AI agents can favor some sources over a better fit

The question

Do AI agents prefer items from some sites even when another item fits better?

The short answer

A new study suggests they can. Across 12 agent models, an item meeting one fewer requirement was picked about two-thirds of the time when it came from a preferred source and the better item came from an avoided one.

Source: Source Preference in the Wild, arXiv, Conversational Capture, arXiv

TL;DR: A new study suggests AI agents prefer items from certain sites, and that preference can beat fit: an item meeting one fewer requirement was still picked about two-thirds of the time (median across models and domains) when it came from a preferred site and the better item came from an avoided one. Hiding the source weakened the effect. Supplying missing information reduced it. My read for a business: put the facts an agent would compare on your own pages, then check how you're listed on the sites AI tools cite in your category.

Your item meets four requirements. A competitor's meets three. You assume an AI shopping agent will pick yours. It's an objective machine, after all. Says so in the pitch deck.

A new study suggests it often won't, when the competitor's item comes from a site the model prefers and yours comes from one it avoids.

What the study found

The paper is Source Preference in the Wild, posted to arXiv on 2 October 2026. The researchers ran 12 AI agent models through end-to-end search tasks in three areas: shopping, hotel booking, and picking research papers. They compared items from different sources that met the same requirements and were shown at the same position, so position couldn't explain the choice.

Here's what they report:

• Every model prefers some sources and avoids others, in every domain. The models largely agree on which. In shopping, all 12 preferred Walmart, but only two preferred Target. In hotels, ten preferred Booking.com, while half avoided Expedia.
• Preference can outweigh fit. The researchers compared pairs where one item met one fewer requirement. When that weaker item came from a preferred source and the better one came from an avoided source, the weaker item was picked 68% of the time (the median across models and domains, counting comparisons where exactly one item was chosen). Reverse the sources, and it was almost never picked.
• The label alone moves the choice. Hiding source information weakened the preference. Relabeling an item with a preferred source raised its selection rate.
• Supplying missing information reduced the preference. The authors test two explanations. Training may have taught models to use some sources as a shortcut for meeting the requirements, and missing information may trigger assumptions about a source. Supplying the missing information, or a prompt that counters those assumptions, reduced the preference.

The requests came from research benchmarks covering shopping, hotels, and scholarly papers, with the agents searching end to end. That's not the same as everyday ChatGPT or Gemini sessions. Note also that the study measures preference for sites, like marketplaces, not for product brands. My read: it's a reason to test this in your own category, not evidence that these preferences hold in your market.

A second paper: how early citations could compound

A separate paper, Conversational Capture, posted 30 September 2026, looks at multi-turn conversations. It argues that a source cited early becomes "substantially more likely to be cited again," partly because the system retrieves based on the conversation so far, and partly because the user's follow-up questions drift toward that source.

The paper supports this with theory and a model-derived illustration. It doesn't report measurements of real users or live engines. In their illustration, single-turn and multi-turn rankings agreed only weakly. The paper is listed for the HAI '26 conference proceedings. It doesn't show how often this happens in real customer searches.

My read: if you only test the first answer to one question, you may miss how a conversation unfolds. Test a short conversation too.

What I see in my own traces

I run buyer-style questions, like "best software for X," through five AI engines with web search turned on (from OpenAI, Anthropic, Google, xAI, and Perplexity) and record what each answer names and cites. Between 25 September and 9 October 2026 (ET), I recorded 233 successful answers across 15 runs. Each run tracked one business in its category. I chose these businesses for outreach research, not at random, so these are counts from that sample, not a rate for anyone else. Anyone selling you an "AI visibility score" to two decimal places from a sample this size deserves a follow-up question.

• The tracked business was named in 163 of 233 answers.
• Of those 163, 68 cited the business's own website as a source.
• 95 named the business without citing its site. In 33 of those, the model had pulled the business's page during its search and still didn't cite it.

These counts don't tell me why a page wasn't cited, or where the model got what it said. They do show that being named and having your own site cited are different results.

What to do about it

My read for a business. The first recommendation is motivated by the study's missing-information result. The rest are my own judgment. None is a tested intervention.

1. Close the information gaps. The authors suggest missing information can trigger assumptions about a source, and supplying it reduced the preference. Put the facts a buyer's agent would compare on the page, in plain text: price or price range, who it's for, specs, service area, delivery times, guarantees. Don't hide them behind a "contact us." Make them available without an inquiry form.
2. Make comparisons easy. Write pages that answer "X vs. Y" and "best for Z" honestly, including where you're not the right fit. An agent comparing requirements needs the requirements written down.
3. Check where AI tools already look. Run your buyer questions through a few AI tools and note which marketplaces, directories, and review sites get cited. Make sure any listings you have there are complete and match your site. Being cited often isn't proof a model prefers a site, so treat this as an audit, not a ranking trick.
4. Test conversations, not single answers. Start with one AI tool and five buyer questions. Run three fresh conversations per question, using the same two follow-ups each time. That's 15 conversations and 45 answers. Save every answer and log the tool and model if shown, the date, the exact question, the turn, whether you were named, whether your own site was cited, and which other sites were cited. Add more tools once that's routine.
5. Count mentions and citations separately. "Named in 4 of 5 answers, own site cited in 1 of those 4" describes your sample. "We're visible in AI" doesn't describe anything.

A quick self-check

• Could an AI agent compare my offer with a competitor's using only my page?
• Are price, fit, and key specs on the page as text, without a download or an inquiry form?
• Do I know which third-party sites AI tools cite in my category?
• Are my listings on those sites complete and consistent with my site?
• Do I track mentions and citations as two numbers, with the sample size?