What happened

When a customer puts a question to ChatGPT, it often adds a few web links under its answer to show which pages it drew on. That’s what an AI “citation” is. So what sets the pages that get cited apart from the ones that don’t? On 22 October, SISTRIX, a German company that makes search and AI visibility analysis tools, published a study that looked at exactly that.

It took pages cited in the second quarter of this year by ChatGPT and by the AI answers in Google Search. On Google’s side, that means the AI Overview at the top of the results and AI Mode, where you can go back and forth with the AI. It paired each one with a page from the same website that wasn’t cited, and compared them in pairs: more than 96,000 URLs in all. Comparing two pages from the same site keeps the site’s reputation out of the picture, so all that’s compared is the pages themselves.

The cited pages were more likely to have three things: a table of contents, a row of subheadings at the top that jump to each section when clicked; an author box with the author’s name, a photo and a line or two on their expertise; and a visible “last updated” date. SISTRIX combined these three with several other presentation features, eleven in all, and pages that met many of them were cited 2.1 times as often as barely optimised pages.

Two things need saying up front. First, the study only shows correlation: pages with these traits get cited more often, which doesn’t mean adding them gets you cited, and SISTRIX says as much itself. Second, all the data comes from AI answers in German-speaking countries (Germany, Austria and Switzerland).

Why it matters to you

Say you run a dive shop in Magong, on Penghu, and your website has two articles you wrote yourself. One, “Snorkelling in Penghu for the first time”, opens with a table of contents and is split into sections on which months are good for getting in the water, what to bring and what to do about seasickness; it ends with a byline from an instructor who has been leading groups for more than a decade, with his photo and a note that it was updated this May. The other explains how to get a diving certificate. It’s just as solid, but it’s one unbroken block of text, with no way to tell who wrote it or when. Going by the study’s figures, pages like the first one turn up more often among the sources listed under AI answers.

But add a table of contents to the second one and AI won’t necessarily start citing it. It’s a bit like busy restaurants: their menus tend to be clearly printed, but reprinting a menu nicely won’t by itself bring customers through the door. Our reading is that a table of contents, an author box and an update date are what a page naturally ends up with when someone tidies it up properly and keeps it current. Bolt a table of contents onto the top of that certificate article without anyone reworking the content or checking whether the exam rules are still right, and you’ve only touched the surface.

As for Chinese-language websites in Taiwan, nobody yet knows whether AI picks its sources the same way: we’ve looked, and as of 26 October we hadn’t seen anyone compare the pages themselves on Taiwanese Chinese-language sites by pairing pages from the same website. Still, ChatGPT and Google’s AI Mode, both covered by the study, are just as available to customers in Taiwan, and their answers come with sources listed underneath too.

What to do now

Act now. Research on Taiwan hasn’t come out yet, but AI won’t wait for it: customers ask it questions every day, and every day it picks sources for them. Just don’t treat this study as a checklist to work through item by item.

Working through it item by item is the easiest way to go in the wrong direction. As we said, the study only shows correlation, and the same feature can cut the opposite way on a different kind of site. Take the update date: on editorial sites (news and magazine sites, say) it has a clear link to being cited, but on online shops, pages with a date were actually cited less often. That doesn’t mean AI penalises online shops for showing dates: the study shows the pattern, not the reason. SISTRIX doesn’t tell online shops to remove dates either; its advice is to open a separate section of the website with practical articles for customers, which it thinks holds more promise than working on product pages; for a shop selling food, that might mean articles on how to choose and store things.

What kind of site you have and what your customers mostly ask decide where to start; and once the work is done, whether AI has switched to citing your pages only shows in whose pages are listed under its answers. Those calls can be handed to our managed GEO service, where we do the whole job for you. GEO stands for generative engine optimisation, and the engines here are AI tools like ChatGPT that write answers on their own. With managed GEO, our team fixes your site’s content, brings what’s said about you elsewhere into line, and regularly checks how each AI engine describes you; you just review the results. What’s said about you elsewhere means places like Google Maps and the business listings on various platforms.

To see whose pages AI lists when it answers questions in your line of business right now, have one of our consultants open them up one by one at a demo. The demo is a session with one of our consultants where they show you live. The consultant puts a question a customer would ask to ChatGPT, such as “When’s the best month for a first snorkelling trip to Penghu?” for the dive shop, then opens each source listed under the answer to see whether any of your pages are there and which one it used. In the same session, the consultant will also show you your website’s GEO audit score. The GEO audit is geoweb.tw’s free website audit, which checks how well AI can understand your site and scores it across 12 areas.

Further reading