Your website was written for a crawler. Short paragraphs, a keyword in the heading, a meta description of exactly the right length, a chirpy paragraph about your commitment to quality.
Something else reads that website now, and it has no use for most of what is on it.
An AI assistant like ChatGPT or Perplexity builds its answer out of sentences it pulls from other people's pages, then names where they came from. If your page has no sentence worth pulling, you are not losing the contest. You were never entered.
The best known research on what gets pulled comes down to three edits. Put real numbers on the page. Quote named humans. Cite where your facts come from.
In the GEO study from Princeton and partners, published at KDD 2024, adding quotations lifted a page's visibility on the researchers' test engine by around 41%. Adding statistics lifted it around 33%. Citing sources lifted it around 28%.
The same study tested keyword stuffing, the old SEO standby. It scored below doing nothing at all.
Almost no business website does any of the three.
Edit one. Put a number in it.
Open your main service page and count the numbers. Skip the phone number and the year you were founded. You are looking for numbers that describe what you do or what happens when you do it.
Most pages have zero. They have adjectives instead. Fast, reliable, affordable, experienced, trusted.
Adjectives are unquotable. A model cannot repeat "we are reliable" as a fact, because it is not one. It can repeat "we respond to commercial callouts within four business hours across the Central Coast" all day long, because that is a claim with an edge on it.
Go through your key pages and swap every adjective for a measurement. These are made-up examples, so use your own figures.
- "Fast turnaround" becomes "most jobs completed within five working days."
- Swap "affordable" for "typical projects run between four and nine thousand dollars."
- "Experienced" turns into "over 600 installations since 2014."
- Instead of "trusted", write "142 Google reviews at 4.9 stars."
An assistant can lift any of those word for word when it explains why it picked you. Leave the adjective in, and the model reaches for the competitor who was more specific.
One rule while you do it: the number has to be true. If you do not know your average turnaround, pull it from your job system or your invoices before you publish anything. An invented number is worse than an adjective, because a model may repeat it to a buyer, and the buyer will hold you to it.
A model cannot quote a feeling. It can only quote a fact. Every vague sentence on your site is a sentence that will never be read aloud to a buyer.
Edit two. Quote a human, by name.
Quotations were the strongest of the three edits in the study, at around 41%. My read on why: a quote in quotation marks with a name after it is a finished, attributable unit. The model can reuse it without rewriting a word.
You do not need to interview an industry analyst. Three sources are sitting in your business right now.
Yourself. A short, opinionated, named quote in the middle of a page. "We stopped offering the cheap option in 2023 because it was generating 80% of our warranty callbacks, says Kendall King, founder." That is quotable, attributable and human.
Your customers. Real testimonials with a full name and a specific outcome, published as text on the page. A testimonial trapped inside an image, or inside a third party widget that loads separately, may never be read at all.
Your team. The person who does the work usually has the most quotable sentence in the building, because they talk about the job and skip the marketing of the job.
Format matters as much as source. Put the quote in the page text, in quotation marks, with the full name and role straight after it. Keep it short enough to stand alone. "Great service, highly recommend" is technically a quote and tells nobody anything. "They rewired our kitchen in a day and left it cleaner than they found it" gives a model something to say about you.
Edit three. Cite where things came from.
This is the one businesses resist hardest, because linking out feels like sending people away. That instinct is a decade out of date.
In the study, the 28% lift came from adding citations to credible sources and changing nothing else. A sourced claim can be checked. An unsourced one is just you saying so.
So when you state an industry figure, say where it came from and link it. If you mention a standard or a warranty term, name the document. Tell the reader what your "typical result" is based on, even if the answer is "our last 40 jobs".
A page that reads like a well researched briefing gets treated as a source. A page that reads like a brochure gets treated as an ad, by people and, in our view, by assistants too.
The structure that makes extraction easy
Beyond the three edits, formatting decides whether a model can cleanly lift what you wrote.
Lead every page and section with a direct answer in the first two sentences. Skip the context and the warm up. Give the answer, then the explanation.
Use headings that are the questions people ask, word for word, then answer them immediately underneath. Somebody types "how much does a bathroom renovation cost on the Central Coast" into an assistant. A page with that exact heading and a dollar range under it is an easy match.
Keep statements self contained. "As mentioned above, this applies here too" means nothing once it is lifted out of the page. Repeat the subject. A human reader will find it slightly redundant, and a machine will find it complete.
Define your terms plainly, once, near the top. A clean one-sentence definition is easy to lift.
A worked example: the accountant's services page
Say you run a hypothetical accounting practice in Adelaide that works with tradies. Your main services page says you are "experienced, affordable and focused on helping small businesses thrive." Ask an assistant who helps tradies with tax in Adelaide and you are not mentioned.
Now apply the three edits. The adjectives become measurements you can verify from your own records: how many trade businesses you currently look after, your typical turnaround for a quarterly activity statement, and a fixed fee range for a standard sole trader return. A named quote goes in from you as the principal, explaining the single mistake you see tradies make most often with vehicle deductions and why it costs them. A second quote comes from a real client, by name, with a specific outcome in their own words. Then you cite properly. When you mention the instant asset write-off or record-keeping obligations, you name and link the relevant ATO guidance instead of paraphrasing it from memory.
Finally, the structure. The page opens with a two-sentence answer to "what does a tradie accountant do and what does it cost?" The headings become the questions clients ask on the phone, and each section's first line answers its heading directly. Nothing about the practice changed. The page now holds a dozen statements a model can lift and attach to your name.
A warning about overcorrecting
One caution, because this advice gets taken too far.
Writing for extraction does not mean writing like a database. Pages stuffed with disconnected facts and no argument read badly to humans, and humans are still the ones who decide to pay you.
You want a page that works twice. A person reads a clear, persuasive argument. A machine finds a dozen clean, quotable facts inside it.
The two rarely conflict. Specific writing with real numbers and named sources was always better writing. Being quoted by an assistant is a bonus for doing what you should have been doing anyway.
Do this on five pages, not fifty
You do not need to rewrite the site. Pick the five pages tied to revenue.
For each, spend an hour. Replace every adjective with a number. Add one named quote. Cite one external source properly. Rewrite the opening two sentences so they answer the question in the title. Convert the headings into the questions your customers ask.
That is five hours of work on the five pages that matter, while assistants are already telling buyers who to call in your industry.
Your competitors are still adding keywords to their meta descriptions. Let them.
How to tell whether it is working
Before you edit, write down five questions a buyer would ask an assistant about your service and your area. Ask them in ChatGPT, Perplexity, Gemini and Google's AI answers. Note who gets named and which sources get cited. Make your edits, then ask the same questions again every month.
Watch for your business starting to appear in answers, and for the description of you getting more specific. The best sign is your own pages showing up in the cited sources. Check your analytics for referral traffic from AI assistants as well. Assistants that search the live web can pick up an edit fairly quickly. Anything that depends on a model's training data takes longer, which is one more reason to start now. For why ranking and being cited no longer move together, see why ranking number one no longer gets you into the answer, and for the wider strategy behind these edits, read the GEO playbook for getting cited by ChatGPT, Gemini and Perplexity.
Make your best pages the ones machines quote
None of this requires a new website. It requires a sharper one, with pages that say specific, true things under a real name. If you want more on the craft of quotable writing itself, write content AI can't help but quote goes further. If you would rather hand it over, our digital marketing team rewrites revenue pages for humans and extraction at the same time, then tracks what the assistants start saying about you.
Frequently asked questions
How do I get my business quoted by ChatGPT?
Give it statements worth quoting. Replace vague adjectives on your key pages with specific, verifiable numbers, add short named quotes from you, your team and real customers, and cite the sources behind any industry claims. Open each page with a direct answer and use headings that match real customer questions. Then check monthly what assistants say about your business and category.
Does adding statistics to a page really help AI visibility?
In the best known test, yes. The GEO study published at KDD 2024 found adding statistics lifted visibility on the researchers' test engine by around 33%, alongside gains from cited sources and quotations. That engine was built by the researchers, so treat the figure as a direction and not a promise for ChatGPT. Only use statistics you can verify, ideally from your own records, and state what each number is based on.
Should I rewrite my whole website for AI search?
No. Start with the five pages most tied to revenue, usually your main service pages and the page that converts best. An hour per page applying the three edits and fixing the structure is enough to test the approach. Measure what assistants say before and after, then roll the method out to more pages once you can see it working.
Will writing for AI make my website worse for customers?
Not if you do it properly. Specific numbers, named quotes and cited sources make pages more persuasive to humans too. The risk is overcorrecting into a page of disconnected facts with no argument. Aim for a page that reads as a clear, confident case to a person while containing plenty of clean, quotable statements for a machine.