ChatGPT Search Can Pre-Select Brands Before Retrieval, Research Shows

研究显示：ChatGPT 搜索可在检索前预选品牌

ChatGPT’s in-chat search process can surface brand preferences before it has retrieved web results. Research tracking the system’s fan-out queries, the sub-queries used to investigate a prompt, found that vendor names sometimes appear at the opening stage. ChatGPT 的聊天内搜索过程可以在检索网页结果之前就显现出品牌偏好。一项追踪该系统“扇出查询”（fan-out queries，即用于调查提示词的子查询）的研究发现，供应商名称有时会在初始阶段就出现。

Separate experiments also found that the language of a query and the user’s exit location can materially change which commercial brands ChatGPT recommends. That does not mean every ChatGPT answer is predetermined, or that a brand mention is proof of a fixed ranking rule. It does mean that AI search visibility is shaped by more than the pages retrieved for a single query. For businesses trying to appear in product recommendations, category alignment, language and market context may influence whether the brand is considered at all. 独立的实验还发现，查询的语言和用户的出口位置（IP 所在地）可以实质性地改变 ChatGPT 推荐的商业品牌。这并不意味着 ChatGPT 的每个回答都是预先确定的，也不意味着品牌提及就是固定排名规则的证据。但这确实意味着 AI 搜索的可见性不仅仅由单次查询检索到的页面决定。对于试图出现在产品推荐中的企业而言，品类对齐、语言和市场背景可能会影响该品牌是否会被纳入考虑范围。

What the evidence says about brand selection in ChatGPT search

关于 ChatGPT 搜索中品牌选择的证据说明

Radyant’s open fan-out dataset, based on 615 buying questions and 3,842 ChatGPT fan-out runs collected from August 11 to 17, 2026, found that 28.6% of opening fan-out queries named a vendor. The reported confidence interval was 25.6% to 31.6%. In the remaining roughly 71.4% of runs, the opening query did not name a vendor. Radyant 的开放式扇出数据集基于 2026 年 8 月 11 日至 17 日期间收集的 615 个购买问题和 3,842 次 ChatGPT 扇出运行，结果显示 28.6% 的初始扇出查询中提到了供应商。报告的置信区间为 25.6% 至 31.6%。在其余约 71.4% 的运行中，初始查询并未提及供应商。

The pattern is more complex than a simple list of preferred brands. When a brand appeared in an opening fan-out query, multiple brands often appeared in the same run. Radyant also found that a brand could appear in a final answer without appearing in any sub-query in that run, at rates ranging from 9% to 29%. Fan-out visibility is therefore a useful signal, but not a complete explanation for why a business is named in an answer. 这种模式比简单的“首选品牌列表”要复杂得多。当一个品牌出现在初始扇出查询中时，通常会有多个品牌出现在同一次运行中。Radyant 还发现，一个品牌可能在最终回答中出现，但并未出现在该运行的任何子查询中，这种情况的发生率在 9% 到 29% 之间。因此，扇出可见性是一个有用的信号，但并不能完全解释为什么某家企业会被提及。

Independent B2B SaaS analyses have identified recurring fan-out patterns, including searches for an official brand page, documentation, domain-restricted pages and multi-brand comparisons. Those analyses repeatedly observed brands such as Braze, MoEngage, Iterable and Insider being selected in relevant category queries. The practical interpretation is that ChatGPT can draw on established brand-category associations while forming its search plan. 独立的 B2B SaaS 分析已经识别出重复出现的扇出模式，包括对官方品牌页面、文档、特定域名页面和多品牌比较的搜索。这些分析反复观察到，Braze、MoEngage、Iterable 和 Insider 等品牌在相关的品类查询中被选中。实际的解读是，ChatGPT 在制定搜索计划时，可以利用已建立的“品牌-品类”关联。

A separate arXiv study of query language and exit IP in commercial recommendations tested logged-out ChatGPT and API runs across languages and countries. Across 234 runs, the researchers found that top recommendations were unstable even for identical runs. They also found that query language and exit location acted as separate factors in brand selection: English queries tended to feature global brands, while other languages and local connections could favor domestic brands or replace global names altogether. 另一项关于商业推荐中查询语言和出口 IP 的 arXiv 研究测试了跨语言和跨国家的未登录 ChatGPT 和 API 运行。在 234 次运行中，研究人员发现，即使是相同的运行，顶级推荐结果也不稳定。他们还发现，查询语言和出口位置是品牌选择的独立因素：英语查询倾向于推荐全球品牌，而其他语言和本地连接则可能偏向于本土品牌，甚至完全取代全球品牌名称。

Observed condition | What the research found | Visibility implication

观察到的情况 | 研究发现 | 可见性影响

• Opening fan-out query names a vendor: 28.6% of Radyant’s tracked opening queries named a vendor. A brand can enter the search process before retrieved results are presented. 初始扇出查询提及供应商： Radyant 追踪的初始查询中有 28.6% 提及了供应商。品牌可以在检索结果呈现之前就进入搜索过程。
• Opening fan-out query names no vendor: About 71.4% of tracked opening queries did not name a vendor. Pre-selection is measurable but not universal. 初始扇出查询未提及供应商： 约 71.4% 的追踪查询未提及供应商。预选是可衡量的，但并非普遍存在。
• Language and exit location change: The arXiv study found that language and location strongly affected which brands were recommended. Testing one English-language result cannot represent every market. 语言和出口位置变化： arXiv 的研究发现，语言和位置强烈影响了推荐的品牌。测试单一的英语结果无法代表所有市场。

Why this is a visibility issue, not just a search-ranking issue

为什么这是一个可见性问题，而不仅仅是搜索排名问题

Traditional search optimization often starts with a results page and asks which sources rank. Generative search adds an earlier stage: the system may decide which concepts, sources or brands to investigate before it composes its answer. A company that is absent from that initial framing may still be found through retrieval or named in the final response, but it may have fewer routes into the answer. 传统的搜索优化通常从结果页面开始，询问哪些来源排名靠前。生成式搜索增加了一个更早的阶段：系统在撰写答案之前，可能会先决定调查哪些概念、来源或品牌。一家在初始框架中缺失的公司可能仍会通过检索被发现，或在最终回答中被提及，但它进入答案的路径可能会更少。

This matters particularly for commercial prompts such as requests for software recommendations, alternatives or the best provider in a category. In these contexts, a model’s existing association between a category and a known set of vendors can shape comparison queries. Retrieved pages remain relevant, but they operate alongside the model’s choice of what to look for. The evidence also argues against treating a single prompt as a reliable audit. Identical runs can produce different leading recommendations, and results can vary by language and exit location. A favorable mention in one test does not establish durable visibility. Equally, one missed mention does not prove a business is permanently excluded. 这对于商业提示词尤为重要，例如请求软件推荐、替代方案或寻找某品类中的最佳供应商。在这种背景下，模型在品类与已知供应商集合之间现有的关联可以塑造比较查询。检索到的页面仍然相关，但它们是与模型对“搜索什么”的选择共同起作用的。证据也表明，不应将单次提示词测试视为可靠的审计。相同的运行可能会产生不同的主要推荐，结果也会因语言和出口位置而异。一次测试中的正面提及并不能建立持久的可见性。同样，一次未被提及也不证明企业被永久排除在外。

How businesses can test and respond

企业如何进行测试与应对

The goal is not to try to reverse-engineer a hidden rule from a handful of chats. Instead, businesses should measure whether their brand is consistently associated with the category and markets that matter to them. A structured test can reveal patterns that a one-off conversation misses. Useful steps include: 目标不是试图从几次聊天中逆向工程出一个隐藏的规则。相反，企业应该衡量其品牌是否与其关注的品类和市场保持一致的关联。结构化的测试可以揭示一次性对话所忽略的模式。有用的步骤包括：

• Test a set of realistic buying, comparison and alternative-to prompts, rather than only searching for your own brand name. 测试一系列真实的购买、比较和“替代方案”提示词，而不是仅仅搜索你自己的品牌名称。
• Repeat prompts across multiple runs and record both cited answers and, where observable, fan-out queries. 在多次运行中重复提示词，并记录引用的答案以及（如果可观察到的话）扇出查询。
• Run relevant tests in the languages and locations where you sell, because those variables can alter recommendations. 在你销售的语言和地区进行相关测试，因为这些变量会改变推荐结果。
• Review whether your site and public materials clearly connect the brand to the specific category, use cases and terminology customers use. 检查你的网站和公开资料是否清晰地将品牌与特定的品类、用例以及客户使用的术语联系起来。
• Build content that supports several plausible discovery paths, including official product pages, clear documentation and category-focused explanations. Clear category signals are especially important for less established brands. 构建支持多种合理发现路径的内容，包括官方产品页面、清晰的文档和以品类为重点的解释。清晰的品类信号对于知名度较低的品牌尤为重要。

The research does not show a guaranteed method for becoming pre-selected, and businesses should not assume that adding a phrase to a page will alter ChatGPT’s internal planning. But consistent, authoritative material that explains what a company offers and where it fits gives AI systems clearer evidence when they do retrieve and synthesize information. AI search visibility is becoming a measurable business channel, not an assumption. 这项研究并没有展示出成为“预选品牌”的保证方法，企业也不应认为在页面上添加一个短语就能改变 ChatGPT 的内部规划。但是，解释公司提供什么以及其定位的一致性、权威性资料，能为 AI 系统在检索和综合信息时提供更清晰的证据。AI 搜索可见性正在成为一种可衡量的商业渠道，而不再仅仅是一种假设。