Reading time: 13 minutes
Original research is a study that a company designs, runs and publishes with a stated method. The goal is to create proprietary data about its market. Since the number doesn’t exist anywhere else, the brand becomes the source for anyone who wants to use it. This guide walks through each step, from the question to tracking citations.
In short
- In a Cision survey of 1,899 journalists, original research is among what they value most in a PR pitch.
- According to AAPOR, the margin of error only applies to probability samples, not to opt-in online panels.
- Costs range from almost zero with usage data to R$307,000 for an in-person Datafolha poll with 2,004 interviews.
- To measure results, track news stories, links and AI answers to questions that ask for the study’s number.
What is original research and why does it become a source?
Original research is any study in which the company produces the data itself. It can be a survey, an analysis of its own database or a market study. The result is proprietary data, that is, numbers no competitor has. So anyone who repeats the number has a reason to cite the company.
Google also asks for originality. One of the questions Google Search Central suggests for evaluating a page is this: “Does the content provide original information, reporting, research, or analysis?”
In its guide to generative AI search, Google goes further:
“Creating 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.”
Google Search Central (2026)
In the academic study that coined the term GEO, replacing qualitative claims with statistics raised visibility in generated answers. However, the test used numbers in general, not original research. The results for each tactic are in the guide to GEO tactics backed by evidence.
The press also wants new data. Cision, which sells PR software, surveyed 1,899 journalists in January and February 2026. The sample covers North America, Europe, the Middle East, Africa and Asia-Pacific. According to the journalists, original research, access to experts and embargoed information are what they value most in a pitch. In a plan for content marketing in the age of AI, original research therefore supplies the data only your brand has.
What kind of proprietary data becomes a source: surveys, usage data or market analysis?
All three types of original research can become a source, as long as the method is public. A survey asks a defined audience, while usage data comes from the company’s operations. Finally, market analysis organizes industry data with its own method. In fact, one report can combine more than one type of proprietary data.
| Type | Where it comes from | Example in Brazil |
|---|---|---|
| Survey | Questionnaire to a defined audience | RD Station Panoramas |
| Usage data | Company data, aggregated and anonymized | Serasa’s Mapa da Inadimplência |
| Market analysis | Industry data processed with a method | FipeZAP Index |
RD Station, which sells marketing and sales software, has published the RD Station Panoramas every year since 2022. The 2026 edition surveyed 2,790 professionals with an online questionnaire, according to the published methodology. In addition, it analyzed proprietary usage data from the company’s own tools, such as more than 1.3 billion lead conversions.
Similarly, Serasa, which offers debt negotiation, has published the Mapa da Inadimplência (Default Map) every month since 2016. The map uses Serasa Experian’s own data. The company calls it the country’s main indicator of consumer default. For example, its 10-year edition became a story on Metrópoles, a Brazilian news site.
The FipeZAP Index, in turn, tracks prices in real estate listings on Grupo OLX websites. Fipe calculates the index and also publishes the methodology and the historical series.
How do you use usage data without exposing customers?
Proprietary usage data needs extra care. Under Brazil’s data protection law, the LGPD, anonymized data is no longer personal data. The exception is when someone can reverse the anonymization with reasonable effort. So publish only aggregated numbers and avoid breakdowns that could identify a customer. Also, when the data comes from a single customer, it becomes a case study, not a research study.
How to do original research in 8 steps
Original research takes eight steps, from the question to measurement. First, you decide what to measure and with whom. Then you make sure the number holds up when someone checks it. Finally, you promote the study and track who cites it.
- Choose the question. Pick a number the market looks for but that has no recent Brazilian data yet. Ask Google and AI: if the answer cites an old or foreign study, there is room for yours.
- Define the audience and the sample. Say who the number describes and how many responses you need.
- Write the questionnaire. Use neutral questions, include “don’t know” and test it with five to ten people before fieldwork.
- Collect data in a fixed period. Record dates, channel and how each person reached the survey.
- Clean and analyze. Remove incomplete or rushed responses and say how many you removed.
- Publish the study page with findings, charts, a methodology statement and the proprietary data to download.
- Promote the study to journalists who cover the topic and on the brand’s channels.
- Measure citations in the press, in links and in AI answers.
How do you set up the sample and method so the data is reliable?
Start with the audience your original research will describe, and choose a sample that represents it. Then state the method: how people entered the survey, when they answered and how many responses survived cleaning. After all, without this, journalists can’t check the number.
The difference that matters most is between probability and non-probability samples. In the first, each person in the audience has a known chance of entering the sample. Household surveys by IBGE, Brazil’s statistics agency, work this way. In the second, however, people choose to take part, as in an online panel or a form sent to customers.
What is the margin of error of a survey?
The margin of error shows how much a result can vary because the survey heard a sample, not everyone. According to AAPOR, the American Association for Public Opinion Research, it only applies to probability samples. So the margin of sampling error does not apply to opt-in online surveys.
| Interviews | Margin of error |
|---|---|
| 200 | ± 6.9 points |
| 400 | ± 4.9 points |
| 1,000 | ± 3.1 points |
| 2,000 | ± 2.2 points |
We calculated the table at a 95% confidence level, for a 50% result in a simple random sample. AAPOR also notes that doubling the sample from 1,000 to 2,000 interviews cuts the margin by only about one point. In addition, each subgroup has its own margin. In a total of 1,000 interviews, a group of 200 people has a margin of 6.9 points.
What goes in the methodology statement of original research?
The methodology statement gathers what someone needs to check the number. In Brazil, Law No. 9.504/1997 makes anyone who runs an election poll for publication register a set of details in advance. They include who commissioned and paid for the poll, the cost, method, period, sampling plan, margin of error and questionnaire. Brand research has no such obligation, but the list works as a model:
- Who ran the research and who paid for it.
- Audience, recruitment method and channel: online, phone or in person.
- Number of responses before and after cleaning.
- Fieldwork period and weighting, if any.
- Margin of error, only if the sample is a probability sample.
- Full questionnaire, in the order people answered it.
How do you present the results so AI and journalists cite them?
Publish your original research as text on a page of your website. Put the findings at the top and the methodology statement right below. Write each finding as a complete sentence, with the number, audience and period. So anyone who summarizes the study, journalist or AI, carries the context along.
A sentence template helps: “X% of [audience] do Y, according to the [name] survey by [company], with N interviews in [month and year].”
Likewise, the page should follow an article structure that Google and AI cite: the answer up front and self-contained sections.
- A title with the main finding, not just the study name.
- Charts showing the base for each question, as the RD Station Panoramas do.
- Data to download in a spreadsheet, plus the full questionnaire.
- Open text on the page, because a PDF behind a form only reaches people who fill it in, and search bots don’t fill in forms.
For outreach, take the finding and the methodology statement to journalists who cover the topic. Offer an embargo until launch and keep a spokesperson available. After all, embargoed information and access to experts are also on Cision’s list. The press release carries the finding in the headline and the methodology statement at the end.
Taking the study to journalists is at the core of digital PR. In addition, the press is among the sources AI cites most, especially for questions about industry trends.
How do you know if the press and AI cited your original research?
Track three fronts: stories that cite the study, sites that link to its page and AI answers that use the number. Start in launch week and repeat every month. For each citation, also note whether the number was correct and whether your brand got the credit.
- Press: set up alerts for the study name and the company name, and log the outlet, date and link.
- Links: in Search Console, the Links report shows the top sites linking to each page. However, Google warns that the report shows a sample, not the complete list.
- AI: use questions that ask for the number, such as “how many Brazilian companies do X?”. Then check whether the answer cites your domain, a news story or another source.
- Traffic: the AI traffic channel in GA4 shows visits to the study that come from assistants.
If AI cites the news story instead of your original research page, the data reached the answer through the press. That also counts. Since answers change with every run, ask each question several times and in more than one AI engine. The protocol to measure brand mentions and citations in AI shows how.
How often should you repeat original research?
Repeat it as often as the data changes. For an opinion survey, one edition a year is usually enough, while usage data can come out every month. In any case, keep the same question and method so each edition compares with the previous one.
For example, the RD Station Panoramas come out once a year. Meanwhile, Serasa’s map and the FipeZAP Index come out every month. As a result, the proprietary data builds a series, and each edition brings a new number for the same story.
Give each edition a fixed name and URL, and keep the old ones online. If you change the method, say so on the page, because an unannounced change breaks the comparison.
How much does original research cost?
The cost ranges from almost zero to hundreds of thousands of Brazilian reais. Proprietary usage data and a survey of your own customers cost little, because there is no panel to pay. On the other hand, online panels charge per response, and polling firms with in-person interviews cost much more.
| Path | Field cost | What the data shows |
|---|---|---|
| Usage data | Analysis hours | What happens in the company’s base |
| Customer survey | Almost zero | What your customers think |
| Online panel | From $0.95 per response | What the panel’s audience thinks |
| Polling firm, in-person interviews | R$307,000 for 2,004 interviews | What the population thinks |
According to the price list of Pollfish, an online panel that sells the service, responses start at $0.95 each. Audience filters and screening questions raise the bill. Full-service projects with analysis and a report, on the other hand, start at $5,000 per market.
At the top of the scale is a national poll by the Brazilian polling firm Datafolha. The poll, with 2,004 in-person interviews, cost R$307,000, according to InfoMoney. The May 2026 poll was registered with Brazil’s Superior Electoral Court (TSE) and had a 2-point margin of error. It was an election poll, but it shows what an in-person sample with national reach costs. For any original research, also count the hours of whoever writes the questionnaire, analyzes the data and builds the page.
Frequently asked questions
Proprietary data is data only the company has, because it collected the data itself: survey responses, usage data and in-house analysis. In paid media, a company’s own customer data, called first-party data, helps target ads. To become a source, however, proprietary data needs to come out with its method published.
Primary data is what the company collects to answer a question, such as the responses to its original research. In contrast, secondary data already exists and comes from third parties, such as public agencies and other studies. Market analysis generally organizes secondary data with a new method.
No. Under the text of Brazil’s Law No. 9.504/1997, registration with the electoral courts applies to published polls about elections or candidates. Therefore, a brand survey about consumer habits falls outside this rule. However, if the questionnaire asks about elections or candidates and the brand publishes the results, the requirement may apply.
Yes, as long as the text says the sample consists of customers. That is because a survey sent to your own base shows what your customers think, not the whole market. Since there is no random draw and each customer decides whether to answer, the sample is not a probability sample. So the results should not come with a margin of error.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- AAPOR, margin of sampling error and credibility interval (2023).
- Cision, State of the Media, survey of 1,899 journalists (2026).
- Google Search Central, Google’s guide to optimizing for generative AI features on Google Search (2026).
- Google Search Central, creating helpful, reliable, people-first content (2026).
- RD Station, RD Station Panoramas, introduction and methodology (2026).
- Serasa, Mapa da Inadimplência do Brasil (Brazil’s Default Map), 10-year edition (2026).
- Planalto, Brazil’s Elections Law, article 33, on registering polls (1997).
- InfoMoney, sample, margin of error and cost of a national Datafolha poll (2026).
