Original research earns citations because it gives other people something they cannot get anywhere else: a number, a pattern, or a dataset that did not exist before you published it. AI assistants and the writers who feed them need specific, checkable claims. If your pages only restate what everyone else already says, there is no reason to name you as the source.
The good news is that "original research" does not mean a survey firm, a six-figure budget, or a year of waiting. It means you collected something yourself and published the method. A small study you can finish in a week beats a big one you never start.
Here is a worked example from our own site, followed by a repeatable process you can run for your brand.
The worked example: 210 robots.txt files
We wanted to know how major websites treat AI crawlers. Not what people say about blocking AI, but what the files actually say. So we fetched the robots.txt of 210 major websites and read them.
That study is here: Who Blocks AI Crawlers? We Read the robots.txt of 210 Major Websites.
Why it works as citable content:
- The scope is concrete. 210 named websites, one file each. A reader can picture exactly what was examined.
- The method is reproducible. Fetch robots.txt, parse the user-agent groups, record which named bots appear and what rules apply. Anyone can rerun it.
- The limits are stated. robots.txt expresses intent, not enforcement. A file can name a bot and the bot can still fetch. The study says so instead of implying the files prove behavior.
- The dataset is published. The raw findings are available, so a journalist, blogger, or another analyst can check the counts or slice the data differently.
That last point matters more than the headline finding. When you publish the underlying data, you become the citation for everyone who builds on it, including people who disagree with your interpretation.
Why AI answers reward this
AI assistants assemble answers from sources they can quote precisely. A page that says "most sites block AI crawlers" gives them nothing to work with. A page that says "we checked 210 sites and here is the breakdown by bot" gives them a claim with a boundary around it.
The mechanics differ by vendor, and it helps to know which surface you are aiming at:
- OpenAI's OAI-SearchBot surfaces websites in ChatGPT search, and sites opted out of it are not shown in ChatGPT search answers. GPTBot is separate and collects content for model training; blocking it does not affect search.
- Anthropic's Claude-SearchBot indexes content for search, while Claude-User fetches pages when a user asks a question. Blocking Claude-User prevents retrieval for those queries.
- PerplexityBot surfaces and links sites in Perplexity search results. Perplexity-User fetches pages for user questions and generally ignores robots.txt because the user requested the fetch.
- Google-Extended controls training and grounding for Gemini, not inclusion in Google Search. AI Overviews are part of Search and follow normal Googlebot crawling.
None of that guarantees a mention. What it does mean is that being crawlable and quotable is a prerequisite, and original data is the part of the page most likely to be quoted.
A process you can run in a week
1. Pick a question your buyers already argue about
The best questions are ones where practitioners disagree and nobody has counted. In our case: are big sites actually blocking AI crawlers, or just talking about it? For your brand, it might be how competitors structure pricing pages, how many support docs answer a question in the first paragraph, or how often a category's top pages mention a specific standard.
If the answer is already well documented, skip it. You want a question where a count would settle something.
2. Define scope tightly enough to finish
"210 major websites" is a scope. "The AI industry" is not. Write down your inclusion rule before you start collecting, and apply it without exceptions. If you bend the rule mid-study, say so in the writeup.
A few hundred rows is plenty. The value comes from the selection being defensible, not from the sample being large.
3. Collect by hand or with a simple script
Many useful studies need no code at all. Open the pages, record what you see in a spreadsheet, one row per item. If you do script it, keep the script simple enough to describe in two sentences.
Record more than you plan to report. Extra columns cost nothing and often become the most interesting finding.
4. State limits before anyone else does
Every small study has them: sample bias, a snapshot in time, a definition that could be argued. Naming your limits makes the rest of the piece more credible, not less. In the robots.txt study, the key limit is that a file's rules describe intent. A site can name a crawler in a disallow rule and the crawler may still fetch, because robots.txt is not access control.
5. Publish the dataset
Put the raw rows somewhere a reader can download or copy. A plain CSV, a public spreadsheet, or a table in the post all work. State the date you collected it, since the data ages.
This is the step most teams skip, and it is the step that generates citations. People cite what they can verify.
6. Write the page so the findings are extractable
One finding per heading. Put the number in the sentence, not in a chart only. A short method section near the top, limits near the bottom, and a table if you have three or more categories. If you want the format that tends to get quoted, see Comparison Pages That AI Assistants Actually Quote.
Turning one study into several citable pages
A single dataset can support a cluster of pages without repeating yourself:
- The main study with the full method and dataset.
- A breakdown by segment, if your data has segments.
- A practical piece on what the findings mean for readers.
- A follow-up with a second collection six months later, which lets you show change over time.
Each page needs its own reason to exist. If two pages would make the same claim, merge them. AI assistants pull from pages that answer a specific question, and duplicate pages compete with each other for the same slot.
What to avoid
- Rounding your way to a bigger claim. If 40 of 210 sites do something, say 40 of 210. Percentages without the base invite correction.
- Presenting a snapshot as a trend. One collection is a snapshot. Say so unless you have a second collection to compare.
- Burying the method. If a reader cannot tell how you got the number, they will not repeat it, and repetition is what makes a citation travel.
- Chasing a viral angle over a useful one. A boring question answered precisely gets cited for years. A dramatic question answered loosely gets corrected.
Make the data the asset
The teams that get named in AI answers tend to own a small body of facts nobody else has. Not a big content library, a specific one. Original research is how you build that, and it compounds: each study gives you something to reference in future posts, something for others to cite, and a reason for readers to come back when you publish the next one.
If you are tracking whether those citations actually show up, that is what PromptRanking measures across ChatGPT, Gemini, and Perplexity.
Next step: pick one question your buyers argue about, write the inclusion rule for a sample of 50 to 200 items, and block two hours this week to collect the first 25 rows. If you can finish the first 25, you can finish the study.