Does publishing our own research help ChatGPT recommend us?

Your firm has been publishing for years. There are articles on the website, a newsletter and a few LinkedIn posts a week from the founders. It is decent work. And yet when a buyer asks ChatGPT or Perplexity a question in your field, the answer quotes a trade body, a large consultancy and a software vendor, and you are nowhere.

The difference is usually not the writing. Those three organisations publish numbers of their own, and when an assistant wants to say something specific, it needs a source it can name. In a Princeton and IIT Delhi study presented at KDD 2024, adding statistics to a page raised its visibility in AI-generated answers by about 34%, and adding quotations from relevant people by about 41%. It measured adding figures to existing pages rather than publishing new research, and it ran on a research engine, so treat it as a direction. But the direction is clear: specific numbers get used.

This guide explains why assistants treat original data differently from commentary, what counts as research for a founder-led firm, how many respondents you need, how to keep the credit when others repeat your finding, and how to run one credible study a quarter. Most of it needs a spreadsheet you already have.

#Why do AI assistants reach for original data?

Because a specific finding cannot be passed off as common knowledge. An assistant can say "outsourcing payroll reduces admin" in its own words without naming anyone, but a finding only one firm has published needs a source.

For example, suppose a payroll firm publishes: "in our review of 140 payroll transfers, most delays were caused by missing pension data, not by the new provider". An assistant that uses that point has a reason to say where it came from. It is a reason to name you, not a promise. In our experience assistants often credit the news site or blog that repeated a figure, or state the number with no source at all, which is why the section below on keeping the credit matters as much as the research itself.

There is a second effect. Google's own guide to its AI search features, last updated in July 2026, says content with first-hand expertise that could not come from anyone else counts for more than commodity content. We explain the wider idea in what E-E-A-T is and why AI assistants care about it.

Buyers respond the same way. In the 2025 Edelman and LinkedIn B2B Thought Leadership Impact Report, a survey of US executives, 64% of "hidden" decision makers such as finance and procurement said they trust thought leadership more than marketing materials when judging a supplier's capabilities.

#What kind of research do AI assistants cite most?

Research that other people have reported, that is recent and that carries specific numbers. Your own study is the start; coverage of it is what spreads it.

Muck Rack's May 2026 analysis of more than 25 million links cited by ChatGPT, Claude and Gemini found that earned media, meaning coverage by others, made up 84% of cited links and paid content 0.3%. More than half of the journalism cited had been published in the previous 12 months, and citations dropped sharply after six months. Profound's 2025 study of 680 million citations found the analyst firm Gartner among Perplexity's most cited sources and the review site G2 among ChatGPT's.

So publish the finding on your own site, get it into the trade press your buyers read, and refresh it every year. How to pitch it is covered in turning your own data into trade press coverage.

#What counts as research for a firm your size?

Far more than most founders assume. Research means a new finding with a clear method, not a five-hundred-person survey or a peer-reviewed paper. For most founder-led B2B firms the richest source is their own operations.

#Your own records

If you deliver work, you have data. For example, a commercial fit-out contractor with two hundred completed projects knows the average overrun, the most common cause of delay and the difference between landlord-led and tenant-led jobs. "Across 200 office fit-outs since 2020, jobs where the landlord's consent arrived after design sign-off took a third longer on average" is a sentence an assistant can quote, and nobody else can publish it.

#A short survey of the people you serve

A recruitment firm that asks sixty finance directors what they expect from a search partner has a finding that a general survey of a thousand office workers does not. Ask eight to ten questions, keep them plain, and include at least one where the answer might surprise you. Be honest about what the sample size supports, which the next section sets out.

#Public data, expert views and benchmarks

The data can be public if the analysis is new: assembling and reading it is what makes the result yours. Ten structured interviews, asked the same questions and written up with care, is research too. And the most valuable study is the one you repeat: after three editions of an annual benchmark, a buyer asking how your field is changing has one obvious source.

#How many survey respondents do we need for credible B2B research?

Fifty to eighty well-chosen respondents is enough for a credible B2B study, as long as you publish the margin of error and only claim what it supports. At fifty, a result of 60% could really be anywhere from about 46% to 74%.

The margin of error shrinks slowly as the sample grows. These figures use the standard formula for a simple random sample at the 95% confidence level, the basis Pew Research Center uses when it says a random sample of 1,067 gives a margin of about 3 points.

Two cautions. First, these figures assume a random sample; a survey of your own clients is not random, so treat the margin as the best case and say who you asked. Second, subgroups have their own, wider margins: if 50 respondents split into three sectors, each sector's result rests on fewer than 20 people on average. Report "about half" rather than "48.7%", and let the sample decide how bold the headline can be.

#What makes a finding quotable?

A quotable finding is specific, attributable and dated. It tells the reader who found what, among whom and when, in one sentence.

A live example: Helm Club, a membership community for founders, can say "1 in 5 new members this year came through an AI recommendation." It is one number, it belongs to one organisation, and it says exactly what was measured. That is the shape to aim for.

• Be specific. "Most delays are caused by missing pension data" is quotable. "Delays have many causes" is not.
• Be slightly surprising. A finding that confirms the obvious is rarely repeated.
• Lead to an action. "Firms that agreed a governance owner before choosing software finished sooner" tells the reader what to do.
• Carry the method. "Based on 140 transfers we managed between 2023 and 2025 for firms with 50 to 500 staff" is credible. "Based on our experience" is not a source.
• Sit on a page an assistant can read cleanly, with the finding near the top and the method close by. The structure is set out in what an article that ChatGPT actually quotes looks like.

Avoid any number bigger than the evidence behind it.

#How do we make sure ChatGPT credits our data and not the site that republished it?

You cannot force it, but you can make your firm's name travel with the figure and make your page the easiest one to cite. Five habits do most of the work.

#Publish one canonical findings page first

Put every study on its own page on your site, at a permanent address, before the press release or LinkedIn post goes out: findings at the top, then method, sample, dates, limits and a named contact.

#Write each headline figure as a sentence that includes your name

Example: "Across 200 office fit-outs since 2020, Example Ltd found that jobs with late landlord consent took a third longer." When a journalist or a blog lifts that sentence, the name comes with it. A bare "a third longer" travels without you.

#Put the figures in text, not only in charts or a PDF

Google does index text-based PDFs, but a web page is easier to cite, link to and update. Keep the key numbers in the page text, with each chart described in a sentence.

#Add Dataset markup

If you publish the underlying table, Dataset structured data tells Google its name, description, creator and formats, and makes it findable in Google's Dataset Search. It helps search engines read your facts correctly; OpenAI and Anthropic have not said they read it, so treat it as good housekeeping rather than a ranking trick. See whether we need schema markup for AI assistants.

#Ask everyone who reports it to link to the findings page

Give trade titles, partners and associations the findings page address and ask them to link to it rather than to your home page or the PDF. When coverage appears, ask politely for a link where one is missing.

#How to run a light research programme

The aim is one solid piece a quarter, not a study a week.

#Step 1: Pick four questions buyers actually ask

Choose four questions your sales team hears and competitors cannot answer with data, one per quarter, where you have the data or could gather it in a few weeks. Building content around the questions customers ask is the place to start if the list is not obvious.

#Step 2: Match each question to the cheapest honest method

Rotate through the methods in the table: records, survey, public data, interviews. Varying the method stops any single dataset carrying the whole year.

#Step 3: Do the work in a fixed window

Put a four-week window in the diary, with a named owner, or it will not happen. The effort for each method is in the table above.

#Step 4: Write the findings page, then the shorter versions

The findings page carries the results, the method, the limits and what the reader should do. Then produce a one-page summary, a chart or two, three short articles that each pick out a finding, and a pitch to your trade title with the most surprising number in the first line.

#Step 5: Date it, and promise the next one

Put the year in the title and say you will run it again. When the next edition comes, link back and describe what changed. Recent findings are cited more, so refreshing an existing study is often better value than starting a new one; see how old is too old for AI assistants.

#How do we make research credible?

State the method plainly: who, how many, when, how chosen and how analysed. Acknowledge the limits: if the sample is your own clients, say so and say what that might bias. Credit anything you built on, for example "this extends the trade body's 2024 survey". And name a person who will answer questions about the data.

A small sample with an honest description beats a large sample with none. "We asked 60 finance directors at UK firms in March 2026; the margin of error is about 13 points" is a source. "A survey found" is not.

#What to do this month

1. List the ten questions your sales team hears most, and mark the three where a specific number would change the conversation.
2. Open the records you already hold, such as your job list, CRM, support log or timesheets, and see which of those three questions the data could already answer.
3. Draft one finding as a single sentence that includes your firm's name, the sample and the date.
4. Create the findings page template: headline findings, method, sample, margin of error, limits and a named contact.
5. Book a four-week window to produce the first study, with a named owner and a date for the trade press pitch, and book the next quarter's window now.

#Frequently asked questions

#How much does it cost to run a B2B survey in the UK?

Cost depends mostly on who you need to reach. Surveying your own clients and contacts costs staff time. Buying responses from a research panel costs more, because senior B2B respondents are scarce, so ask panel providers for a price per completed response for your exact audience before you commit. An analysis of your own records is usually the cheapest credible option. Our own fees are on what it costs.

#Can we publish anonymised client data under UK GDPR?

Yes, if it is anonymised properly. The ICO says data protection law does not apply to anonymous information, but pseudonymised data, where someone could still be identified using other information, is still personal data. Small numbers and specific details can make people or firms recognisable, so publish totals and averages across enough cases, check your client contracts for confidentiality terms, and take advice if in doubt.

#Do we need a designer or a PDF?

No. A clear web page with the findings near the top, a simple chart and the method underneath is more useful. Google indexes text-based PDFs, but a page is easier to cite, link to and update. Produce a designed PDF for people who want one, and publish the page first.

#What if our findings are not surprising?

Publish them anyway if they are specific and useful. A confirmatory finding still gives an assistant a number to attach to your name. Then look harder at the data for the one result that is not what people assume, and make it the headline.

#Can we outsource the research?

Yes, an analyst or research firm can gather and analyse the data, and the findings are yours to publish if your agreement says so. Keep a founder in the room for the judgement about which questions matter and what the results mean for a buyer.

#How is this different from a white paper?

A white paper is usually an argument dressed as a report. Research is a finding with a method someone could check. If your white paper contains a number nobody else has and explains where it came from, it is research; if it contains a point of view and borrowed statistics, it is not.

If you want help choosing the first study and getting it into the trade press, Reputation That Wins is built around this work. To find out where your data already gives you an edge, start with a free Growth Review: 30 minutes with David Adams and three fixes in writing.

#Sources