Short answer, as of October 2026: the most dependable way to be cited by ChatGPT and other AI assistants in your industry is to publish original figures from your own work — what jobs actually cost, how long they take, what goes wrong most often — stated plainly, dated, and with a short note on the sample behind them. It costs time, not money: usually a day or two with your own records. Tracking whether it worked starts at $99 a month and runs to $999 a month for the largest plans.
Most advice on getting cited is about formatting: answer first, use tables, add schema. All of that helps, and all of it can be copied by every competitor in an afternoon. What cannot be copied is information only you hold. An assistant choosing between ten pages that say the same thing has little reason to pick yours. An assistant looking for an actual number has to cite whoever published one.
Why original numbers get cited
The research that named generative engine optimization tested which changes to a page raised its visibility in AI-generated answers. Adding statistics, adding quotations and citing sources were among the most effective, with improvements of up to about 40 per cent in the authors' benchmark (Aggarwal et al., "GEO: Generative Engine Optimization", arXiv:2311.09735). The mechanics of how assistants choose are covered in how ChatGPT, Perplexity and Gemini pick sources.
The practical reading is straightforward. When someone asks "how much does it cost to…" or "how long does it take to…", the assistant needs a figure. Pages that offer a specific, sourced figure are usable. Pages that say "it depends on many factors" are not.
There is a second benefit. Other sites cite numbers too. A trade blog or a local journalist writing about your industry needs a figure as much as an assistant does, and those mentions are the third-party corroboration assistants weigh, as described in third-party sources and AI citations.
Five kinds of data you already have
You do not need to commission a survey. Look at what the business records in the course of ordinary work.
| Data you already hold | Where it lives | The question it answers | Example of a citable statement |
|---|---|---|---|
| Completed job prices | Invoices, accounting software | "How much does X cost?" | "Across 180 jobs in 2026, X cost between A and B, with most near C" |
| Time taken | Booking system, timesheets | "How long does X take?" | "The median visit lasted N minutes; one in ten ran over two hours" |
| Common faults or requests | Job notes, tickets | "What usually goes wrong with X?" | "Three causes accounted for most of the call-outs we attended" |
| Seasonal demand | Booking dates | "When is the best time to book X?" | "Enquiries roughly doubled between November and January" |
| Customer questions | Inbox, call logs, chat | "What should I ask before hiring for X?" | "The five questions customers asked us most before booking" |
The examples use letters in place of figures on purpose. Yours must come from your own records.
Each row is a page. Price data is usually the strongest, because cost questions are among the most common things people ask assistants and the least often answered with real numbers. The case for publishing prices at all is made in publishing prices and AI citations.
How to turn records into a publishable figure
1. Pick one question. "What does a boiler service cost in our area?" A single question gives a single clean answer.
2. Pull the records. Twelve months is a sensible window. Export the relevant jobs into a spreadsheet.
3. Clean them. Remove jobs that are not comparable: warranty work, bundled jobs, obvious data-entry errors. Write down what you removed and why.
4. Summarise honestly. Report the range, the middle and how many jobs it covers. A median is usually more representative than an average when a few large jobs would skew it. If the data splits naturally into groups, such as by property size or vehicle type, report each group.
5. Write the method note. Three sentences: how many jobs, over what period, what was excluded. This is what makes the figure trustworthy.
6. Anonymise. Aggregates only. Nothing that identifies a customer. If your records include personal data, check your privacy obligations before you use them, even in aggregate.
Publishing so the number is quotable
Lead with the figure. The first two sentences should state the finding, the sample and the date. "In our review of 142 jobs completed between January and September 2026, the typical cost was…" An assistant can lift that sentence whole.
Name yourself in the sentence. "In our review" becomes "in [Business]'s review" when quoted. Putting your name next to the number is how attribution survives.
Use a table. One row per group, with the range and the count.
Show the method. A short section headed "How we got these numbers". Do not hide it at the bottom in small print.
Date it. State the period covered and the date last updated.
Give it a stable address. Publish at a URL you will keep and update each year. Do not create a new page for each edition; update the same page and note the previous year's figure. Pages that go stale lose their place, for the reasons in citation decay.
Make it reachable. Link to it from related service pages and from your main navigation or resources section. A figure on an orphan page is rarely found.
Mistakes that get figures ignored
Rounding into mush. "Prices range from $50 to $5,000" is true and useless. Break the data into groups until the ranges are narrow enough to help someone.
No sample size. A percentage with no count behind it looks invented. Always give the number of cases.
Claiming more than the data shows. Your figures describe your jobs, in your area, in that period. Say that. "The national average is…" from 140 local jobs will be challenged, and it should be.
Marketing dressed as research. "97% of our customers are delighted" is not data anyone will cite. Report things a neutral reader would find informative, including results that are not flattering.
Burying it in a PDF. Publish on a normal web page. A downloadable report can accompany it.
Borrowed statistics presented as your own. If a number came from elsewhere, cite the original source. Never present an estimate as a measurement.
If you have very little data
A new business may have thirty jobs, not three hundred. There are still honest options.
- Publish what you have and say so. "Based on our first 30 jobs" is acceptable when stated.
- Run a small structured test. Time yourself doing the same task under different conditions, or compare three approaches side by side, and document it.
- Count something public. Tally how many businesses in your directory category publish prices, or how local regulations differ between neighbouring towns. Original analysis of public information is still original.
- Record from today. Set up the spreadsheet now and publish in six months.
More starting points for sites without a track record are in GEO for a new site with no authority.
Checking whether you were cited
Before publishing, write down the prompts your data answers, such as "how much does [service] cost in [area]". Run each several times across the assistants your customers use and record who is cited. Publish, wait a few weeks, and run them again. You are looking for two things: your page among the cited sources, and your figure appearing in the answer. Expected timings are in how long GEO takes.
Doing that by hand across six engines gets tedious quickly. DigiRank Expert's AI Visibility Tracker runs the prompts on a schedule across ChatGPT, Claude, Gemini, Perplexity, Grok and Google AI Mode and keeps each answer, so you can see when your number starts to appear and which page it was taken from. The Content Engine scores drafts against the Princeton GEO signals, statistics and citations included, before they go live. See the features page and the pricing page for what each plan includes.
Frequently asked questions
How can I get cited by ChatGPT in my industry? Publish specific, checkable information that other sites do not have, and state it plainly at the top of a page AI crawlers can read. The most dependable form is original data from your own work: typical costs, timings, failure rates or customer questions, with a note on how many cases it covers and over what period.
Do I need a large sample for original data to be useful? No. A clearly described sample of 50 to 200 real jobs is more useful than a vague claim about thousands. State the number, the period and what was included. Being open about a small sample makes the figure more credible, and readers can judge it for themselves.
What kind of data can a small business publish? Price ranges from completed jobs, how long work takes, the most common problems or requests, seasonal patterns in demand, and the questions customers ask most. All of these already exist in invoices, booking systems, call logs and inboxes.
Is it safe to publish data from customer jobs? Yes, when it is aggregated and nothing identifies a customer. Publish ranges, averages and counts. Remove names, addresses, registration numbers and any detail that could single out an individual, and check your privacy obligations before using personal data in any form.
How much does it cost to produce original data content? Mostly time. Pulling a year of invoices into a spreadsheet and writing it up is typically a day or two of work with no outside spend. As of October 2026, tracking whether assistants then cite it costs from $99 a month with DigiRank Expert, with larger plans up to $999 a month.
How often should published data be updated? At least once a year, and sooner if prices or conditions change sharply. Show the date range the data covers and the date it was last updated. Figures without a date tend to be passed over in favour of ones that have one.
Will AI assistants credit my business when they use my figures? Often, but not always. Assistants that show sources usually link the page a specific figure came from. Naming your business in the sentence that states the figure, such as 'in our 2026 review of 140 jobs', makes attribution more likely to travel with the number.