Citation Sets: Which Sources Shape AI Answers

Anyone choosing software today increasingly starts with a question to ChatGPT, Gemini, Perplexity or Copilot: Which vendors are there, and which one fits us? Google's AI Overviews and AI Mode also answer right above the search results. The shortlist takes shape in the answer window, often without anyone opening the pages it comes from.

Two numbers make this measurable. On Google results pages that carried an AI summary, only 8 percent of visits led to a click on one of the familiar results, on pages without one the figure was 15 percent. The numbers come from the Pew Research Center, which analyzed real browsing data in 2025.

For media relations, this has a surprising consequence. A piece in a specialist magazine might find fewer human readers than its print run implies, yet machines read it closely and reuse it in their answers. Independent coverage, earned media, gains value because answer engines draw much of their material from independent sources. The open question is which pieces actually work and how to check.

Between question and answer: the short path to a source

Every product handles the details differently, but the core is the same: for a specific question, the system searches the web or its own index, reads a handful of results and writes its answer from them. Anchoring an answer in retrieved text is called grounding. It lowers the risk of invented claims without removing it, as our story What AI Summaries Leave Out, and When It Matters shows.

This story goes one step further: to the sources that prevail across many answers. Ask the same kind of question again and again and a pattern emerges. The citations change, but not at random. Some addresses show up constantly, others never.

Citation sets: every topic has its regular sources

Ask twenty questions about one software category, say inventory management for online retailers, and note the citations. A compact set quickly takes shape: one or two industry portals, a comparison site, an association with guides, a reference work, plus a few vendor websites. We call it the citation set here. For another topic it looks completely different.

What matters is who belongs to it. Small trade titles are often cited for specific questions because they answer exactly those questions. Large general outlets with generic coverage rarely appear, despite many times the reach. Reach and domain authority therefore say little about whether a title plays a role in AI answers.

The citation set also depends on the type of question:

• Questions on industry news, such as “What is changing in warehouse logistics right now?”, rest chiefly on news outlets and the trade press.
• How-to questions, such as “How do I move my shop from spreadsheets to inventory software?”, lean more on reference works and on what companies publish themselves: websites, guides, documentation.

Earned media and owned media therefore serve different questions, and both belong in the plan. The same principle holds inside a company, as Support Chatbots Can Only Repeat What the Help Center Says explains: what is not written down cannot be answered.

Freshness: when the date has a say

A second pattern concerns time. For questions about current developments, opinions in the market or events, answer engines favor fresh content. An interview from three years ago helps little when someone wants to know which vendors are in demand right now.

Older pieces still keep their value: thorough features, profiles and case studies are the foundation for answers to fundamental questions. What is new is the weighting over time. A single big hit fades, regular coverage keeps a brand present.

One brand, one description

Before an answer engine recommends a brand, it has to understand what the brand is. To do so, it collects descriptions from press texts, the company website, team bios, directory entries and platform profiles. If they match, a clear picture emerges. If they contradict each other, the system becomes unsure and may borrow a competitor's way of describing the category or repeat an outdated claim.

A usable article therefore answers a real buyer question and, at the same time, the question of who the company is. That takes four elements:

• Name, always spelled the same way.
• Category, meaning the offer in the words buyers use too.
• Audience, meaning who the offer is for.
• One proof point, plain and concrete: a feature, a field of use, a verifiable result.

A quote like “We are excited about the partnership” gives a language model nothing to work with. A sentence like “The software syncs stock levels and the online shop in real time, without retailers writing any code” does.

An answer engine only recommends what it can clearly place.

Answer variance: why a screenshot proves nothing

Ask a question twice and you rarely get the same text twice. Language models generate answers word by word based on probabilities. On top of that, search results change constantly, wording, location and settings make a difference, and new model versions arrive. The same question names three vendors today, four tomorrow, and the citations change with them.

A single answer is therefore a snapshot, not a measurement. The screenshot with your own brand at the top proves as little as the one where it is missing. A finding only becomes reliable when many questions are asked repeatedly and compared over months. Visibility in AI answers is a frequency, not a fixed position.

Measuring by sampling: question set and monthly log

The familiar PR metrics remain useful: reach and impressions, share of voice in the media, message pull-through, the number of placements. They do not show whether coverage shapes AI answers. The additional layer can be built without a big budget:

• Question set: 25 to 50 questions that real buyers ask, spread over three phases: getting oriented, comparing, choosing. Write them yourself, based on conversations with sales and customers.
• Rhythm: Once a month, run the full set through the major assistants: ChatGPT, Gemini, Perplexity, Copilot and AI Overviews in Google.
• Record: Note per answer: named or not, position, cited sources, whether the description is right, and the competitors mentioned with you.
• Evaluation: What counts is the development across several months, not any single run.

A spreadsheet is enough to start, specialized tools pay off as the question set, markets and languages grow. The most valuable output is often the source list: it shows which titles belong to the citation set of your own topic.

A fictional example: inventory software from Hanover

Take a fictional vendor: Kontorix from Hanover, inventory management software for mid-sized online retailers. The marketing team writes down 30 questions, from “When does inventory software start to pay off?” to “Kontorix or a spreadsheet, which suits a shop with three warehouses?”, and asks them in five systems in the first month.

The first log shows three things. Kontorix appears almost only in choosing questions that already contain its name. The citations are mainly two trade portals for e-commerce logistics and a retail association's guide, while a big founder interview in a business paper appears nowhere. And one system takes Kontorix for point-of-sale software, because an old directory entry still says so.

The team then aligns the description on the website, in profiles and in directories: “Kontorix from Hanover is inventory management software for mid-sized online retailers and syncs stock and shop in real time.” The media list is re-sorted by citation frequency. New articles answer the orientation questions where Kontorix is missing, spread across the year.

Whether a system names Kontorix more often afterwards, nobody can predict. But the log shows whether the wrong description disappears, whether the new articles turn up among the citations and in which question phase something moves.

What you can now ask of your PR

This adds up to a short self-check. Rephrase the points for your company and answer them with yes, no or don't know:

• Do we run the same buyer questions on a regular schedule?
• Does our brand appear in category and comparison answers, with our recent coverage cited as a source?
• Is our media list sorted by citation frequency rather than by reach?
• Does every article carry our name, what we offer and one concrete proof point, and do such articles appear regularly?
• Is the picture AI systems give of us accurate, and do we get outdated claims fixed?
• Do we know the titles cited for our competitors, even ones nobody on our team has ever contacted?
• Does our reporting connect placements with AI visibility?

If no and don't know dominate, your media relations are not yet reaching the places where buyers now prepare their decisions. Ask your PR partner to back story ideas with citation data, to build stories around the questions buyers put to AI, and to show in reporting whether the brand's position in the answers changes.

This is where pr4geo comes in, like axaix a service of NexTao GmbH. pr4geo places PR articles and other formats across the network's magazines and portals, more than a hundred of them, chosen by topic. The editorial team produces them, you approve every text before publication, and live reporting documents every placement. Every package comes with multilingual PR articles, frontlinks as citable core sentences, a personal teaser.md and six tools.

The Starter package, with 3 PR articles and 24 placements a year, is €99 monthly or €990 annually. Premium brings 12 PR articles, 96 placements and a promo video for every article, at €290 monthly or €2,990 annually. Exclusive, at €990 monthly or €9,990 annually, adds to Premium, among other things, entries in the askiana chatbot and an ad-free landing page that is optimized for AI. The term is twelve months, all prices plus VAT.

No one can promise that ChatGPT or any other system will name a company, pr4geo included. What is promised are the agreed placements, and whether a title belongs to the citation set of your topic is something your own measurement shows. You can request an initial call via pr4geo.com at any time.