October 9, 2026 · ChimpanSEO

What did the new study find about source preference in AI agents?

Source preference in AI agents means a generative engine can pick a familiar source over a better fitting answer. A study posted to arXiv on 2 October 2026, titled Source Preference in the Wild by Song, Park, Shim, Song and Jo, finds that 12 agent models prefer some sources and avoid others in every domain tested, and that this pull can outweigh fit. For anyone doing SEO or GEO in 2026, being the most complete answer is no longer automatically enough. The lever a publisher controls is completeness, because the authors find that missing information triggers preconceptions about a source.

What did the new study find about source preference in AI agents?

A new arXiv paper studies source preference in AI agents across 12 models and finds that each model favors some sources and avoids others.

According to the paper Source Preference in the Wild, the authors Song, Park, Shim, Song and Jo ran end to end search tasks with 12 agent models across three domains: products, hotels, and papers to cite. Every model showed preferences in every domain, and the models largely agreed on which sources they favored.

Source identity works as a signal of its own. The authors report that hiding the information that identifies a source weakens the preference, while relabeling an item with a preferred source raises how often that item gets selected. The paper also tests two routes to the behavior: training that rewards better items can turn a source into a shortcut, and missing information can trigger preconceptions about the source. Supplying the missing information or adding a prompt that counters those preconceptions reduces the preference.

How can source preference outweigh a better fitting answer?

The study finds that an item from a preferred source wins roughly two thirds of the time even after missing one requirement.

That single result carries most of the weight in the paper. According to the authors, an item that satisfies one requirement fewer is selected about two thirds of the time when it comes from a preferred source and the better item comes from a dispreferred one.

Condition tested in the study	Selection behaviour reported
Item missing one requirement from a preferred source against a better item from a dispreferred source	Selected about two thirds of the time
Item missing one requirement from a dispreferred source against a better item from a preferred source	Reversal almost never occurs
Source identifying information hidden from the agent	Preference weakens
Item relabeled with a preferred source	Selection rate rises
Missing information supplied, or a counter prompt added	Preference reduces

Why does source preference matter for SEO and GEO?

Source preference in AI agents means an unfamiliar page must be more complete to win selection against a familiar source.

Our reading is that this moves part of the competition away from relevance alone. When an agent weighs candidate items, it appears to carry a prior about the source, and that prior can do work that page quality alone cannot undo. In generative engine optimization, a page can cover a request better than a competitor and still lose the slot inside Google AI Overviews, Perplexity or ChatGPT Search.

Because the models in the study largely agreed on which sources they favored, the effect looks consistent rather than random noise. That matters for AI search visibility, where a single weak signal rarely explains a lost citation. Our interpretation of the experimental result, not a measured effect on any ranking system:

Being the best answer is not enough on its own, because completeness competes with source familiarity.
Gaps on a page are where source preconceptions take over, following the mechanism the authors test.
Source identity is a signal agents react to, so consistent brand labeling across the web gains weight.
Agreement between models suggests the effect can show up across assistants, not just one.

What should publishers do today, and what should they avoid?

Publishers can weaken source preference by stating every fact a request may require in plain text, tables, and structured data.

What to do today

Audit key pages for missing requirement level facts: specifications, prices, availability, policies and dates. State each one explicitly in plain text.
Put the attributes users filter on into tables or structured data so an agent extracts values without guessing.
Test how AI assistants describe and choose your offer against well known competitors for the same request.
Track whether closing information gaps changes how often assistants cite you, using AI citation tracking.
Strengthen consistent brand and source identification across your site and third party mentions, so entity density and topical authority point to the same entity.

What to avoid

Do not claim the paper shows how Google, ChatGPT or Perplexity rank websites. It studies agent models in controlled end to end search tasks.
Do not name which sources were preferred unless the names come from the paper itself.
Do not invent percentages beyond the about two thirds figure stated in the abstract.
Do not relabel or misrepresent a source to exploit the bias.
Do not treat structured content for AI as a substitute for explicit facts on the page.

Frequently Asked Questions

Source preference in AI agents raises practical questions about citation, completeness, and how publishers should measure AI visibility in 2026.

Is source preference the same across different AI models?

According to the paper, the 12 agent models largely agree on which sources they prefer and which they avoid. That agreement points to a shared pattern in selection behaviour rather than random variation between individual assistants.

Does this study prove that some domains get ranked higher by Google or ChatGPT?

The paper studies how agent models select items in controlled end to end search tasks covering products, hotels and papers to cite. Read the findings as evidence about agent selection behaviour, and keep them separate from ranking rules for search engines.

How do I find missing facts on a page that an agent might need?

List every requirement a buyer or researcher would filter on for that topic, then check whether the page states each one in plain text. Prices, specifications, availability, policies and dates are the usual gaps, and tables or schema markup make the values easier to extract.

Can structured data reduce source preference?

Structured data helps extraction, and the mechanism described in the paper gives a reason to close information gaps anyway. When a page states every relevant detail, an agent has less reason to fall back on what it assumes about the source.

How should I measure whether these changes work?

Run the same prompts across several assistants before and after you update a page, log whether your domain gets cited, and compare results month over month. AI citation tracking turns an abstract preference into something you can observe on your own pages.

Our Reading: What Source Preference Means for GEO

Source preference in AI agents means completeness and clear source identity now matter as much as relevance in generative search results.

Our reading is that the paper describes a selection behaviour, not a ranking system, so source familiarity looks like a competitive force publishers cannot simply outrank. Completeness seems to be the most direct lever, since the authors link missing information to preconceptions about a source and show that supplying it reduces the preference. We expect teams that state every requirement level fact explicitly to earn more citations in 2026, while thinner pages keep losing to sources agents already treat as known. Start with one page, fill its information gaps, and rerun the same prompts a month later to see whether your citation rate moves.