Is original research worth it for a startup?

1. Blog
2. Discoverability & Content

By Saroj Jha · November 2, 2026 · 13 min read

Yes, if you can answer a question your buyers argue about with data nobody else has, and you are willing to show your method. A startup does not need a big survey: an analysis of its own anonymised product data, or of public data no one has counted, is often more credible and cheaper. What makes research worth the effort is not the headline number but whether a journalist, a writer or an AI system can check it, which means disclosing the sample, the dates and the definitions, and publishing the data under a licence that lets people reuse it with credit.

Those three numbers frame the argument: data helps a source get used, small samples are less precise than they look, and disclosure is what makes a finding citable.

Running case: Example Co. Fictional

Example Co. is a fictional seed-stage startup selling help-desk software to support teams at companies with 50 to 500 employees. It has 180 customers, a newsletter with 6,000 subscribers and a marketing team of the founder plus one marketer. The founder wants a piece of research that support leaders will quote, that trade press might cover and that AI answer engines might cite when someone asks about support response times. Every number about Example Co. in this article is invented for illustration.

Why does original research get cited when opinion pieces do not?

A writer who cites you needs something they cannot get elsewhere and can defend to their editor. An opinion can be paraphrased without credit; a number with a method behind it has to be attributed. That is why, in my view, a modest study with a clear method earns more references over time than a dozen well-written opinion posts on the same topic.

Google’s own guidance points the same way. Its self-assessment questions for helpful content ask, “Does the content provide original information, reporting, research, or analysis?” and whether it provides “substantial value when compared to other pages in search results”^[4]. Its guide to generative AI features in Search tells creators, “Don’t just recycle what others on the internet have already said, or could easily be produced by a generative AI model”^[5]. The same guide notes that a page must be indexed and eligible to be shown with a snippet to appear in those features, so research that search engines cannot crawl earns nothing there^[5].

The lab evidence on AI citations is narrower than it is often reported. Aggarwal and colleagues’ paper “GEO: Generative Engine Optimization”, accepted to KDD 2024 according to its arXiv record, tested rewriting methods on a benchmark of 10,000 queries^[1]. On the 1,000-query test set, with GPT-3.5 Turbo generating the answers, adding statistics to a source raised its position-adjusted word count, a measure of how much of the answer is attributed to that source, from 19.3 to 25.2, a 30.6% relative gain; adding quotations raised it to 27.2, a 40.9% gain, and keyword stuffing lowered it^[1].

The documents were already in the model’s context, and no clicks or citations in live products were measured. What the full body of GEO evidence does and does not show is covered in What is GEO?, so I will only draw one conclusion here: specific, attributable evidence helps a source get used, and original research is the most specific evidence you can own.

Example Co., step 1 Fictional

The founder lists the questions support leaders argue about on community forums and in sales calls: how fast weekend tickets get a first reply, whether AI drafts cut handling time, and how many tickets one agent can carry. None has a recent public answer based on real ticket data.

Do this

Write down five questions your buyers argue about where no one has published a number with a method, and keep only those you could answer with data you can get.

What can a startup credibly research?

There are three routes, and each has a different credibility problem to solve.

A survey. You ask a group of people the same questions. It is the most familiar route and the easiest to get wrong, because a startup’s respondents usually come from its own list, social followers or a paid panel, none of which is a random sample of the market. A survey is credible when the questions are neutral, the sample is described honestly and the findings are framed as “among respondents” rather than “among all support leaders”. The Survey Playbook covers question design and fieldwork in depth.

Your own product data. You aggregate anonymised behaviour across customers: response times, usage patterns, prices paid, outcomes. This is the route where a startup has a real advantage, because no one else holds the data. It needs customer permission under your terms and privacy policy, aggregation large enough that no customer can be identified, and a clear statement that the data describes your customers, not the whole market.

Public-data analysis. You collect and count something publicly visible that no one has counted: public pricing pages, job postings, documentation, regulatory filings, open datasets. It is the most reproducible route, because anyone can repeat it, and it needs a written collection rule, a date and a list of what was excluded.

In my view, own product data is the best first study for most startups after seed, and public-data analysis is the best first study before you have enough customers. A survey is the right choice when the question is about opinions or intentions that no behaviour data can show.

Example Co., step 2 Fictional

A newsletter survey would expect about 6,000 × 5% = 300 responses (an assumed response rate), all from people who already know Example Co. Example Co.’s own data covers 2.1 million tickets, and 142 of its 180 customers, 78.9%, have accepted terms that allow anonymised, aggregated use. The founder chooses product data for the first-response question.

Do this

Check what your terms and privacy policy allow before you plan a product-data study, and ask counsel if they are silent.

How big does the sample need to be, and what must you disclose?

Sample size matters less than sample honesty. The margin of sampling error, the range within which the true figure probably falls, applies only to probability samples, where everyone in the population had a known chance of selection^[2]. The American Association for Public Opinion Research states that it does not apply to opt-in online surveys and other non-probability polls^[2], which describes almost every startup survey. If your sample is opt-in, report how many people answered and who they were, and do not print a margin of error.

For probability samples, precision drops quickly in subgroups. AAPOR’s example is a survey of 1,000 adults with a margin of about plus or minus 3 points, where a subgroup of 200 has a margin of about plus or minus 6.9 points^[2]. The standard formula gives the same answer, and it shows that you need about 385 respondents for a margin of plus or minus 5 points on a simple random sample. My rule of thumb, not a benchmark, is not to report a percentage for any subgroup with fewer than 100 responses, whatever kind of sample you have.

Disclosure is what lets others trust and cite the result. AAPOR’s code lists 11 items to disclose immediately when results are released, including who sponsored and conducted the research, the exact wording of questions, the population, how the sample was drawn and whether it was probability-based, the mode, the dates, the sample sizes and precision, weighting, data processing and limitations^[3]. A startup should add one more: a plain statement that it sells a product related to the topic. That is the line readers look for first.

Example Co., step 3 Fictional

Example Co.’s methods note states: 2.1 million tickets from 142 customers who allow aggregated use, 1 January to 30 June; first response defined as the first agent reply, excluding automated acknowledgements; median reported because a few large accounts skew the mean; no customer contributes more than 10% of tickets; Example Co. sells help-desk software. The finding: the median first response on weekend tickets was 9.8 hours against 3.4 hours on weekdays, 9.8 ÷ 3.4 = 2.9 times longer.

Do this

Write the methods note before you look at the results, and publish it unchanged except for corrections you log.

How do you publish it so others can cite it?

Publish so that a busy writer can take one sentence and its source in under a minute. That means one stable page per study with the key findings written as complete sentences near the top, each with its number, sample and date; charts whose titles state the finding; and a methods section on the same page, not in a PDF behind a form.

Then make the data itself reusable. Put the figure data and, where privacy allows, a row-level file up for download, and choose a licence; Creative Commons Attribution 4.0 lets others copy, redistribute and build on the material, even commercially, as long as they give appropriate credit, provide a link to the licence and indicate if changes were made^[6]. Mark the dataset up with schema.org Dataset structured data: Google requires a name and description and recommends a licence, creator, distribution and temporal coverage, and says the purpose is to make datasets easier to find in Dataset Search^[7]. Add a suggested citation and a version history, and log every correction with its date.

AAJ’s own reports show what this looks like in practice, as examples of format rather than results. The Agent Readiness Study 2026 is a public-data analysis with a methods section, downloadable CSVs and a CC BY 4.0 licence. The AI Answer Visibility Index publishes its prompt set and answer-level data. The Marketing Budget Benchmarks 2026 is a different kind of study, a synthesis of published surveys that states each source’s sample.

Distribution is the other half. Pitch the finding, not the company, to a short list of reporters who cover your buyer’s problem; How to pitch a journalist covers that, and the Earned Media Playbook goes further.

Example Co., step 4 Fictional

The study page opens with three sentences, each a finding with its sample and date. The figure data and aggregated tables, with no customer-level rows, are downloadable under CC BY 4.0, the page carries Dataset markup, and a suggested citation sits under the title. The founder pitches the weekend finding to three trade reporters with the dataset offered in advance.

Do this

Before launch, ask someone outside the company to find one finding, its sample and its date on your page; time how long it takes.

What does it cost, and when is it worth it?

The main cost is time, not money. A first study from your own product data takes a few weeks of one person’s part-time work plus a data pull from engineering; a survey adds question design, fieldwork and cleaning; a public-data analysis adds collection and checking. My rule of thumb, not a benchmark, is to budget as much time for publishing and promotion as you would for a strong blog post, on top of the research itself.

It is worth it when three things are true, in my view. The question is one your buyers argue about and will search for; you hold data or access that makes your answer better than anyone else’s; and you will repeat the study, because an annual series builds a reference point while a one-off ages within a year. It is not worth it when the only data available is a thin opt-in survey that would need heavy caveats, or when the question is one no buyer is asking.

The risk to manage is credibility, not cost. A study with a hidden sponsor interest, an unexplained sample or a headline the data does not support can do more harm than no study, because it is the kind of claim writers check. The GEO & AEO Playbook covers how research fits into a wider plan for being found in AI answers.

Example Co., step 5 Fictional

Internal time is costed at $90 an hour. Survey route: 70 hours plus a $1,500 charity donation as the incentive, 70 × 90 + 1,500 = $7,800. Product-data route: 45 hours, 45 × 90 = $4,050. Public-data route: 80 hours plus $600 of tools, 80 × 90 + 600 = $7,800. Adding 20 hours of publishing and promotion to the chosen product-data route gives 65 hours, 65 × 90 = $5,850 for the first edition. The founder commits to repeating it every January.

Do this

Cost all three routes in hours before choosing, and include the promotion time.

What changes by stage?

• Pre-seed — You have little data of your own. A public-data analysis on a narrow question is the realistic option, and it can double as market research.
• Seed — Product data starts to be useful once enough customers allow aggregate use to protect anonymity. Choose one question and plan to repeat it yearly.
• Series A–B — You can afford a professionally fielded survey alongside product data. Consider a probability-based sample for any finding you expect to be quoted widely, and give one person ownership of the series.

Free tool

Design the survey before you send it

Survey Studio helps you draft neutral survey questions and plan the fieldwork before you send anything, so the method you publish is one you designed rather than one you explain away afterwards.

For how AI answer engines treat evidence, read What is GEO?, and for getting the finding in front of reporters, read How to pitch a journalist. The Survey Playbook covers fieldwork if you choose the survey route.

If you want me to look at your own funnel, positioning and channels, my Growth Audit ends with three prioritized fixes and a 30-day action plan.