Publishing original research that holds up

Several pages in this hub end at the same instruction: publish something only you can publish. That is correct, and it is not a method. This page is the method: what counts as original data, what makes a study credible, what disclosure a vendor owes, and how to publish so the result can be cited.

Why first-hand data is the strongest citable asset

One argument is measured. The GEO research paper tested page-level changes against generative engines. Its three best-performing changes were citing sources, adding quotations and adding statistics, with a reported relative improvement of 30 to 40 percent on its position-adjusted word count metric for that group. Statistics you generated yourself are the version nobody can copy.

The structural argument matters more. A paraphrase of common knowledge is available somewhere else, so an engine assembling an answer has no reason to name you. A number that exists in one place has to be attributed to that place, or it cannot be used at all. Google's guidance on generative AI features draws the same line, describing commodity content as material "based on common knowledge, which could originate from anyone" and contrasting it with content built on first-hand experience. Which page attributes earn citations is covered in how to get cited by AI engines.

What counts as original data

Original research does not require a research budget. Five forms are open to most companies:

• Your own operational data. Aggregate figures from the system you already run: throughput, failure rates, price movements, seasonality in your order book. Publish nothing traceable to an individual customer, and define every field.
• A survey of your customers or your market. Cheap to run and easy to do badly. It is legitimate if the population is described honestly: your customers are a population, not a proxy for everyone.
• A structured test you run. A benchmark, a repeated measurement, or a controlled comparison under conditions you write down in advance.
• An analysis of a public data set. Public data becomes original when you join it to something nobody has joined it to, or cover a period nobody has covered. The novelty is the analysis; credit the source.
• A documented case with permission. One customer, in numbers, published with written consent. A single case is a sample of one and should be labeled that way.

The method decides whether it is credible

The finding is not the asset. The finding plus a method a stranger can inspect is the asset. Apply this checklist before you publish.

• State the question before you look. Write down what you are testing and what result would count as an answer. A question chosen after the data arrives is not a question, it is a description of the data.
• Define the population and the sampling frame. Who were you trying to describe, and which list did you actually draw from? The distance between the two is the first limitation to name.
• Report the sample size and the collection window. Both, in the summary and not only in a footnote. "412 responses collected between 3 and 17 March 2026" is a study. "Our research shows" is not.
• Publish the instrument. The full survey wording, the test protocol, the query set, the script. Wording moves answers, and a reader who cannot see yours cannot judge your result.
• Report what you excluded and why. Duplicates, incomplete responses, bots, outliers. State the rule and how many records it removed, and set it before you see which way it pushes the answer. Exclusions decided afterward are how careful people fool themselves.
• Give the uncertainty, not a bare point estimate. Attach a margin or a range to every headline figure. A percentage from 200 responses carries a lot of room around it, and the room belongs in print.
• State the limitations, including who funded it. What the study cannot show, who paid for it, and who ran the analysis. Name the funder, especially when the funder is you.

Extra rules when the publisher sells something

A study whose result happens to support what you sell needs more disclosure, not less. Pre-register the question in public before you have the answer: a stated question is a constraint you cannot quietly move later, and moving it after the fact is the most common failure in commercial research. Report the result you got, including the one that does not help you, because a study that only ever confirms the sponsor's product is evidence about the sponsor, not the subject. Never publish a headline number your own method cannot support: if the sample is your own customers, the headline says customers, and if the difference is inside the margin, the headline claims no difference.

Publishing it so it can be cited

• A stable URL. One permanent address that does not move when the campaign ends. Citations accumulate on a URL, not on a study.
• A plain summary sentence that carries the finding. Near the top, self-contained, naming the subject, the figure, the population and the date, so it survives being lifted off the page.
• The method on the same page. Visible in HTML, not locked in a download or behind a form.
• The data downloadable where possible. A CSV of the aggregated results, with a data dictionary, where privacy allows.
• A clear license. State what others may do with the figures and charts. Permissive terms with an attribution requirement turn reuse into citation.
• A date. When the data was collected and when the page was last updated, as separate facts.

Then keep it alive. A study rerun each year becomes a series, and a series is cited as a reference rather than as news; the maintenance argument is in freshness. If the finding is also a fact about your company, it belongs in the canonical set described in the grounding page.

Distribution

Coverage is earned, not bought. Send it to people who already write about the subject, lead with the finding and the method rather than with your company, and have the data and the instrument available before anyone asks. The mechanics are in digital PR.

The honest part

Most published marketing research does not survive scrutiny: small sample, undefined population, a question written after the data was in, a headline claiming more than the numbers can bear. The fix is almost never a bigger sample. The fix is a smaller claim: one finding, stated precisely, about a population you can describe. A modest result with a visible method gets cited for years; an impressive one with no method is ignored by everyone qualified to check it.