Here's how Dupe does Deep Research
Total time saved by Dupe users
Every Dupe report runs through the same six-step pipeline: AI agents gather and screen evidence from across the web, candidates are scored on criteria we publish, and the Dupe team reviews the findings and selects the top picks. This page walks through each step and what flows between them.
368 Consumer Reports7 product categoriesMeet the research team
At a glance
Step by step methodology breakdown
Data only moves forward when it has earned it. A query becomes a buyer profile, the profile directs the sweep, the sweep produces sourced evidence, the evidence is screened, survivors are scored, and people approve what gets published. Each section below shows what flows in and out of that step.
The exact scoring weights and tooling are ours; everything a reader needs to interrogate a report (the sources, the criteria, the dates, the money) is on the report itself.
- Step 01Every report starts with a questionA search queryA buyer profile and research plan
- Step 02AI agents sweep the open webThe research planEvidence, kept with its source
- Step 03Evidence is screened before it countsRaw gathered evidenceA vetted candidate set
- Step 04Candidates are scored on category criteriaVetted candidatesA ranked shortlist
- Step 05Humans make the final callThe ranked shortlistApproved top picks
- Step 06We monitor the reports for the latest products and developmentsApproved top picksA live report, and corrections back into the pipeline
Step 01
Every report starts with a question
A report begins the way your shopping does: with a search. Before any research runs, we work out who is actually asking. A shopper hunting the lowest defensible price behaves differently from one who wants the closest match to a designer original, or one who only trusts what other owners say.
That buyer profile decides everything downstream: which products are even candidates, which criteria matter for this category, and what evidence would settle the question. It becomes the research plan the rest of the pipeline executes against.
- In
- A search query
- Out
- A buyer profile and research plan
Step 02
AI agents sweep the open web
Our research agents read what a careful human shopper would read (retailer product pages, current pricing, manufacturer specs, owner reviews, YouTube reviews, and Reddit threads), but at a scale no human team could: typically hundreds of sources for a single report.
Nothing is collected as a loose fact. Every claim is stored with the source it came from, which is why each published report can list its sources at the bottom. If we cannot point at where something came from, it does not get used.
- In
- The research plan
- Out
- Evidence, kept with its source
Step 03
Evidence is screened before it counts
Raw web data is messy: duplicate listings, outdated prices, reviews about a different variant, products that look right in a thumbnail but fail the category definition. Before anything is scored, the evidence is screened: deduplicated, checked against the category’s hard constraints, and stripped of claims that cannot be tied to a source.
Products are cut here too. A candidate without enough independent evidence behind it does not advance, no matter how good it looks. We would rather publish a shorter report than pad one with products we cannot back up.
- In
- Raw gathered evidence
- Out
- A vetted candidate set
Step 04
Candidates are scored on category criteria
Every surviving candidate is scored against the criteria that actually matter for its category: things like visual fidelity to the original, the real price difference, what owners report after months of use, and how materials and specs compare. The criteria change by category; a sofa is not judged like a serum.
We publish the criteria, not just the verdict: each report’s "How we picked" section states what its picks were scored against, so you can disagree with our weighting instead of having to guess at it.
- In
- Vetted candidates
- Out
- A ranked shortlist
Step 05
Humans make the final call
The Dupe team reviews the findings and selects the top picks before a report is published. Agents are good at coverage; people are accountable for judgment. A pick that contradicts its own evidence, a price that looks wrong, a product that should not have survived screening: this is where it gets caught.
The same accountability applies after publication: when readers or our own checks surface an error, we correct the report or take it down. The people behind this are on the research team page, not hidden behind invented bylines.
- In
- The ranked shortlist
- Out
- Approved top picks
Step 06
We monitor the reports for the latest products and developments
A published report shows its work: the sources behind it, the date it was last updated, the criteria it used, and a disclosure of how we make money. Buy links may earn us a commission, but reports are researched and ranked before monetization is considered; a product never ranks higher because it pays more.
Publication is not the end of the pipeline. Reader feedback and our own checks flow back to the start: reports get corrected, re-researched, or unpublished when the evidence changes. A recommendation is only as good as its maintenance.
- In
- Approved top picks
- Out
- A live report, and corrections back into the pipeline
The output
What the pipeline produces
The pipeline publishes into research hubs. Each topic gets two: a transactional Top Picks page for readers ready to choose, and an informational Guide for readers still working out what matters. Both are built on the same underlying reports and the same sourced evidence described above.
Get to know the Dupe Team
Reports are published under the Dupe Research Team: AI-assisted, human-reviewed, never under invented author names. When you buy through links on a report we may earn a commission at no extra cost to you; reports are researched and ranked before monetization is considered, and a product never ranks higher because it pays more.