighter searches to discover the vocabulary and principal sources of a field, then assign the genuinely complicated question to Research. The resulting report can also become a shareable page.
Search-centred design does not make sources reliable by nature. Search systems are good at locating accessible, indexable material. The most important source may be inside a database, an attachment, a scanned record, a paid service or a page that resists extraction. Several apparently independent webpages may repeat a claim from the same unverified origin. Citations and a progress display create routes for inspection; they do not remove the need to open the sources that carry the conclusion.
Speed must also be kept in its proper place. Perplexity says that many Research tasks produce a report within several minutes. Rapid iteration is useful. It does not establish that a slower product is necessarily deeper, or that the faster result meets publication standards. Research time ends when the important sources have been checked, conflicts have been addressed and the conclusion can support its intended use—not when the first report appears.
A citation can look correct without supporting the sentence
The most common mistake in using deep research is treating a link at the end of a sentence as proof that the sentence has been established. Several failures are possible.
The linked page may discuss the same subject without supporting the stated number. It may report correlation while the generated report writes in causal language. A webpage may cite another report whose method and sample are absent from the current page. The original source may contain a regional or temporal limitation that disappears during synthesis.
Verification need not apply equal effort to every line. Identify the load-bearing claims first: if this statement is wrong, does the conclusion change? Market size, legal thresholds, risk rates and product prices usually deserve priority. Open the original material and check units, date, sample and qualifications. Then sample some secondary claims to see whether the report shows a recurring tendency to exaggerate. If one problem is found, examine similar claims rather than repairing only that sentence.
A research product may provide a long bibliography while relying on only part of it in the report. Source count should not become a quality score. A better report lets the reader see which evidence carries which judgement and preserves questions for which dependable evidence was not found.
Three kinds of work, three plausible starting points
When a subject has a clear hierarchy of authoritative sources and you want to control websites, edit the plan and carry the report into local or Word-based editing, ChatGPT Deep Research is a sensible first tool to try. This recommendation follows from the workflow design; it is not a general claim about accuracy across every subject.
When the decisive material already resides in Gmail, Drive, Docs or Gemini Notebook, and the research needs to return to Google Docs for collaboration, Gemini’s reduction of material handling may be more valuable. Account type and actual regional availability should be confirmed before subscribing for that purpose.
When the main activity is mapping a fast-changing public-web field and repeatedly adjusting the focus, Perplexity Research’s search-centred environment is a natural starting point. If the result will support publication or a high-consequence decision, the essential sources should still be saved and verified separately.
Many tasks should not begin with deep research at all. A single well-defined fact is often faster to locate with ordinary search. If the entire assignment concerns three nominated documents, a source-grounded Notebook or Project may fit better. When the question itself remains unclear, a preliminary conversation can prevent an automated system from producing a polished report that answers the wrong question.
A more dependable sequence
Before starting deep research, write a short commission: the question, the decision it will inform, relevant geography and dates, sources that deserve priority, and distinctions the final report must preserve. When the tool proposes its plan, inspect more than the elegance of the outline. Look for missing counter-evidence, methodology and definitions.
Once the report arrives, read the conclusions and sources before polishing the prose. Put the load-bearing claims into a small verification list and record what could not be confirmed. Writing, presentation or decision should begin only after the factual scope is stable. The research system can carry a large share of discovery and organisation in the first pass. It should not simultaneously become the sole author, evidence store and final reviewer.
Conclusion: choose the part of the chain you need to control
ChatGPT, Gemini and Perplexity can all produce impressive deep-research reports. The most useful distinction is not their visual finish. It is the part of the research chain over which the user needs control.
If you already know where dependable evidence should be found, source restrictions and plan editing matter. If the material lives in Google services, calculate the real labour saved by integration. If the assignment begins with a rapidly changing public web and will develop through repeated exploration, a search-centred environment is more natural. In every case, a citation supplies an entrance to verification; it does not perform verification for the user.
Deep research is most valuable when it releases people from repetitive searching so that attention can move to source judgement, conflicting evidence and correction of the question. If it merely produces an unchecked long report more quickly, speed has increased while the research remains unfinished.
Primary sources
- OpenAI: Deep research in ChatGPT
- Google: Use Deep Research in Gemini Apps
- Perplexity: What is Research mode?
- Perplexity: What’s New in Advanced Deep Research
Continue reading: Choosing the Right AI Tool
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.