Skip to content
SilktideHelp

Original research and statistics

Original research is content whose primary value is data you produced: a survey of your customers, a benchmark from your product, a measurement study, a longitudinal dataset. Everyone else who wants those numbers has to cite you - journalists, competitors' roundups, and AI assistants alike.

This is the technique Google's helpful-content guidance puts first when it asks whether content provides original information, reporting, research, or analysis. It is also what Silktide's AEO reports judge as Offers original information: when an assistant can get the same facts from a hundred pages, it cites the ones that offer something the others do not.

Throughout this page, suppose Fernwood runs expense management for thousands of companies and can see - uniquely - how long reimbursements actually take across its customer base.

Why it works

Most web content about a topic is paraphrase of other web content. Assistants and search engines both discount that pile: there is no reason to pick your paraphrase over the next one. Original data breaks the tie.

  • You become the source, not a retelling. A journalist writing "average reimbursement time" and an assistant answering the same question both need a number with a provenance. If Fernwood publishes "median reimbursement time across 12,400 companies was 2.1 business days in Q2 2026," that sentence gets copied; the copies point back.
  • It compounds. One well-structured study seeds comparison pages, blog posts, sales decks, and earned press for years. Journalists quoting your numbers are downstream of having numbers worth quoting - which is why this technique pairs with digital PR.
  • It is difficult to fake well. Fabricated statistics eventually contradict reality, get challenged, and destroy the trust this technique is meant to build. Real data, method disclosed, is durable.

The cost is real: designing a study, collecting clean data, and writing it up properly is more work than another blog post. That is also why it works - most competitors will not do it.

Deciding what to research

Pick questions buyers and assistants already ask, where you have privileged access to the answer:

  1. Operational data from your product. Fernwood can report median reimbursement time, approval-cycle length, or fraud-flag rates across its customer base - numbers no analyst without Fernwood's data can invent.
  2. Surveys of a population you can reach. Your customers, your community, a purchased panel - asked the questions your category argues about ("what share of expense reports still need manual rework?").
  3. Controlled measurements. Time a workflow in your product vs. a documented alternative; measure error rates; benchmark integrations. Publish the method so a sceptic could attempt a replication.
  4. Longitudinal series. The same metric, updated quarterly. A living dataset is cited repeatedly; a one-off chart is cited once.

Avoid research theatre: surveys of 40 people presented as industry truth, "studies" whose only finding is that your product is popular, or statistics restated from someone else's report with your logo on top. Google's helpful-content questions specifically penalise content that mainly summarises others without adding value - do not be that page.

Anatomy of a research page

Structure the page so a skimmer, a journalist, and an assistant all get what they need in that order. This is answer-first page structure applied to data.

1. The headline finding, first

Open with the one or two numbers that answer the question, dated and scoped:

Across 12,400 companies using Fernwood in Q2 2026, the median time from expense submission to reimbursement was 2.1 business days. Direct-deposit reimbursements were 1.4 days faster than cheques.

That paragraph is the quote. Write it as the sentence you want repeated on other sites and in assistant answers.

2. Method, early and honest

Within the first screen or immediately after the finding, state how the number was produced: population, time window, exclusions, sample size, and anything that would change the interpretation. Method is not an appendix journalists skip - it is what makes the number defensible enough to cite.

Method: all Fernwood customers with at least one reimbursed expense in April–June 2026 (n = 12,400 companies; 3.1M expense lines). Median of per-company medians, to avoid large accounts dominating. Excludes reimbursements delayed by missing receipts. Currency converted to USD at month-end rates.

If you surveyed people, say how they were recruited and whether they were incentivised. If you measured a workflow, say what you timed and on which versions.

3. Extractable figures

Put every citable statistic in a shape machines lift cleanly - tables for comparisons, short labelled paragraphs for single figures, not a chart that is the only place the number appears. Charts are for humans; the number must also exist as text. Example:

Reimbursement methodMedian business days (Q2 2026)
Direct deposit1.8
Corporate card settlement2.0
Cheque3.2

Assistants quote table rows. A PNG of that table quotes nothing.

4. Depth the finding earns

After the headline and method: breakdowns by segment, caveats, what surprised you, what you are changing because of the data. This is where expertise shows - and where Author pages and bylines matter. Name the researcher.

5. Reuse hooks

Offer a downloadable CSV or a stable URL for the dataset where you can, a short "how to cite this" line, and 2–3 suggested chart captions journalists can paste. Making citation easy is part of the technique.

Technical checklist

  • Declare dates in metadata - datePublished when the study went live, dateModified when figures or method change. See Machine-readable dates. A statistic without a date is a statistic assistants hesitate to repeat.
  • Use structured data for the article and its author. There is no universal "Dataset" rich result you can count on in search, but Article / Dataset markup still helps machines parse what the page is.
  • Make every figure readable without JavaScript - see Content readable without JavaScript. A study whose numbers only render in a client-side chart might as well not exist for AI crawlers.
  • Stable URL. fernwood.example\/research\/reimbursement-times-2026 (and a \/research\/reimbursement-times that always points at the latest). These URLs get bookmarked and cited for years; do not bury the study under a campaign slug.
  • Back every claim on the page with the figure or the method. This is what Claims backed by evidence judges - and it is the standard you are asking everyone else to meet when they cite you.

Keeping it honest

  • Do not cherry-pick to a predetermined marketing conclusion. If the data says cheques are fine for small teams, say so. A finding that concedes a limit is more citable than a finding that only flatters you - the same credibility pattern as comparison pages.
  • Do not quietly rewrite history. When you update a quarterly series, keep prior figures visible (a table of past quarters, or versioned reports). Assistants and journalists who cited Q1 need Q1 to still be findable.
  • Label estimates and opinions as such. Modelled figures, small samples, and editorial interpretations must not be presented as measured fact - Tone cannot be misread as fact exists for exactly this failure.
  • Privacy and consent. Aggregate and anonymise. Do not publish customer-identifiable data. If a survey involved people, say how consent worked.

How to tell if it worked

  • Search and ask assistants for the question your study answers ("average expense reimbursement time," "how long do reimbursements take") and see whether your page is cited and whether the number quoted matches what you published.
  • Track inbound links and press mentions of the specific figures - not just traffic to the page.
  • Watch competitors' roundups: when they start citing your number, the technique is paying rent.

How Silktide helps

Last updated

Was this page helpful?

Original research and statistics | Silktide Help