Sources checked 8 November 2026.
Nobody has shown it in a controlled test. What the evidence supports is narrower: a finding that exists only on your page gives an AI engine a reason to retrieve it, because the answer can't be stated accurately without it. Whether the engine then credits you is a separate question, and the evidence on credit is poor. This site is built on that bet, and on 2 January 2027 it will publish the first scored test of it.
What the companies say
Google's guide to its AI features asks for non-commodity content, which it describes as providing "unique expert or experienced takes that go beyond common knowledge and the ordinary" (Google guide, last updated 10 July 2026). Bing's guidelines say content should "be original and authoritative" (Bing).
Both are statements of what the companies say they want. Neither is a measurement of what their systems do.
What the research shows
The most quoted study is the GEO paper, accepted at KDD 2024. It reports that content edits can raise a source's visibility in generative engine answers "by up to 40%", with results varying by domain (Aggarwal et al.). It tested edits to existing pages on a benchmark, not the effect of publishing new research, and it predates the current versions of every engine named in this article.
On links and authority, this site reviewed 16 studies and found all 16 correlational, with no raw citation data published (backlinks and AI citations). I haven't found a controlled study, before or after, of what publishing original research does to AI citations. If you know of one, send it to me.
The argument for it
I call this a lone-source finding: a result that exists on one page because the data behind it was collected only there. "Nine of the UK's ten biggest newsbrands block at least one AI crawler, checked 2 October 2026" is one (crawler study). Nobody else checked those ten files on that day with that definition.
My argument (original analysis, not an established finding): an answer engine asked a question that only a lone-source finding answers has three options. It can retrieve the page that holds the finding, retrieve a copy of it, or answer without it and be vague or wrong. Commodity content never forces that choice, because the same answer is available from a hundred pages. That's the whole case for original research in AI search, and it's about being retrieved. It says nothing about being credited.
The argument against, or at least for caution
Credit is where the evidence turns bad. In Tow Center's test of eight AI search tools, more than 60% of queries asking them to identify a news article's source were answered incorrectly, and tools often cited syndicated copies on Yahoo or AOL instead of the original publisher, including publishers with licensing deals (Tow Center, CJR). If that holds for research, the likeliest outcome for a small site is that its finding gets used, and the credit goes to whoever repeated it with more authority, or to nobody.
There's a second objection. The GEO paper's gains came from editing existing pages, not from publishing anything new. If engines respond to how specific and checkable a page's claims are, wherever they came from, then a page that repeats your figure with a citation to you could do as well as your page. Original research would then help the field but not necessarily the researcher.
I think both objections are serious, and I've written my own predictions to reflect them.
The test
This site registered a study plan on 14 October 2026, before any results existed: nine findings from its own studies, nine questions that only those findings answer (none naming this site), four engines, two runs each, 72 answers per audit. Each answer is scored on whether the finding appears, whether the number is right, whether the date is given, and whether jacobwhitmore.com is linked, named or not credited (findings recall audit). The first audit is due 2 January 2027.
Two calls on the predictions board rest on it:
- P21: no more than 3 of the 72 answers will credit jacobwhitmore.com by link or name.
- P22: at least one answer will state one of the nine findings in full without crediting the site.
So my own bet is that research gets used before it gets credited. If P21 misses because the site is credited more often than I expect, that's evidence for original research I didn't predict. If both hit, it's evidence for the caution above. Either way the result will be on the board, in red if I'm wrong.
What to do if you publish research
- State the finding in the first sentence with its date and denominator: "7 of 40", "checked 10 October 2026". That's the sentence that gets lifted.
- Put it in body text and an HTML table, not only in a chart or a PDF. In this site's extraction test, body text and tables survived all four tools, and JSON-LD survived none.
- Publish the data and the method. It's what separates a finding from a claim, and it lets others cite you accurately.
- Give each finding one stable URL. If you update it, keep the old figure visible with its date.
- Plan for being used without credit. Check quarterly with exact-phrase searches and by asking engines the question your finding answers.
Limits
Everything here about why research should help is an argument, not a measurement. The recall audit will be one site, nine findings, four engines and one quarter. It won't settle the question for anyone else, and I'll say so when it's published.
Conclusion
Original research is the one kind of content an AI can't answer around, which is a reason to expect it to be retrieved. It isn't a reason to expect credit, and the best evidence on credit says engines are bad at it. I'm publishing research anyway, and I've written down in advance how I expect it to go.
Frequently asked questions
Does original research improve AI citations?
There's no controlled study showing it does as of 8 November 2026. Google and Bing both say they want original content, which is intent, not evidence.
What is a lone-source finding?
A result that exists on one page because the data behind it was collected only there. It's a term used on this site, chosen on 8 November 2026 after a quoted Google search found no other use of it.
Can editing a page help it appear in AI answers?
The GEO paper (KDD 2024) reported visibility gains of up to 40% from content edits on a benchmark, varying by domain. It didn't test today's live engines.
Will AI credit my research if it uses it?
Not reliably. Tow Center found AI search tools often credited syndicated copies over original publishers.
How is this site testing it?
With a registered quarterly audit of nine of its own findings across four engines, first due 2 January 2027, and two predictions written down beforehand.
Sources
All opened 8 November 2026.
- Google, Guide to optimizing for generative AI features, last updated 10 Jul 2026
- Microsoft, Bing Webmaster Guidelines
- Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024
- Tow Center / CJR, AI Search Has a Citation Problem, 6 Mar 2025
- jacobwhitmore.com, Study plan: findings recall audit, registered 14 Oct 2026
- jacobwhitmore.com, Predictions board (P21, P22)
- jacobwhitmore.com, Backlinks and AI citations: what 16 studies measured, 30 Oct 2026