The one thing on a page that makes AI cite it
One edit raised a page's share of a machine written answer by 41 percent across 10,000 test questions, in the GEO study by Aggarwal and coauthors, and the edit was not a keyword. It was a quote from a credible source. Numbers with a source came next. Stuffing in keywords was the only edit that made a page lose ground.
You run a service business. You do not care how ChatGPT works. You care that when a customer asks it who to call, your page is the one it reads from. This is what makes that happen, with the numbers behind it.
What the test measured
A research team rewrote pages 9 ways and watched which version the answer engine used more.
Aggarwal and coauthors built a test set called GEO-bench with 10,000 questions across 25 topic areas, as reported in the GEO study. They took the pages an answer engine drew from and rewrote each one a single way. Then they asked a test engine built on a GPT model the same question again. They measured how much of the answer came from the changed page.
| Edit made to the page | Change in share of answer | Change in rater impression |
|---|---|---|
| Add quotations | +41.0% | +28.0% |
| Add statistics | +29.2% | +22.8% |
| Cite sources | +27.0% | +13.5% |
| Rewrite for fluency | +27.0% | +13.5% |
| Add technical terms | +16.7% | +10.9% |
| Make easier to read | +13.5% | +6.2% |
| Authoritative tone | +9.8% | +18.7% |
| Unique words | +6.7% | +6.2% |
| Keyword stuffing | minus 8.7% | +4.7% |
Basis: Aggarwal and coauthors, GEO study, GEO-bench, 10,000 questions across 25 domains, Table 1. Share of answer is their position adjusted word count. Rater impression is their subjective impression score. The same edits run on the live Perplexity.ai engine gained up to 37 percent in the same study.
Quotes won the test
A quote from a named source beat every other edit.
Adding quotations from credible sources lifted a page's share of the answer by 41.0 percent and its rater score by 28.0 percent, in the GEO study by Aggarwal and coauthors. Adding statistics came second at 29.2 percent and 22.8 percent in the same test. Citing sources came third at 27.0 percent and 13.5 percent, tied with a plain fluency rewrite. Tone words, fancy terms and rare words all landed under 17 percent in the same table.
Keyword stuffing was the only edit that hurt, cutting the page's share by 8.7 percent in the GEO study. The pattern is simple. Claims that can be checked win. Claims that just sound good do not.
That is the one thing. Put a checkable claim on the page. A quote with a name on it, a number with a source on it, or a link to where the number came from. Say who measured it and on what, in the same sentence, the way this article does.
Google says no tricks
Google's own docs say there is no special file or markup for its answer features.
Google Search Central states there are 0 extra requirements to show up in AI Overviews or AI Mode, only that the page is indexed and eligible for a snippet. Google also says no special schema markup or AI text file is needed. So anyone selling you a magic file for this is selling you air.
Google does say its answer features use query fan-out, running multiple related searches on subtopics to build 1 response, per Google Search Central. That means your page can get pulled for a small question inside a big one. A page that answers 1 clear question with a checkable fact is easier to pull than a page that vaguely covers 10.
Google Search Central also asks for 1 visible date labeled Published or Last updated, matched to the page's structured data, and says to keep other dates off the page. Keep that date honest. The study Do Large Language Models Favor Recent Content fed 7 language models identical passages with different dates and found a newer date moved a passage as many as 95 ranks and flipped preference between equal passages by up to 25 percent on average. That is a bias in the model, not a judgment of quality, so do not build your plan on it.
Being cited pays little
The click you get from a cited answer is small, so the page has to carry the sale itself.
Pew Research Center tracked 900 U.S. adults across 68,879 Google searches and found 18 percent of searches showed an AI summary. When a summary appeared, people clicked a regular result on 8 percent of visits, against 15 percent when no summary appeared, in the same Pew study. Only 1 percent of visits clicked a link inside the summary itself.
Question searches with who, what, when or why triggered a summary 60 percent of the time, and searches of 10 or more words did 53 percent of the time, per Pew Research Center. Your customers ask questions. So the summary is where they meet you. Make sure what it repeats is your number, your name and your town.
Machines misquote pages
Getting cited does not mean getting quoted right.
Liu, Zhang and Liang at Stanford audited answers from 4 answer engines and found only 51.5 percent of sentences were fully backed by the pages cited, and only 74.5 percent of citations backed the sentence they sat under. The Tow Center at Columbia ran 1,600 queries through 8 answer tools asking them to name the source of a news excerpt, and more than 60 percent of answers were wrong. Perplexity missed 37 percent and Grok 3 missed 94 percent in that Tow Center test. More than half of Gemini and Grok 3 answers pointed to fake or broken links in the same test.
Allaham and Diakopoulos audited 4 engines on 712 real questions and found about 16 percent of cited sources looked machine written. Each engine leaned on a narrow set of sites and cited the rest once. So a page with a plain, checkable fact stands out from a pile of vague filler.
How to write the page
One question, one checkable answer, one named source, near the top.
Pick 1 question a customer really asks, like what a water heater swap costs in your county. Answer it in the first 2 sentences with a number and say where the number came from, such as your last 40 jobs. Add a quote from a real person with a name, like your lead tech or a supplier. Put a link to any outside source you lean on.
Then stop. Do not stuff the town name in 30 times, since keyword stuffing cut share by 8.7 percent in the GEO study by Aggarwal and coauthors. Keep 1 labeled date on the page, as Google Search Central asks. Give the page a plain title that matches the question.
What we got wrong
I set out to prove one trick and the evidence gave me a pattern instead.
I wanted to tell you the one thing was a number or a fresh date. The only controlled test I trust, the GEO study by Aggarwal and coauthors, put quotations first at 41.0 percent and statistics second at 29.2 percent. So the honest answer is that checkable claims win, and no single kind of claim is the whole story. Google says in writing there is no special trick, and I take that at face value.
I also have to admit the prize is small. Pew Research Center found 1 percent of visits clicked a link inside a summary. And the GEO numbers came from a test engine and Perplexity, judged by raters, not from counting live citations today. Treat them as lift in a lab, not a promise.
Do it yourself
The method is free and it fits on one line.
Take your best page and add 1 quote with a name, 1 number with its source and 1 link, then cut the fluff. That is the whole method and it costs nothing. If you would rather not do it yourself, the VuseDesk website and SEO install builds the page this way for you.
Or call (252) 666-7217 and ask the 1 question this article did not answer. Email [email protected] if you would rather write it down.