By maxaeo.ai ｜ Published 2026-10-25 ｜ Updated 2026-10-25
How OpenAI Deep Research selects whitepapers has no published ranking formula. A whitepaper becomes a plausible source when a research task leads to it, the system can access and interpret it, and its evidence helps answer the question. Those are practical conditions, not a guaranteed citation recipe. For publishers, the goal is to make original findings easy to discover, inspect, and verify.
What does OpenAI disclose about source selection?
Deep Research is a multi-step research tool, not a whitepaper index. OpenAI says it can search the public web, examine uploaded files, use permitted connected sources, and work from a research plan that the user can revise. Users can also restrict research to selected websites or prioritize those sites while allowing a broader search. Its reports include citations or source links for verification. OpenAI does not disclose a whitepaper-specific scoring algorithm or promise that any eligible document will be cited. (help.openai.com)
That distinction matters. A public PDF, a file uploaded by a user, and a document retrieved through a connected source enter the task through different routes. Public discoverability helps only when the task permits public-web research; an uploaded document may be considered because the user supplied it. Even then, appearing among the materials examined is different from supporting a claim in the finished report.
What must happen before a whitepaper can be cited?
A useful working model is a four-gate evidence funnel: task fit, access, extractable evidence, and report use. This is an editorial framework, not OpenAI’s internal architecture. It helps identify where a whitepaper may fail without claiming to reverse-engineer a proprietary system.
| Gate | Question to ask | Publisher-side check |
|---|---|---|
| Task fit | Does the paper address a question a researcher actually posed? | Match its title and summary to specific buyer or research questions. |
| Access | Can the researcher reach the relevant document? | Test the public landing page and PDF without a login or form. |
| Evidence | Can a reader identify the method and the exact supporting passage? | Expose sample, timeframe, definitions, limitations, and page references. |
| Report use | Does the finding add something the final answer needs? | State a distinct result rather than repeating general advice. |
OpenAI says Deep Research can interpret text, images, and PDFs, but that capability alone establishes neither retrieval nor citation. Google separately confirms that PDFs are among the file types Google Search can index; Google indexability is a useful discoverability check, not proof of how OpenAI retrieves a particular file. (openai.com)
How should a publisher package the PDF and its landing page?
Treat the landing page as a clear route into the full study, and the PDF as the evidence of record. Give both a descriptive title. On the page, summarize the research question, the principal finding, who conducted the work, and what the reader can download. Link directly to the document rather than making visitors guess which button leads to the study.
Inside the PDF, make claims auditable: state the population studied, data-collection dates, sample size, exclusions, definitions, and limitations. Put a concise explanation beside each important chart. A statement such as “teams preferred option A” is hard to assess unless the paper says which teams, how many, and under what question. These details improve a human reader’s ability to check the work; they are not asserted Deep Research ranking factors.
A form-gated paper presents a different access problem from a publicly reachable one. If a publisher wants its findings considered in public-web research, it can publish a substantive, ungated summary while keeping any gated asset separate. The trade-offs are explored in what happens to a whitepaper behind a form. Similarly, a readable page is worth checking when a report repeatedly misses a source: ChatGPT browsing fetch errors explains how access failures can complicate diagnosis.
Why might a credible whitepaper still be omitted?
Omission does not establish that Deep Research rejected a paper’s quality. The task may be too narrow, the user may have restricted source domains, or another source may support the same claim more directly. A paper can also be useful during research without appearing in the final citations. OpenAI describes reports as the result of a planned, revisable investigation—not as an exhaustive bibliography of everything examined. (help.openai.com)
This suggests a better diagnostic question than “Why didn’t the AI like our PDF?” Ask: Which gate failed for this particular task? If the report cited a secondary article for your finding, compare that article with your original. Does it explain the result more plainly, identify the relevant figure sooner, or provide context missing from your landing page? If so, improve the original source rather than merely repeating its title across more pages. The distinction between an original and a cited copy is developed in syndication and AI citation confusion.
Finally, audit any citation before treating it as validation. OpenAI cautions that AI systems can produce incorrect facts or references; a source link should be checked against the sentence it appears to support. (help.openai.com)
How can you test whether changes help?
Use a controlled citation audit, not a single favorable report. The following is a proposed test protocol, not a claim about measured citation rates. It separates access problems from evidence problems and makes results more interpretable.
- Choose five concrete research questions. Each should correspond to a distinct finding in the paper; avoid prompts that simply demand a citation to your brand.
- Record the source conditions. Note whether the task allows the full web, prioritizes selected sites, or uses uploaded files. Do not combine those settings into one result.
- Save each report and source list. For every question, record whether the paper was cited, whether another source was cited for the same finding, and whether the cited passage supports the claim.
- Change one thing at a time. For example, add a methods summary to the landing page, then rerun the same questions under the same source conditions.
- Compare claim-level outcomes. A citation to your homepage is not the same as a citation to the whitepaper’s original finding.
The audit measures what happened in those tasks, not a universal probability of selection. For a broader view of how multi-step reports assemble sources, see how Deep Research citations differ from short AI answers. MaxAEO tracks brand mentions, citation sources, and competitor comparisons across AI engines; a free AI visibility diagnosis can help identify where to investigate next.
Frequently asked questions
Does Deep Research prefer PDFs over web pages?
OpenAI has not published a general PDF-over-HTML preference. It says Deep Research can examine PDFs and web material. Choose a format that lets readers reach the original evidence and check its methods; a clear landing page and a readable PDF can serve complementary purposes. (openai.com)
Will adding more keywords make a whitepaper get cited?
There is no disclosed keyword-density target for Deep Research citations. Descriptive titles and summaries help communicate relevance, but repeated phrases cannot substitute for an accessible document or a finding that supports the task at hand. OpenAI does not publish a whitepaper citation formula. (help.openai.com)
Can a gated whitepaper appear in a report?
It depends on how the task obtains the document. A user may upload a file or permit a connected source even when the public cannot reach the PDF. For public-web discovery, a form can prevent direct access to the underlying paper; test that route separately from uploaded-file research. (help.openai.com)
Is a citation proof that the whitepaper’s claim is correct?
No. A citation identifies a source to inspect. Check that the linked passage supports the precise claim, that its figures match the report, and that any limitations survived the summary. OpenAI advises verification because AI-generated answers can contain factual or citation errors. (help.openai.com)