Which Sources Do AI Assistants Choose, And What Does The Research Say?

A Search Happens Before The Answer Is Written

In AI-assisted search, the answer usually does not come from the model’s memory alone. The system first runs a search, gathers a few pages and writes the answer by drawing on them. So being named is a two-step event: first entering that small pool of sources, then actually being used in the text that gets written. Getting into the pool is not enough on its own; some pages sit inside it and never appear in the answer.

The paper that gave this field the name GEO was published in 2023 by Aggarwal and colleagues and presented at KDD 2024¹. The authors stress that generative search engines leave content owners very little say over whether their content appears in the answer. In a test bed built from queries across different domains, they found that changes in how content is presented could raise its visibility in answers by up to 40%, and that what works differs from domain to domain¹.

That last detail may be the most instructive part for a business. A firm that makes stainless steel valves and an online store selling cosmetics are not playing the same game. For “what pressure can this valve take,” the assistant leans on technical documents; for “which moisturiser for sensitive skin,” it turns to comparisons and user experiences. There is no single universal rule for source selection; the type of question decides the type of evidence sought.

Citing Is Not The Same As Citing Correctly

In 2023, Liu, Zhang and Liang examined four generative search engines with human evaluators: Bing Chat, NeevaAI, Perplexity and YouChat². The answers were fluent and looked informative, but only 51.5% of generated sentences were fully supported by citations, and only 74.5% of citations actually supported the sentence they were attached to².

Two years later, the Tow Center team at the Columbia Journalism Review asked a similar question using news content³. Eight AI search tools were given excerpts from articles chosen from 20 publishers and asked to find the source article; across 1,600 queries, the answer was wrong more than 60% of the time³. The gap between tools was wide: Perplexity was wrong on 37% of queries and Grok 3 on 94%, and 154 of Grok 3’s 200 citations led to error pages³. Perhaps the report’s most striking observation was that the tools rarely said when they were unsure, delivering wrong answers in the same confident tone³.

For a business, this means your name can appear next to a sentence you never said, or a text you wrote can be summarised with the link handed to a directory site. The aftercare notes a dental clinic writes for its patients might form the backbone of an answer and never show up in the citation. Being named and being named correctly are separate questions, and each needs to be watched on its own.

Your Own Voice Carries Less Weight Than Others’

A 2025 study by Chen and colleagues compared AI search services such as ChatGPT, Perplexity and Gemini with classic Google search⁴. Their clearest finding is that AI search shows a systematic and strong preference for earned media, meaning third-party sources independent of the brand, over a brand’s own content and social media⁴.

The study also shows the services behave differently from one another: there are marked differences in source diversity, freshness of content, consistency across languages and sensitivity to how a question is phrased⁴. Asking the same question in Turkish and in English, or changing a few of its words, can be enough for other brands to appear on the list. The authors also point out that big brands start these systems with an advantage⁴.

Picture a company building bookkeeping software for small businesses. On its own site it describes its product as the easiest solution; the assistant, meanwhile, listens to an industry comparison article, a professional association’s review or user forums. A party testifying in its own case is heard, but does not weigh as much as a neutral witness. Assistants weigh things roughly the same way.

The Link To Classic Search Has Not Been Cut

An analysis Ahrefs published in summer 2025 examined 1.9 million citations in 1 million Google AI Overviews⁵. Of the cited pages, 76.10% were in Google’s top 10 results for the same query, 9.50% sat between positions 11 and 100, and 14.40% did not appear in the top 100 at all⁵. It is worth noting this is an industry analysis by a tool vendor; but the sample is large and the method is clearly described.

This picture says that, at least for Google’s own summaries, the source pool is fed largely by ordinary search results. SEO is therefore not an outdated field; it is the entry gate on the road to AI answers. On the other hand, ranking at the top does not explain everything: a not insignificant share of cited pages do not appear in classic rankings at all⁵.

Picture an education provider. If its exam preparation page sits on the second page of search results, its chances of being named in an AI summary drop accordingly. But if there is a long, detailed discussion about the same provider on an independent parents’ forum, its name can enter the answer through another door. The two routes do not exclude each other; most of the time they are two faces of the same reputation.

Being Named Does Not Mean Being Clicked

In March 2025, Pew Research Center examined the browsing data of 900 US adults and 68,879 Google searches; 18% of those searches produced an AI summary⁶. On pages with a summary, users clicked a classic result link in 8% of visits, against 15% on pages without one; clicks on links inside the summary itself came to just 1% of visits⁶.

The same study looks at how sources are distributed. Wikipedia, YouTube and Reddit together make up 15% of sources in AI summaries and 17% in classic results; government sites account for 6% in summaries and 2% in classic results⁶. Long, question-style searches trigger summaries far more often: 53% of searches of 10 words or more produced one⁶. The data is from the US; user behaviour in Türkiye has not been measured separately, but it gives a sense of direction.

Value is shifting from the click to the mention. Someone who sees a name in an assistant’s answer for a locksmith or an air-conditioning service may simply pick up the phone without ever visiting the site. That visit never shows up in the analytics panel. A measurement that looks only at traffic misses most of this new kind of visibility.

The Selection Process Is Fragile, And So Are Shortcuts

Kumar and Lakkaraju studied language models’ product recommendations using a catalogue of fictitious coffee machines⁷. They showed that when a specially generated text sequence was added to a product page, even a rarely recommended product became markedly more likely to be the top recommendation⁷. The authors also warn plainly that this could undermine fair competition⁷.

Two lessons follow from this finding. First, these systems are less stable than they look; a small intervention can change the ranking. Second, gaps of exactly this kind are what platforms work to close. Visibility that leans on a gap lasts as long as the gap; visibility built on what independent sources say about you lasts as long as your reputation.

What All This Tells A Business

Put side by side, the studies reveal a chain: being found in classic search, entering the answer’s source pool, being judged credible by independent voices, and being summarised correctly. Which link in the chain is weak varies from sector to sector¹⁴. For an online store, comparison articles may be decisive, while for an industrial manufacturer technical documents and trade publications carry more weight.

The answer to “why does the assistant name our competitor and not us” is usually hidden in the sources the assistant shows. Who those sources are, and how they talk about your brand or do not talk about it at all, says far more than a rankings report. That is what GEO is really about: reading not a single page, but the whole trace a brand leaves online.

None of these studies offers a formula that says in advance which brand will be named. What they offer is more modest but more useful: an understanding of which question gets answered by looking at which sources. A business owner who sees this picture finds it easier to grasp that visibility works less like a line in an advertising budget and more like an account of trust built up over years.