Ask a research question and get an answer built from papers, encyclopedia entries and official statistics that you can open and check. With the Firecrawl plugin, ChatGPT pulls those records from Firecrawl Alexandria: papers from the Firecrawl Research Index, articles from Wikipedia and country data from the World Bank.
Best for
- Students using ChatGPT for academic research who need a starting reading list where every paper has a DOI, PMID or arXiv ID.
- Analysts who want a sourced background brief on a topic before a meeting.
- Writers checking a statistic against the agency that publishes it.
Plugins and data providers
Plugin: the Firecrawl plugin. It adds web search, page scraping and the Firecrawl Alexandria catalog to ChatGPT, and any priced lookup is charged to the Firecrawl account you connect.
Data providers in Firecrawl Alexandria
| Provider | Tool | What you get back |
|---|---|---|
| Firecrawl Research Index | Search papers | Ranked papers for a question with title, abstract and IDs such as DOI, PMID or PMCID. Filters cover author, subject category and publication dates. |
| Firecrawl Research Index | Read paper passages | Excerpts from one indexed paper that are relevant to what you asked. |
| Firecrawl Research Index | Find related papers | Similar work, papers that cite a seed paper, or the papers it references. |
| Wikipedia | Structured article | The current revision of an article by exact title, parsed into abstract, infobox fields, sections and references, through Wikimedia Enterprise. |
| World Bank | Data | Indicator values by country and year, such as GDP or labor force participation, with the indicator's full name. |
| PubMed and PMC | Search | Biomedical records found with PubMed syntax, listing PMID, journal, DOI and the date it was published. |
Explore Firecrawl Alexandria data providers across categories
Starter prompt
Replace the topic with your own question and keep the structure:
@firecrawl I'm researching how remote work affects employee productivity.
Use Firecrawl Alexandria and tell me the price of each tool before calling
it.
Build me a short research brief:
- Background: the Wikipedia article "Remote work", summarized in three
sentences, with its last-modified date.
- Evidence: six papers from 2025 or later from the Firecrawl Research
Index, each with its DOI or PMID and one line on what it found.
- Numbers: one relevant World Bank indicator for two countries I name,
with the year of each value.
Only cite papers and figures the tools returned. If something is missing,
say so instead of filling the gap.Before you start
- Add the plugin. Use the button below to install Firecrawl in ChatGPT.
- Link a Firecrawl account. Paper searches in the Firecrawl Research Index cost nothing, while Wikipedia, World Bank and PubMed calls use a few credits. You can sign up for free.
- Know your plan. A few providers in the catalog are limited to paid Firecrawl plans, and ChatGPT will tell you when a tool isn't open to your account.
Limitations of using ChatGPT for research without Firecrawl
Is ChatGPT reliable for research on its own? It falls short on traceable sources and current data in five ways.
- Made-up references. OpenAI lists fabricated quotes, studies and citations among the mistakes ChatGPT can make, and says the model may sound confident even when wrong (OpenAI). In one study of medical articles written by ChatGPT, 47% of 115 references were fabricated and only 7% were both real and accurate (Cureus).
- Poor recall on a literature search. When researchers asked GPT-4 to reproduce the reference lists of 11 systematic reviews, it found 13.7% of the relevant papers and 28.6% of what it returned was hallucinated (JMIR). The authors advise against using these models as the main tool for a systematic review.
- A training cutoff. Answers don't cover events after the model's training data ends unless a tool fetches newer material (OpenAI). A study published last month won't show up in what ChatGPT remembers.
- Paywalls and blocked sites. OpenAI notes ChatGPT may not reach a page because of technical problems, paywalls or robots.txt settings (OpenAI). The University of Maine's guide to deep research names paywalled journals and licensed databases among the sources it may not access (University of Maine).
- ChatGPT deep research has gaps too. OpenAI says deep research can hallucinate facts, make incorrect inferences and struggle to tell authoritative information from rumor, and plans cap how many queries you get each month (OpenAI).
How ChatGPT gets papers, facts and figures with Firecrawl
A research request in ChatGPT runs through these parts:
- Firecrawl Alexandria: the data layer behind the plugin. It gathers 100+ official providers, each under a paid agreement, alongside indexes Firecrawl runs for the web, research papers, developer docs and government records. In a Firecrawl test of 845 tasks that kept the model and prompts constant, agents using Firecrawl Alexandria gave answers rated 21% better than agents that only had the standard web tools.
- Papers with IDs: the Firecrawl Research Index returns papers with their DOI, PMID or arXiv identifiers, so every entry in a reading list can be opened. Our search on remote work and productivity, limited to 2025 onward, returned PLOS One and Ergonomics papers with DOIs and PubMed IDs.
- Background and numbers from the source: Wikipedia arrives as structured fields with the article's last edit date, and the World Bank tool returns values tagged with country, year and indicator name. A query for India's female labor force participation gave 32.4% for 2025, labeled as a modeled ILO estimate.
- You see the bill first: ChatGPT can list tools and prices before running anything, and the Firecrawl Research Index searches are free.
Researching in ChatGPT with and without Firecrawl Alexandria
| ChatGPT alone | ChatGPT with Firecrawl Alexandria | |
|---|---|---|
| Where papers come from | Memory of training data, which can include invented titles | Indexed records, each with a DOI, PMID, PMCID or arXiv ID |
| Recent work | Stops at the training cutoff without web search | Searchable by date window, including papers from this year |
| Reading a paper | Depends on whether the page loads | Passages pulled from the indexed paper that answer your question |
| Statistics | A remembered figure with no year | World Bank values tagged with country, year and indicator name |
| Background | A summary from memory | The current Wikipedia revision with its edit date |
| Source | Hard to trace | The Firecrawl Research Index, Wikipedia and the World Bank, all available through Firecrawl Alexandria alongside 100+ other providers. Explore the catalog |
| Follow-up | Start a new search | Find papers that cite or are cited by one you liked |
Build your own research workflow
These ChatGPT prompts for research fit two different jobs. For a student writing a literature review:
@firecrawl Search the Firecrawl Research Index for peer-reviewed papers
from 2024 onward on social media use and adolescent sleep. List eight
with DOI, journal and year, then take the one closest to my thesis and
find papers that cite it.For an analyst preparing a market brief:
@firecrawl Give me background on the electric vehicle industry from the
Wikipedia article "Electric vehicle", then pull the latest World Bank
values for CO2 emissions per capita for China, Germany and the US.
Cite each number with its year.Both prompts name the sources, a date range and the kind of ID each result should carry, which leaves ChatGPT less room to fill gaps from memory.
Take it further
Quote a paper directly
@firecrawl From the Firecrawl Research Index, find the PLOS One paper on
experienced teammates and remote work productivity, and quote the
passages that report the productivity effect.Trace a paper's influence
@firecrawl Take the paper I just picked and find ten later papers that
cite it, ranked by how closely they relate to remote work productivity.Check a statistic
@firecrawl Is India's female labor force participation really around 32%?
Check the World Bank indicator and give the value and year.Compare encyclopedia and research
@firecrawl Summarize what the Wikipedia article "Four-day workweek" says
about productivity, then find three recent papers that test the claim.Ask for a price list
@firecrawl Which tools can search research papers, read Wikipedia or pull
World Bank data? List each with its price per call. Don't run any yet.What it costs
Looking around the catalog is free, and so is asking what a tool costs. Firecrawl credits are spent only when ChatGPT runs a priced tool, and Firecrawl pricing breaks down each plan.
| Tool | Answers | Credits per call |
|---|---|---|
| Search papers (Firecrawl Research Index) | Which papers address my question | 0 |
| Read paper passages (Firecrawl Research Index) | What does this paper say about X | 0 |
| Find related papers (Firecrawl Research Index) | What cites or relates to this paper | 0 |
| Structured article (Wikipedia) | Background on a topic with sections and references | 1 per article |
| Data (World Bank) | An indicator's value by country and year | 5 |
| Search (PubMed and PMC) | Biomedical records matching a query | 5 |
You'll see which tool ChatGPT ran, and its credit cost, at the end of each answer.
[ Research in ChatGPT ]
Install the Firecrawl plugin and build a reading list from real papers. Firecrawl Research Index searches are free.
Related use cases
- How to search arXiv and PubMed in ChatGPT: a deeper look at preprints, PubMed syntax and OpenReview peer reviews.
- How to find papers that cite a study in ChatGPT: trace a single study to the later work that cites it.
- How to use ChatGPT at work: the same approach applied to company, market and competitor research.
Coverage limits
- Indexed papers only. The Firecrawl Research Index covers what it has indexed, mostly abstracts plus arXiv, and an empty passage read doesn't prove a paper is silent on your question.
- Wikipedia needs an exact title. The tool looks articles up by title rather than searching, so a near miss returns nothing.
- Modeled statistics. Some World Bank series are estimates, such as ILO-modeled labor figures, and the indicator name says so.