Research and Evidence

Research and Evidence

OmniRank's design rests on published research into how search and generative engines select sources. This page cites that work precisely, states what each study actually measured, and marks the limits of each finding. Every number here is observational rather than causal, so treat the direction as useful and the magnitude as unproven.

That caveat is not modesty. It is the single most important thing on this page. The literature below is drawn overwhelmingly from English-language, business-to-business datasets, and several results come from models that have since been replaced. A tool that cited these figures as guarantees would be doing exactly what OmniRank refuses to do elsewhere.

A note on what "uses it" means. OmniRank acts on some findings — omnirank fix prints a real diff for 4 mechanical finding ids out of 67 total (see Fix-Preview) — but every citation on this page still describes research the content of a page needs to satisfy (statistics, quotations, primary-source citations, entity mentions), not the structural edits fix makes today. fix never drafts prose; it only ever splices a constant string or resolves an existing value. The research below still awaits aeo-onpage (roadmap), which is the skill that will actually draft against it — see Fix-Tiers-and-Applicability for why that boundary is deliberate.

What the research supports

On-page elements that lift AI visibility

Source: GEO: Generative Engine Optimization — arXiv 2311.09735, published at KDD 2024 (Aggarwal et al., Princeton).

The study tested content modifications against generative engines and found that adding statistics, direct quotations, and inline citations to primary sources produced the largest gains in how often a page was surfaced and quoted. Reported lift reached roughly 40% on their benchmark.

How OmniRank uses it: the aeo-onpage skill (planned) will treat those three elements as required when drafting, not optional polish. The shipped audit skill does not yet check for them, and omnirank fix does not draft content of any kind.

Limits: a single benchmark, on 2023-era models, in English. Whether the effect transfers to Bengali news content or Bangladeshi commercial queries is untested by anyone.

A structured criteria framework

Source: GEO-16 — arXiv 2509.10762.

Proposes a sixteen-criterion framework for evaluating a page's readiness for generative engines, covering extractability, factual density, structure and provenance.

How OmniRank uses it: it informs which gates exist in the AEO and GEO layers. OmniRank does not implement GEO-16 wholesale and does not claim conformance to it.

Brand mentions correlate with AI citation more than backlinks

Source: Ahrefs' analyses of AI Overview citations against their web index.

Reported correlation of roughly 0.66 for brand web-mentions versus roughly 0.22 for backlinks, with brand search volume the single strongest individual predictor.

How OmniRank uses it: it is why offsite-entity (planned, v0.4) prioritises unlinked mention breadth and entity resolution over link acquisition, and why the avoid-list forbids link buying.

Limits: correlation, not causation, with an obvious confound — large brands have both more mentions and more citations because they are large. Nothing here establishes that manufacturing mentions produces citations.

Answer-engine retrieval leans on a conventional index

Source: Seer Interactive's comparison of ChatGPT citations against Bing organic results.

Roughly 87% of ChatGPT citations matched pages ranking in Bing's top organic results, consistent with a retrieval layer built on a conventional index.

How OmniRank uses it: it is the reason indexing (planned, v0.3) treats Bing Webmaster Tools submission as the highest-value early action for a new domain. A page absent from the index has approximately no chance of being cited.

AI citation is largely decoupled from Google rank

Source: Ahrefs, across roughly 15,000 queries.

Only about 12% of AI-cited URLs also ranked in Google's top 10. Per-engine alignment varied widely: Perplexity approximately 28.6%, Copilot 8.6%, Gemini 8.2%, ChatGPT roughly 6–8%.

How OmniRank uses it: two ways. First, a site with modest rankings can still be cited, so the GEO layer is worth doing early. Second, and more importantly as a guard, the measure skill (planned, v0.5) must never diagnose a citation change from a rank change. They move independently.

What the research does not support

Claim	Status
"Publishing llms.txt causes AI engines to cite you"	Unproven. It is a public proposal. No engine has documented reading it as a ranking or retrieval input. See FAQ#is-llmstxt-a-real-established-standard.
"These techniques guarantee rankings"	False. Nothing guarantees rankings. OmniRank fixes signals engines demonstrably use; outcomes are downstream and uncontrolled.
"A 40% lift is what you will get"	Misreading. That figure is one benchmark under one set of conditions, not a forecast for your site.
"Brand mentions cause citations"	Not established. The correlation is real; the causal direction is not demonstrated and brand size confounds it.
"This works the same in every language"	Untested. No study in this list covered non-English content.
"omnirank fix applies this research automatically"	False. fix splices constants and resolves existing values for 4 mechanical finding ids; it drafts no prose and adds no statistics, quotations or citations.

Measuring it yourself

Because the published evidence does not cover most real situations, the honest method is direct observation. The measure skill (planned, v0.5) will query ChatGPT, Perplexity, Gemini and Copilot with a fixed question set and record whether your brand is named.

Two calibrations matter when reading those numbers:

• Treat referral analytics as a floor. Agentic browsers and native apps strip referrers, so a large share of AI-originated visits arrive as Direct. One published dataset put this near 70%. Whatever your analytics reports, the real figure is higher.
• Do not infer citation from rank. Per the decoupling result above, they are close to independent for every engine except Perplexity.

Citing this project

OmniRank ships a CITATION.cff, so GitHub's "Cite this repository" button produces a correct reference. Code is MIT; documentation and the generated corpus are CC BY 4.0.

---

See also: Glossary · GEO-Artifacts-Skill · Fix-Tiers-and-Applicability · Roadmap · FAQ