How one paragraph on sample, limits and error bars gets your research cited by people and by AI answer engines.
Most company research ships with a headline number and a prayer. The sample size is buried, the dates are missing, and the word “limitations” shows up about as often as a unicorn at a board meeting. That is a wasted opportunity. In the studies below, work that showed its math drew 9% to 25% more citations, and web passages carrying statistics and quotations were up to 41% more visible in AI answers. Here is what goes on the page, how to confess a weak spot with some grace, and what the evidence says.
Q1. What is a methodology page, and why would an AI answer engine care?
1. It is one URL, or one clearly labeled section, that says who you studied, how many, when, how you counted, and where the numbers get shaky. One screen is plenty.
2. Answer engines quote passages, and they lean toward passages with numbers and attribution. In the Princeton-led GEO study, researchers rewrote web pages nine different ways and measured visibility on a 10,000-query benchmark. Table 1 shows what happened.
3. Smaller sites gained the most. A source sitting at search rank 5 gained 115.1% visibility from citing sources and 97.9% from adding statistics, while the rank 1 source lost 30.3% and 20.6%. Receipts are the underdog’s best weapon.
4. Getting retrieved is only half the trip. Peec AI reports that 64% of the URLs Perplexity retrieves get left out of the final answer (February 2026, 1 million+ citations). Something on the page has to earn the quote.
Table 1. What rewriting a page did to its visibility in AI answers
| Change made to the page | Visibility score | Change vs. baseline |
| No change (baseline) | 19.3 | n/a |
| Keyword stuffing | 17.7 | -8.3% |
| Cite sources | 24.6 | +27.5% |
| Add statistics | 25.2 | +30.6% |
| Add quotations | 27.2 | +40.9% |
Source: Aggarwal et al., GEO: Generative Engine Optimization, Table 1 (GPT-3.5, 1,000 test queries). Score is position-adjusted word count; percent changes are my math.
Q2. Is there real evidence that disclosure increases citations?
1. Yes, with honest asterisks. Table 2 lines up four studies you can open and check yourself.
Table 2. Four studies on showing your work
| Study | Sample | What they found |
| Colavizza et al., PLOS ONE, 2020 | 531,889 PLOS and BMC articles | Papers linking to data in a repository had up to 25.36% (ยฑ1.07%) higher citation impact |
| Piwowar & Vision, PeerJ, 2013 | 10,555 gene expression studies | 9% more citations with public data (95% CI: 5% to 13%); about 30% for 2004 and 2005 papers |
| Christensen et al., PLOS ONE, 2019 | 1,660 articles in two top economics and political science journals | Posting data added roughly 97 citations on a mean of about 100 |
| Kidwell et al., PLOS Biology, 2016 | Psychological Science, before and after open data badges | Data sharing rose from about 3% to 39%; comparison journals showed no change |
2. Now the asterisks. All of these are observational. And the famous early figure shrank: a 2007 study of 85 papers found a 69% citation boost, with a confidence interval of 18% to 143%, wide enough to park a truck in. With 10,555 papers and more controls, the same lead author landed on 9%. Smaller, far more believable, still positive. That revision argues for error bars better than I can.
3. Christensen’s team also found that across 30 journals, a data policy by itself produced no significant citation change, because only 35% of articles on average complied. The payoff went to authors who actually posted the data.
Q3. What belongs on the page?
1. Borrow from pollsters. The AAPOR disclosure standards list 11 items owed with any survey release, ending with “A General Statement Acknowledging Limitations of the Design and Data Collection.” Table 3 adapts them for a company study.
Table 3. The seven-line methodology page
| Element | What to write | Example line |
| Who ran it, who paid | Sponsor, authors, conflicts | “Run by our team. We sell software in this category.” |
| Sample | Size, population, recruitment | “812 U.S. HR managers from a paid panel.” |
| Dates | Collection window | “Fielded March 3 to 21, 2026.” |
| Definitions | Question wording, how terms were coded | “‘Adopted’ means in use for 90+ days.” |
| Precision | Margin of error or interval | “ยฑ3.4 points at 95% confidence.” |
| Limits | What the design cannot show, and which way it leans | “Panel members skew toward larger firms.” |
| Data | Raw file or full tables | “Download the CSV.” |
2. Error bars take one line of arithmetic. Using the formula that Statistics By Jim walks through (0.98 divided by the square root of your sample, at 95% confidence), Table 4 gives the worst case.
Table 4. Maximum margin of error at 95% confidence
| Sample size | 100 | 400 | 1,000 | 2,000 |
| Margin of error | ยฑ9.8 pts | ยฑ4.9 pts | ยฑ3.1 pts | ยฑ2.2 pts |
3. Then add humility. Jim puts it plainly: “The margin of error only accounts for random sampling error.” A Stanford, Microsoft and Columbia team checked 4,221 polls across 608 elections and found stated margins of about 3 points next to actual error of 6 to 7. Only 73% of Senate poll results landed inside the stated margin, where 95% should have. On Andrew Gelman’s blog in January 2026, one of the authors cautioned against treating “double it” as a general rule. Fair. The safe move is to state your margin and say out loud what it leaves out.
Q4. How do I write about my own study’s weaknesses without sinking it?
1. Put the finding first and the flaw second. In Elliot Aronson’s 1966 pratfall experiment, a quiz contestant who got 92% of answers right became more likable after spilling coffee on himself. One who got 30% right became less likable. Richard Shotton reran the idea with 626 people and two identical cookies: 66% preferred the one with the rough edge.
2. Keep the flaw specific and modest. Stanford research on the “blemishing effect” found that “small doses of mildly negative information may actually strengthen a consumer’s positive impression,” provided the negative arrives after the positives.
3. Write it in three beats: the limit, which way it probably pushes the number, and what you did about it. Example: “Respondents came from our own email list, so they likely know more about AI search than a typical marketer. Read our 61% as a ceiling. The raw responses are posted so you can cut them your own way.”
4. Watch a pro do it. In January 2026, SparkToro and Gumshoe had 600 volunteers run 12 prompts 2,961 times across ChatGPT, Claude and Google’s AI. The chance of getting the same brand list twice was under 1 in 100. Rand Fishkin then wrote that his starting hypothesis “was (probably) wrong,” added “No, I/we aren’t professional researchers or credentialed data scientists,” flagged his co-author’s conflict of interest, and posted the raw data. Search Engine Journal covered it within days anyway.
5. Whitespark’s 2026 Local Search Ranking Factors surveys 47 experts on 187 factors, then says flatly: “None of the experts have special access to the internal workings of Google’s local search algorithm.” It even admits its new AI visibility question “was a bit ambiguous.” It is still one of the most referenced studies in local search.
6. Make it findable. Uri Simonsohn at Data Colada counted that about 43% of 2022 papers in four psychology journals included a preregistered study, yet many links were broken or anonymous. His test: if a reader needs “more than 20 seconds” to find it, “the paper is not ready for publication.” Link your methods from the first paragraph.
Q5. What does skipping it cost?
1. Errors travel. Sayash Kapoor and Arvind Narayanan of AI Snake Oil documented data leakage errors in 17 fields, affecting 329 papers, then helped 19 researchers write a 32-item reporting checklist. The Retraction Watch database now holds 63,000 retractions.
2. Contradictions slip through. One AI citations report covering 1 million+ citations credits community platforms with 52.5% of citations in its summary and brands with 52.5% in its body, as I read it. A two-line definitions section would settle which is right. I say this fondly; the dataset is valuable, which is why the definitions matter.
3. Numbers age fast. Ahrefs reported that about 76% of AI Overview citations came from top-10 pages in July 2025. By March 2026, across 863,000 keywords and 4 million URLs, the figure was 37.9%. Ahrefs noted that its parsing method had changed in between. That one sentence keeps readers from drawing the wrong conclusion.
4. Buyers already discount you. In a SurveyMonkey and Reddit survey of 1,202 U.S. business decision-makers (ยฑ3 points), 73% trusted peer insights, 55% vendor websites and 39% AI chatbots. The Reuters Institute Digital News Report 2025, spanning 48 markets, found audiences expect AI-produced news to be less trustworthy (net score of -18).
5. Few competitors bother. Orbit Media’s 2026 survey of 1,042 marketers found only 13.9% reporting strong results, a record low, and noted that publishing original research “improves the likelihood of performance by 50%” even as fewer marketers do it. SEOTesting adds a rule worth framing: “Never end a test early just because you see good results!”
Q6. Where should the page live so AI systems can actually use it?
1. Give it a plain address such as /methodology and link to it from every chart and every headline number. Use real text. A methods note trapped inside an image or a slide is invisible to a crawler, and Otterly’s report claims 73% of sites already have technical barriers blocking AI crawlers.
2. Write sentences that survive being lifted out alone. Answer engines quote one passage at a time, so the sample, date and margin should travel with the number. Table 5 shows the difference.
3. Date it and keep a change log. SparkToro warned that “newer/updated models may have more or less difference than what we found.” Two lines of version history do that job for you.
Table 5. The same finding, written two ways
| Version | How it reads | What a reader or AI can do with it |
| Naked claim | “61% of HR managers have adopted AI screening.” | Quote it blind, or skip it |
| Citable claim | “In our survey of 812 U.S. HR managers (March 3 to 21, 2026; ยฑ3.4 points), 61% had used AI screening for 90+ days. Panel members skew toward larger firms, so treat this as a high estimate.” | Quote it whole, with sample, date and caveat attached |
Illustrative example with invented figures, matching the sample lines in Table 3.
Q7. What can this article not prove? (Yes, I am taking my own medicine.)
1. I found no published controlled test of “add a methodology page, then count AI citations.” The GEO experiment ran on GPT-3.5 with a 200-query check on Perplexity, and engines change monthly.
2. About half of my sources were chosen from a list of independent publishers, so the mix leans toward practitioners over journals. The percent changes in Table 1 and the margins in Table 4 are my arithmetic.
3. Still, the direction is consistent across medicine, economics, psychology, polling and search. I found no study in which disclosure reduced citations. The price is a paragraph.
Sources

Discover more from 1000 Startups
Subscribe to get the latest posts sent to your email.
Claude Penland builds the marketing and business strategy for companies that are good at what they do and hard to find. Thirty years operating, one exit, eight of them as a practicing casualty actuary.
The free two-page read is genuinely free. Email claude@1000startups.com and I'll send back what I can see from the outside. Or see the work samples and how to work with me.