Obsurfable"Publish original research to earn AI citations" has become standard GEO advice. Blog posts cite figures like 11.3 citations per page for primary research versus 3.4 for ordinary content, or 4.31× more citation occurrences for websites hosting first-party data.
A careful methodology review published by Liam Vi on 3 October 2026 argues those figures measure something different from what marketers think they measure — and the only published study that compared cited pages against uncited controls from the same search found no originality advantage in Google AI Overviews.
The claim vs the evidence
The popular narrative has three legs:
- Kevin Indig and Amanda Johnson (July 2026): 8 of 301 already-cited pages were primary research, averaging 11.3 citations each versus 3.4 for the rest.
- Yext (Q4 2025): Business-owned websites generate 4.31× more citation occurrences per URL than managed listings.
- Princeton GEO paper (KDD 2024): Adding statistics, quotations, and citations to existing pages raised visibility in generated answers by up to 40%.
Each sounds like proof that original data wins. Vi's review shows they answer different questions — and none directly tests whether publishing original research increases the chance of being cited in the first place.
The selection bias problem
Indig and Johnson's analysis used Gauge's dataset of 301 live pages that AI systems had already linked in answers to 316 prompts across seven verticals, with 1,075 citations total.
Of the 301, 8 qualified as primary research — "the original source of the data and methodology are on the page, rather than a writeup of someone else's numbers." Those 8 pages took 90 of 1,075 citations — 11.3 each versus 3.4 for the rest.
The finding is real within its frame: among pages AI already cited, research pages were cited more heavily.
But the frame excludes every page AI never cited. The question most marketers ask — "will publishing original research get me into AI answers?" — requires comparing cited and uncited pages. Indig and Johnson's set contains only winners.
The authors acknowledge the concentration: benchmarks of cloud data warehouses account for 75 of 90 research citations. A single Fivetran warehouse benchmark accounts for 44. Strip the benchmark cluster out, and first-party research barely registers.
Their own conclusion narrows the claim: the win is not "we published original data." What drew citations was a benchmark that compared named products on measurable terms with methodology shown in full.
The only cited-vs-uncited comparison
On-Page.ai's test, published 15 July 2026, is the only published comparison that scored both pages Google AI Overviews cited and pages they skipped — all from the same search results.
| Metric | Cited pages (median) | Skipped pages (median) | Significant? |
|---|---|---|---|
| Originality score (0–100) | 52 | 55.5 | No (p = 0.07) |
| Unique numeric data points | 4 | 5 | No |
Across 50 keywords in 10 verticals, ranking pages AI Overviews cited had a lower median originality score than pages they passed over. The gap was statistically indistinguishable.
The most original ranking page appeared among an Overview's citations in 15 of 46 searches where at least one scored page was cited. Random selection would include it about half the time.
Legal content leaned the other way — cited legal pages scored 62 versus 49 for skipped ones. But that rests on five keywords with no separate significance test. The study calls it "a direction, not a law."
Limits: one answer engine, one day, 50 hand-picked keywords, a vendor's proprietary originality score. ChatGPT was tested separately but searched the web in only 32 of 150 runs — too little overlap with Google rankings to run a cited-vs-uncited comparison.
The Yext 4.31× misread
ZipTie's September 2026 article stated: "Yext's analysis of 17.2 million AI citations found that websites hosting original research or first-party content generate 4.31× more citation occurrences per URL than listings."
Vi's review traces the figure to Yext's Q4 2025 study of 17.2 million citations across Gemini, Claude, Perplexity, and SearchGPT — at the location level, not the research level.
Yext's actual comparison: business-owned websites generate 4.31× citation occurrences per URL; managed listings generate 2.46×. The comparison is website vs listing — not research vs non-research.
Yext's top content tier is "content entirely created and hosted by the business": websites, blogs, and newsrooms. It does not isolate original research from ordinary blog posts. Reading 4.31× as a research multiplier is a category error.
The GEO paper measures editing, not publishing
The Princeton-led GEO paper (Aggarwal et al., KDD 2024) tested rewriting existing pages nine ways — adding statistics, quotations, citations — and measured visibility shifts in generated answers. Cite Sources, Quotation Addition, and Statistics Addition produced 30–40% relative gains.
That is evidence that adding statistics to an existing page can increase how much a generative engine draws on it. It is not evidence that publishing an original research report increases citation likelihood. Editing ≠ publishing.
What ChatGPT does differently
In On-Page.ai's ChatGPT runs, the model searched the web in only 32 of 150 test queries. Non-searching runs produced zero citations — original or otherwise.
When ChatGPT did search, it cited 7.8 sources per answer on average, but only 3 of 213 distinct URLs were first-page Google results for the query. 85% of cited URLs appeared in only one of three identical runs of the same keyword.
Two practical implications:
- Check whether the assistant searches at all for your topic before investing in citation-targeted content.
- Single-run citation checks are noise. Any report of "we got cited" needs to state how many runs stand behind it.
What both findings can be true at once
Indig and Johnson measured citation density among already-cited pages. On-Page.ai measured originality among ranking pages cited vs skipped by AI Overviews. Neither refutes the other:
- Research can be cited heavily once an assistant finds it (Indig's finding).
- Originality may play no role in what an Overview picks from the rankings (On-Page.ai's finding).
- Originality may matter more in sources fetched beyond the first page — On-Page.ai found off-ranking sources scored higher (median 57 vs 54), though the sample was small.
Vi hypothesizes — without testing — that originality counts in what assistants retrieve through query fan-out, not in what they select from page-one results. A benchmark answering a buyer comparison (Fivetran's warehouse benchmark) fits that pattern. A blog post with a novel statistic may not.
What Google actually says
Google's guide to optimizing for generative AI features (updated 10 July 2026) states: content "that people find unique, compelling, and useful will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide."
That is long-run guidance, not a measurement. It does not say "publish original research." It says publish content people find useful. Google's guide also explicitly rejects AI-specific schema, llms.txt, and content chunking — see Google Says AEO and GEO Are Still SEO.
What to do instead
Publish research for the right reasons
Amplistory's article — one of the sources Vi reviews — puts it well: "The research should still have a reason to exist if no AI system ever cites it." Bad surveys produce unique statistics that are still bad research.
If AI citation is part of the motivation, the narrowest evidence-supported path is Indig and Johnson's own description: a benchmark that compares named products on measurable terms with methodology shown in full. Not a press release with one survey question.
Measure before and after, with enough runs
No published study measures what happens to a site's citation rate after publishing research. If you publish a benchmark, track citation outcomes across multiple prompt runs over weeks — not a single check on day one.
Obsurfable is built for exactly this: repeated prompt runs that show whether your brand appears in answers and which sources get cited, over time.
Separate retrieval from absorption
Getting retrieved (into the candidate pool) and getting cited (selected as a source) are different gates. Original research may help with the second once you clear the first — but the published evidence does not show it helps with the first. Domain authority and relevance dominate retrieval — see Page Optimization Only Helps High-Authority Domains.
Do not conflate vendor marketing with evidence
On-Page.ai sells the originality score it tested with. Gauge, ZipTie, and amicited sell citation monitoring. Yext sells listings and AI search software. The Growth Memo is a paid newsletter. Read claims with that context — including Obsurfable's interest in citation measurement.
How Obsurfable fits
Obsurfable tracks whether your content actually appears in AI answers — not whether a vendor's proprietary score predicts it might. If you publish a benchmark or research report, Obsurfable shows whether AI engines cite it, name your brand, or ignore both.
The gap between "published research" and "got cited" is where most GEO programs fail silently. Obsurfable makes that gap visible.
Should we stop publishing original research?
No. Research earns links, press coverage, sales credibility, and reader trust — regardless of AI citations. The correction is to stop treating "original research gets cited" as proven when the published evidence compares winners with winners.
Does the 11.3 vs 3.4 figure mean anything?
Yes, for a narrow question: among pages AI already cites, research pages accumulate more citations per page. It does not mean publishing research increases your chance of entering the cited set.
What about earned media citing our research?
Valid path. AI engines cite journalism at meaningful rates in some studies — see Earned Media Drives 84% of AI Citations. Publishing research journalists can reference may earn citations indirectly. That is a distribution strategy, not an on-page originality strategy.
Is the On-Page.ai originality score reliable?
It measures how much of a page's text is new relative to other ranking pages for the same keyword — a vendor product, not an industry standard. Its null result on AI Overviews is directionally useful; treating the score as ground truth is not.
Bottom line
The popular claim that original research guarantees AI citations rests on figures that compare already-cited pages with each other — not cited with uncited. The only study that counted both found no originality advantage in Google AI Overviews. The Yext 4.31× figure compares websites to listings, not research to ordinary content. The GEO paper tests page editing, not research publishing.
Original research can still be worth doing. Just do not budget for it as a proven AI citation lever until you measure whether your specific research actually gets cited — with enough runs to account for the volatility every engine shows.