Original data and research: creating the content AI chooses to cite

Imagine asking an AI assistant how quickly small B2B teams respond to inbound leads. If every page on the subject repeats the same secondary claims, the engine can support its answer with any of them and needs none in particular. A page that reports a measured answer, with sample, period and method stated, offers what the others cannot: a fact with a single, attributable origin. That is the case for original data and research in generative engine optimization. Here, original research means content whose central findings come from data the publisher collected or generated itself, such as operational records, a survey, a benchmark or an experiment, and which therefore exists nowhere else in the same form.

The first half of this piece examines the evidence: what the academic study that introduced the term generative engine optimization tested, where its conclusions stop, and what "information gain" does and does not mean. The second half sets out a production method, from choosing the question to design, disclosure, publication and ethics.

Why would an engine prefer a page with first-hand data?

Google's documentation for AI Overviews and AI Mode states that appearing in them requires no additional optimizations beyond being indexed and eligible for a snippet, a point worth keeping in mind when reading about how AI Overviews changed the results page. It refers site owners to Google's guidance on helpful, people-first content, which asks whether content provides "original information, reporting, research, or analysis" and substantial value compared with other pages in search results. Neither statement reveals how an engine weighs sources, but both agree with a simple argument about redundancy.

When fifty pages state the same claim in different words, each is substitutable and none is necessary. A page containing a measurement nobody else has published is the only possible source for it, so citing that page is the natural way to attribute the measurement. The wider signals that make an engine willing to rely on such an origin are discussed in our article on trust and authority signals, and the underlying discipline in our introduction to GEO.

What did the GEO study actually test?

The paper that introduced the term is "GEO: Generative Engine Optimization" by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, posted to arXiv in November 2023 and published in the proceedings of KDD 2024. The authors built GEO-bench, 10,000 queries from nine sources across 25 domains, and their own engine, which fetched the top five Google results for each query and had gpt-3.5-turbo write an answer with inline citations. For each query, one source was chosen at random and rewritten by a language model with each of nine methods in turn, from an "Authoritative" tone and "Keyword Stuffing" to "Statistics Addition", "Quotation Addition" and "Cite Sources". Visibility was measured as a position-adjusted count of the answer's words attributed to the source and as a "Subjective Impression" score assigned by a language model.

The evidence-oriented methods did best: on the 1,000-query test split, the paper reports relative improvements of 30 to 40 percent on the word-count metric and 15 to 30 percent on the subjective one for Cite Sources, Quotation Addition and Statistics Addition. Keyword stuffing offered little to no improvement, and an authoritative tone no significant gain. In the per-category breakdown, the top three categories for Statistics Addition were law and government, debate and opinion, and for Quotation Addition people and society, explanation and history. When all five sources were rewritten at once, the evidence-oriented methods shifted visibility toward sources ranked lower in Google's results: with Cite Sources, the source ranked fifth gained 115.1 percent in visibility, while the top-ranked source lost 30.3 percent on average. On a 200-query subset run through Perplexity, with the source texts supplied as file uploads rather than retrieved by Perplexity, Quotation Addition improved the word-count metric by 22 percent and Statistics Addition the subjective metric by 37 percent, while keyword stuffing scored about 10 percent below the baseline on the word-count metric.

Where do the study's conclusions stop?

The authors note that methods may need to adapt as engines evolve, that real queries change, that effects on search rankings were not evaluated, and that some of the benchmark's query labels may be wrong. Three further limits matter in practice. The main engine was a 2023-era model answering from five supplied texts, and even the Perplexity test answered only from uploaded files, so neither setting involved a commercial system's own retrieval. The metrics measure how much of an answer draws on a source, not whether the answer is accurate. And the rewrites were machine-generated: the prompts in the study's public code repository allow the added statistics to be "hypothetical" and the added quotations and sources to be invented.

The study therefore shows that, in a controlled engine, adding the form of evidence raised a source's share of the answer; it does not show that any evidence will do, and it gives no license to fabricate. Commercial engines, journalists and customers can cross-check claims, and an invented statistic attached to your brand becomes a liability once repeated. The defensible lesson is narrower: supply real numbers, quotations and sources, and for claims about your own niche, the most dependable supply of real numbers is often research you conduct yourself.

Original research is cited for the same reason it is trusted: it states something verifiable that no other page can state.

What is information gain, and what does Google's patent say?

In information theory and machine learning, information gain is the reduction in uncertainty about one variable gained by observing another, formally an expected reduction in entropy; content strategists use the term loosely for what a page adds beyond what a reader has already seen. Google's US Patent 11,354,342, "Contextual estimation of link information gain", filed in 2018 and granted in June 2022, describes scoring documents by how much information they add beyond documents the user has already viewed, in the context of automated assistants and search results. A patent records an invention its owner chose to protect; nothing in it shows whether, or how, the method is used in Google's products, so it is not evidence of ranking behavior.

As an editorial test, the idea is useful regardless. List the pages already cited for your target prompts, which a source gap analysis produces, and ask what your page will state that none of them does. If the honest answer is nothing, the page is a paraphrase.

Which question can only you answer?

A strong research question meets three conditions: your audience needs the answer to decide something, no credible answer has been published, and you have privileged access to data that can answer it. Software and service companies often hold such data as a by-product of operations, such as aggregated response times or seasonal patterns, usable within what contracts and privacy law allow; where they do not, a survey, a controlled benchmark or a small experiment may. Write the question down as precisely as the angle in a content brief before choosing a method, because a vague question produces a vague dataset.

Which research design fits which question?

Each design trades a particular strength for a characteristic risk, which a credible methodology section names rather than hides.

Research type	Data source	Main strength	Main validity risk
Operational data analysis	Aggregated records from your product or service	Large samples of actual behavior	Your customers are not the market; definitions drift
Customer or market survey	Questionnaire to a defined population	Measures attitudes that leave no records	Unrepresentative sample, nonresponse, leading wording
Benchmark	Standardized tests of products or processes	Comparable figures under identical conditions	Test design favoring one option; fast aging
Controlled experiment	Randomized comparison against a control	Strongest basis for causal claims	Small samples; limited generalizability
Public-data reanalysis	Official statistics or open datasets	Inputs anyone can verify	Misread definitions; interpretation posing as data
Expert interviews	Structured conversations with practitioners	Explains mechanisms behind numbers	Selective interviewees; anecdote presented as prevalence

How do you design a survey that withstands scrutiny?

Define the population the findings should describe, by role, company size, market and period, and name the sampling frame respondents are actually drawn from; a survey of your newsletter subscribers describes your subscribers, not the industry. The American Association for Public Opinion Research (AAPOR) requires researchers to state whether a sample is probability-based, and its code says non-probability surveys should report measures of precision only with a detailed description of the underlying model, so an opt-in sample should not carry a conventional margin of error.

For probability samples, the margin of error is narrower than it sounds. As a Pew Research Center methodologist explains, it reflects sampling error only, not coverage, nonresponse or wording effects; it is larger for subgroups; and for the gap between two figures from the same poll, such as two candidates' shares, it is roughly double. For a simple random sample of 400, the 95 percent margin for a proportion near 50 percent is about plus or minus 4.9 points, enough to erase many headline "differences". Wording is the other half of the job, covered in our guide to survey questions that measure something; pretest every questionnaire on a few members of the target population.

What should the methodology section disclose?

A methodology section turns a claim into a citable finding, because it lets readers, journalists and models check what a number means. AAPOR's disclosure standards are a sound template: sponsor and conductor, exact question wording and response options, population and recruitment including incentives, mode and languages, dates of data collection, sample sizes, weighting, data-quality checks and exclusions, and a statement of limitations. For operational data, state the time window, inclusion criteria, number of accounts or events, metric definitions and treatment of outliers. Keep the section on the same page as the findings, version and date it, and keep earlier versions accessible.

How should findings be written and published so they are quoted correctly?

Write each key finding as a self-contained sentence carrying number, population, period and sample, so it survives being lifted out of the page. Consider a hypothetical: "In a September 2026 survey of 412 operations managers in Türkiye, 38 percent said they review supplier contracts quarterly." One finding per sentence, absolute figures beside relative changes, no adjective the data cannot support; the answer-first structure from our article on FAQ pages for AI applies here too.

Publish numbers as text, not only as images: the W3C's accessibility tutorial says header cells must be marked up as headers and data cells as data, and a table published as text, unlike a chart image, can also be parsed by software. For downloadable datasets, Google documents schema.org Dataset markup, with name and description required and properties such as license, creator, temporal coverage and distribution recommended, to help datasets appear in Dataset Search; implementation is covered in our guide to structured data markup. Set an update cadence before the first release and keep the method stable between editions; freshness is among the criteria an AI readiness audit checks.

Which ethical lines cannot be crossed?

People whose data you use must know the purpose: Türkiye's Personal Data Protection Law No. 6698 (KVKK) requires controllers to inform them, at collection, of the controller's identity, the purposes of processing, possible recipients and the purpose of any transfer, the method and legal basis of collection, and their rights under Article 11. Publish only anonymized aggregates. KVKK defines anonymization as making data impossible to associate with an identified or identifiable person even when matched with other data, and Recital 26 of the GDPR notes that pseudonymized data remain personal data, whereas truly anonymous information falls outside the regulation. Suppress categories small enough to identify a single company, see our guide to AI and customer data privacy, and take legal advice for your case.

Integrity matters as much as privacy. Fix the questions and analysis plan before seeing results, publish unflattering findings with the rest, and never round, reweight or omit to improve the story. Never fabricate statistics or quotations, whatever a benchmark of rewritten texts suggests, and do not mass-produce thin "research" pages: Google's spam policies classify generating many pages primarily to manipulate rankings as scaled content abuse.

To see whether the work registers, Source Intelligence in Rocketly GEO Suite shows which domains and URLs engines cite for your tracked prompts, with each source's history; Google AI Overview is measured weekly as standard, ChatGPT and Perplexity are in a controlled rollout, and Gemini, Claude and Grok are not measured. Content drafts in the product go through draft, review and approval by a different person and are never published automatically, and the research remains your team's work. Create a Rocketly account to track the prompts your study is meant to answer.

Frequently asked questions

Does original research guarantee that AI engines will cite my page?

No. No engine publishes its source-selection criteria in detail, and most of the published experimental evidence comes from a controlled benchmark rather than from commercial systems. Original data makes your page the only source for a specific claim, which removes the main reason to cite someone else for that claim; measure the outcome on your own prompts.

How large does a survey sample need to be?

It depends on the precision you need and the subgroups you will report. For a simple random sample, about 400 responses give roughly plus or minus 5 points for a proportion near 50 percent, and each subgroup has a wider margin. For opt-in samples, report the sample size and recruitment method instead of a margin of error.

How often should a study be updated?

Match the cadence to how fast the subject changes, for example annually for attitudes and quarterly for an operational metric, and announce it with the first edition. Date every version and keep earlier editions accessible, so anyone citing an older figure can still find its source.