By Elizabeta Kuzevska, Co-Founder, Revenue Experts. About a 12-minute read.
How do you get cited by AI?
To get cited by AI, put a checkable fact next to every claim on the page an assistant reads. Assistants lift sentences they can verify, such as a dated number, a named case, a documented method or a specification, and skip adjectives like “world-class”. Two plain checks do exactly this: an evidence check, which asks whether your page carries first-party evidence, and a claim-support check, which asks whether that evidence sits next to the claim it supports.

Imagine you ask ChatGPT why a buyer should choose your company, and the answer is polite, vague and could describe any competitor. That is because every sentence it found on your page could describe them too. Your AI is already answering for your company, and nobody has checked what it says.
In our field service software study, the 108 answers from ChatGPT, Claude and Perplexity carried 1,493 citations, and 1,181 of them went to pages outside the 18 vendors’ own websites [1]. The pages cited most were not home pages. Each one was a comparison, a price breakdown or a direct answer to one narrow buying question. One field service company that was not even on our list of 18 had its blog cited 63 times, close to the 68 citations of the category leader’s own website [1]. The assistants quoted the pages that said something checkable.
This guide gives you both checks as a 45-minute job you run by hand on one page. For a plain-English warm-up on whether assistants can even read your site, see our Academy Lab on whether AI search crawlers can read your site; for how assistants choose which pages to cite, see how AI assistants decide which pages to cite. You need one page of your website, a text editor and about 45 minutes.
Every Thursday we send The Revenue Signal, our weekly newsletter: one study or how-to, with the numbers and the sources. Subscribe here: https://revenueexperts.beehiiv.com/
Why do adjectives give an assistant nothing to quote?

An AI assistant does not read your page top to bottom. It cuts the page into passages of a few hundred words, picks the passage that best matches the buyer’s question, and lifts a sentence from it. “World-class service” and “trusted by leading companies” survive that cut, but they carry no information, so there is nothing in them the answer can repeat with confidence. A sentence such as “Support by phone, Monday to Friday 8:00 to 18:00 Central time, median first reply 40 minutes in Q3 2026” can be repeated, checked and credited.
Grade each claim in two passes. First the evidence check: does the page carry proof that comes from you, such as your own numbers, a named case, a documented method or concrete specifications? Mark it clear, partial or missing. Then the claim-support check: does that proof sit next to the claim it supports? Mark it clear, partial or missing again. The two are separate because a page can have strong proof in a “Results” block at the bottom and bare claims in the banner at the top. To the assistant, those are two unrelated passages.
Google asks for the same thing in its guidance for site owners. Its self-assessment questions include whether your content offers “original information, reporting, research, or analysis”, and whether it earns trust through “clear sourcing, evidence of the expertise involved” [2]. None of this needs new markup or a developer.
What changed this week in how ChatGPT counts mentions?

On 10 Nov, OpenAI opened Businesses in ChatGPT in beta. A business that claims its listing gets a Business Hub showing how often ChatGPT mentioned it, how many clicks went to its website, and “discovery topics”, the themes of conversations where its site was cited [3]. OpenAI’s own tooltip warns that a recorded click does not confirm a website visit or a conversion [3].
A count is useful, and it will start conversations in many leadership meetings. What it will not show is which sentence on your page the answer used, or why a competitor’s page was used instead. That part stays in your hands, and it is exactly what the two checks look at. A second signal from the same fortnight points the same way: a research note from an AI search analytics vendor traced the statistics quoted in AI answers and found that fewer than half led to a study saying the same thing [4]. A number published with its source and date in the same sentence is the rarer and safer kind.
What are the five steps on one page?

Pick the one page your buyers land on most from search: usually a service, product or pricing page. Copy its visible text into a plain text document. Keep the browser tab open beside it. You will edit the text document first and move the changes to the site at the end.
Step 1. Find every claim a buyer might doubt (10 minutes)

Search the text for these words: leading, best, #1, world-class, seamless, trusted, thousands, millions, proven, cutting-edge, industry. Mark every hit. Then read the page once more and mark the claims those words missed: anything about speed, savings, reliability, popularity or expertise. A useful test: if a sentence could appear unchanged on a competitor’s site, it is not evidence.
Do not fix anything yet. Write each marked sentence on its own line in a list. That list is your work for the next two steps.
Step 2. Write the fact behind each claim (15 minutes)

For every line on the list, ask one question: what fact is behind this? There are four kinds of first-party evidence, and a service company without public client names can still pass the evidence check on the last two.

| Kind of evidence | Weak (clear it or back it) | Strong (example, made up) |
|---|---|---|
| Your own numbers, with source and date | “Thousands of happy customers” | “412 active client accounts on 30 September 2026, from our billing system.” |
| A case with an outcome and a date | “A leading manufacturer loves our service” | “A 140-person metal fabricator moved its quoting to us in March 2026; quotes now go out in one day instead of four.” |
| Your documented method | “A proven, rigorous approach” | “Every project runs the same 12-step checklist, and we send the client the finished sheet.” |
| Concrete specifications | “Enterprise-grade security, fast support” | “Data hosted in Frankfurt, Germany; support by email and phone, Monday to Friday 8:00 to 18:00 Eastern time.” |
Table 1. Four kinds of first-party evidence the evidence check looks for. The strong examples are made up to show the shape; write your own facts.
Four rules keep your numbers honest. Use the real figure: “412” beats “over 400”. Give the date or period, because an undated number becomes wrong without anyone noticing. Give the sample size for any survey. And never invent a number: if you have no data, describe what you do, and start counting this quarter. If you cannot name a client, make the case specific (size, industry, what changed, when); specificity moves the label towards clear even without a name.
If a claim has no fact behind it at all, delete it. Removing “world-class” costs you nothing, because it was never quotable.
Step 3. Move each fact next to its claim (10 minutes)

This is the claim-support check, and it is the fastest fix on the page. Put the evidence in the same sentence as the claim, or in the sentence straight after it. “We close month-end faster” followed, three screens later, by a survey figure is only partly supported. “We close month-end faster: in our 2026 survey of 48 clients, the median close fell from 9 days to 5” is clearly supported, because one passage now carries both.
Four page layouts break the pairing most often:
A banner with bold claims and the proof in a “Results” section far below. Repeat the key number in the banner sentence, or move one case up under it.
Benefits as three icons with one-line labels. Add one evidence sentence under each label.
A statistics bar (“400+ clients, 99.9% uptime, 5-star rated”) floating on its own. Turn each figure into a sentence with its source and date, placed beside the claim it supports.
Claims on the service page and proof only on a separate case studies page. Keep at least one evidence sentence per claim on the page itself and link to the full case. An assistant reads the page it lands on, not your whole site.
Step 4. Name outside sources inside the sentence (5 minutes)

When a claim rests on someone else’s research, name the publisher and the year inside the sentence and link that name, so the source travels with the retrieved passage instead of sitting in a footnote list at the foot of the page. Replace “studies show”, “experts agree” and “research suggests” with the real publisher, or delete them. These phrases borrow authority and present none.
One note. Outside statistics help the claim-support check, because they connect a claim to support. They do not count as first-party evidence, which wants proof that comes from you.
Step 5. Keep your proof near the top (5 minutes)

Assistants, and the buyers reading your page, start at the top. Put your strongest proof in the first screen or two, not only in a “Results” block at the foot of the page. Highlight every sentence that holds a number with its source and date, a case with an outcome, a described method or a specification. Two or more such sentences that support the page’s main claims are a good sign. Text inside images, infographics and dashboard screenshots cannot be read as text, so write the number in a sentence beside any graphic you keep.
Then publish the edits. A week later, ask the same buying question in ChatGPT, Claude and Perplexity and see whether your new sentence is the one they quote.
What does a page that gets cited look like?

Here is one paragraph before and after, from a made-up logistics software company.

| Before (evidence missing, support missing) | After (evidence clear, support clear) |
|---|---|
| “FleetDesk is the leading dispatch platform for mid-size carriers. Trusted by hundreds of fleets, our seamless tools save you hours every week.” | “FleetDesk is dispatch software for carriers with 20 to 300 trucks. 186 carriers ran dispatch on it in October 2026, from our account records. In our May 2026 survey of 54 dispatchers, the median time to plan the next day’s loads fell from 95 minutes to 40.” |
Table 2. The same paragraph before and after. Company, figures and survey are made up to show the shape.
The after version is not longer by much, and it does three things the before version cannot: it says who the product is for, it gives a number with its source and date, and it puts the proof inside the claim. An assistant can lift any of its sentences and the sentence still makes sense on its own.
What are the common mistakes?

Fixing only the first claim on the page. The page stays only partly supported when other major claims stand alone.
Rounding up. “Over 500” when the real count is 412 is a claim a buyer can later catch.
Moving all proof into one testimonial carousel. Sliders often load late or as images, and a quote without an outcome is weak proof.
Using a logo wall as proof. Images are not quotable; add one sentence naming the customers if you have permission.
Quoting a big outside statistic in place of your own evidence. It helps the claim, but it is not your proof.
What is your plan for this week?

Today: copy one page’s text and run Step 1 (10 minutes).
Tomorrow: write the facts for each marked claim; ask finance or operations for the two numbers you do not have (Step 2).
Before Friday: pair every claim with its fact, name outside sources in the sentence, keep your proof near the top, then publish the edits (Steps 3 to 5).
Short on time? Do Step 3 alone. Moving existing proof next to its claim takes minutes and can move a claim from partly supported to clearly supported.
Do not spend money yet on new content for this. The fix is in the sentences you already have.
How Revenue Experts helps

You can run both checks yourself with the steps above. If you would rather see the result, our AI citation audit shows what ChatGPT, Claude and Perplexity actually tell your buyers today. It costs $1,495 and takes 2 to 5 days: 50 buyer questions, three assistants, three runs each, 450 saved answers, how often you are named and whose pages are cited instead. You get the findings, then a short call with John Bush to work through them. Everything we build can be checked.
To scope one, book a 20-minute call with John Bush, or tell us about your page through our contact page.
Which claim on your best page would you be least comfortable defending in front of a buyer who asks, “Where does that number come from?”
If this was useful, the next one arrives on Thursday. Subscribe to The Revenue Signal: https://revenueexperts.beehiiv.com/
FAQ

We cannot name any clients. Can the page still pass the evidence check? Yes. Your own dated numbers, a documented method and concrete specifications all count as first-party evidence.
Do outside statistics count as first-party evidence? No. They help the claim-support check when they sit in the same sentence as the claim, but the evidence check wants proof that comes from you.
Will this get our page cited? Nobody can promise that, and we do not. It removes the most common reason a page gives an assistant nothing to quote. What the assistants then say is a separate measurement, which we make with saved answers.
Does the new ChatGPT business listing replace this work? No. It counts mentions and clicks [3]. It does not edit your page or show which sentence was used.
Related
- AI visibility: is your company the answer when buyers ask AI?
- Academy Lab: can AI search crawlers read your site?
- The AI Visibility Prompt Playbook
- Book a 20-minute call with John Bush about an AI citation audit
Sources
[1] Revenue Experts, “Who do AI assistants cite when buyers ask which field service software to buy?” (18 vendors, 12 questions, 108 answers, 1,493 citations), 3 Oct 2026. https://revenueexperts.ai/ai-assistants-field-service-software-citations-study/ (opened 14 Nov 2026)
[2] Google Search Central, “Creating helpful, reliable, people-first content”. https://developers.google.com/search/docs/fundamentals/creating-helpful-content (opened 14 Nov 2026; no change date shown on the page)
[3] Search Engine Roundtable (Barry Schwartz), “Businesses in ChatGPT Business Hub Live With Click Data”, 10 Nov 2026, quoting OpenAI’s Business Hub and its click tooltip. https://www.seroundtable.com/chatgpt-business-hub-42505.html (opened 14 Nov 2026)
[4] Trakkr Research (Mack Grenfell), “Zombie Stats”, traced 1,285 statistics quoted in AI answers; fewer than half reached a study saying the same thing, 28 Oct 2026; Trakkr sells AI search analytics software. https://trakkr.ai/trakkr-research/zombie-stats (opened 14 Nov 2026; used as a signal only, no numbers quoted in the body)