How ChatGPT Decides Which Suppliers to Recommend: Observable Signals vs. Uncontrollable Factors

Whenever someone asks “what do we have to do so our company shows up in ChatGPT recommendations,” we start with one sentence: nobody can guarantee this. That is not modesty—it is the result of our own testing. In September 2026 we ran 72 buyer questions in a clean environment. Across the AI answers, 18 companies were named in total; our own site was cited 3 times, and every one of those times it was as a source for a “how to choose a vendor” reference list, not as the company being recommended. Being cited is not the same as being recommended.

This article breaks ChatGPT recommendations into three layers: which signals are visible, which you can act on yourself, and which are entirely out of your hands. After reading it, you will at least be able to tell whether a vendor is talking to you about observable engineering or selling you a promise it cannot keep.

First, separate three things: appearing, being cited, being recommended

Most discussions lump these three together, which is why the conclusions turn into mysticism.

• Appearing: whether the question triggers an AI-generated answer at all. In September 2026 we tested 500 Chinese-language B2B queries, and 80.9% returned a Google AI Overview. In other words, four out of five times your buyer’s question will surface an AI summary.
• Being cited: whether your domain appears in the source list under the answer. This layer can be measured and tracked.
• Being recommended: whether the body of the answer names you, and does so in the tone of “you can go to them,” rather than “here is a list you can consult when choosing.”

The third layer is the most valuable and the least controllable. All 3 of our citations fell into the second layer. That gap is the biggest hole in the “guaranteed AI recommendation” business.

The measurable part: one question panel, four fields

To talk about optimization, you first need repeatable measurement. Our approach is to write 50 to 100 real buyer questions of our own, run them logged out, in the same region, in the same interface language, all on the same day, and then record only four fields: whether an answer appeared, which domains were cited, which companies were named, and in what role they were named.

The key is “same day, same environment.” Models and retrieval sources get updated, so answers from different days, different accounts, or different regions cannot be compared directly. A single result is a data point, not a trend; if you want to talk about change, rerun the same panel once a month. We wrote about how to track AI citations down to the URL level in the article on tracking ChatGPT citation sources.

Signals you can see and also act on

The following things affect whether AI can understand you and whether it is willing to cite you, and all of them live inside your website and your external content:

• Entity consistency: your company name, what you do, where you are, and which markets you serve should be described the same way on your website, in directories, and in third-party articles. If the same brand has three different descriptions across three sources, it is hard for a model to assign you to a single entity.
• A crawlable and indexable foundation: canonical tags, hreflang, sitemap, internal links, and whether pages render properly. If this layer fails, citations further down the line are out of the question.
• Vendor-selection lists and comparison content: all 3 of our citations came from roundup articles of the “how to choose a vendor” kind. AI cites third-party lists more often than it cites your own self-introduction page.
• Direct-answer paragraphs and FAQ: one question paired with one direct answer is easier to excerpt than a long build-up. Structured data (schema) helps engines understand the page, but it does not by itself produce citations.

Pay attention to that last sentence. Many guides present schema as “the key to getting AI to cite you,” and our own measurements do not support that claim.

The part that is not in your hands

No matter how much you optimize, there are things you cannot control: model versions and their update cadence, which sources the answer engine decides to retrieve at that moment, the platform’s own commercial logic, the user’s location and personalization, and how strong your competitors’ content is on the same batch of questions. These variables make “guaranteed recommendations” technically impossible.

Google states it plainly in this document: nobody can guarantee a number one ranking on Google, and you should be especially wary of vendors claiming a special relationship with Google or guaranteed rankings. The same holds if you swap in ChatGPT.

If someone says they can get ChatGPT to recommend you, ask three questions first

1. What does your measurement panel look like? How many questions, what environment, which fields do you record? If they cannot answer, there is nothing further to discuss.
2. Can you tell the difference between being cited and being recommended? Anyone who treats the two as one thing is selling you the uncontrollable part.
3. If AI still does not name me three months from now, what does the contract say? A promise without written boundaries is no promise at all.

When we discuss this topic among ourselves, we always split it into “checks and fixes we can promise” and “outcomes we cannot promise.” The former is engineering; the latter is probability.

So should you still do SEO

Yes, and the order does not change. Take our own site as an example: over the past 16 months, tianwenwangluo.com’s Google Search data shows 543,305 impressions, 7,460 clicks, a CTR of 1.37%, and an average position of 15.8. What these numbers say is this: impressions can be generated, but with an average position past the second page, the number of clicks actually available is limited. AI Overviews will eat another share of those clicks, so “first get pages indexed, understood, and ranking” remains the baseline, and GEO adds a layer of measurement on top of that baseline.

As for which practices are implementation-level and which are just talk, we have collected that separately in the risks of guaranteeing AI recommendations and competitors being named by AI while you are not. The full statistics for those 500 Chinese queries, including the query list and the calculation method, are in our AI Overview coverage study, so you can rerun it yourself.

Back to the original question. The ChatGPT recommendation list is generated by the model at that moment, and nobody can dictate it. What you can do is: describe your entity clearly, shape your pages into something citable, take care of third-party sources, and then use a repeatable panel to check once a month whether you are moving from “being cited” toward “being recommended.” Moving is a result; not moving does not mean anyone owes you a ranking.

This page was machine-translated from the Traditional Chinese version and is pending human review.