How ChatGPT decides which brands to recommend

ChatGPT recommends brands from two sources: what its underlying model learned in training, up to a fixed cut-off date, and the web pages it finds when it searches for that question. Which route an answer takes changes which brands appear, and only the second route can be changed quickly.

That explains most of the surprises marketing teams report: ChatGPT naming a competitor with confidence, missing a product you launched last year, or quoting fees you changed long ago. Below is what OpenAI's own documentation says, as we read it on 11 October 2026, and what it does not say.

In short

• ChatGPT names brands from memory (training data) and from search (live web pages it cites).
• Memory has a cut-off date. OpenAI lists a knowledge cut-off of 1 June 2024 for GPT-4.1, one of its API models.
• OpenAI says sites that block its OAI-SearchBot crawler will not be shown in ChatGPT search answers. Blocking GPTBot is a training opt-out, not a search one.
• OpenAI's crawler and developer documentation does not say how it ranks the sources it cites. Be wary of anyone who claims to know the formula.
• The only reliable way to know what ChatGPT says about you is to ask it the same buyer questions repeatedly and count.

Where does ChatGPT get the brands it recommends?

From what the model learned in training

A language model learns from a very large body of text collected up to a cut-off date. Brands that were widely and consistently described before that date are easier for it to recall. OpenAI's model documentation shows these dates: GPT-4.1, for example, has a knowledge cut-off of 1 June 2024. Events after a model's cut-off are not in its training data, so it can learn of them only by searching. That memory changes only with a new model version.

From what it finds when it searches

When ChatGPT searches the web, websites can be shown in its answers as linked sources. OpenAI's developer documentation for the web search tool in its API shows how this works there: answers carry inline citations to the pages used, and a separate list shows every URL the model consulted. The list of pages consulted can be longer than the list cited, so being read is not the same as being credited.

When does ChatGPT search instead of answering from memory?

OpenAI's documentation for its API's web search tool says the model can choose whether to search, based on the content of the prompt. The documentation does not list which questions trigger a search.

For a marketer, the practical consequence is that the same question can be answered from memory on one run and from search on another, or differently by different models. You cannot choose the route, so plan for both: be well described across the web for the memory route, and be readable and quotable for the search route.

Which OpenAI crawlers matter, and what do they do?

OpenAI runs several crawlers with different jobs. Each obeys its own robots.txt rule, so you can allow one and block another.

Crawler	What OpenAI says it is for	If you block it in robots.txt
OAI-SearchBot	Search in ChatGPT	Your site will not be shown in ChatGPT search answers, though it can still appear as a navigational link
GPTBot	Training OpenAI's generative AI foundation models	Your content should not be used for training; search is a separate setting
ChatGPT-User	Actions a user asks for in ChatGPT and custom GPTs	robots.txt rules may not apply, because a person started the request; OpenAI says it is not used to decide what appears in search
OAI-AdsBot	Checking pages submitted as ads on ChatGPT	It visits only pages submitted as ads

Two details from the same documentation matter in practice. If you allow both OAI-SearchBot and GPTBot, OpenAI may use one crawl for both purposes. After you change robots.txt, OpenAI says its search systems can take about 24 hours to adjust. OpenAI also recommends allowing requests from its published IP ranges, so check firewall and CDN bot rules as well as robots.txt. Our guide to robots.txt for AI crawlers covers the settings line by line.

How does ChatGPT choose which sources and brands to cite?

OpenAI's published crawler and developer documentation does not set out ranking factors. What it does show:

1. Access comes first. A page OAI-SearchBot cannot crawl will not be shown in ChatGPT search answers.
2. Location can change the sources. OpenAI's web search tool accepts an approximate user location, such as a country and city, to refine results. A question asked from Johannesburg may draw on different pages from the same question asked from London.
3. Products have their own route. OpenAI's commerce documentation lets merchants submit product data for discovery in ChatGPT, says eligibility does not guarantee display, and treats ads as a separate integration from discovery.

The best public evidence on what makes a page more likely to be used comes from research rather than from OpenAI. In the paper that coined the term generative engine optimisation (Aggarwal et al., KDD 2024), the authors rewrote web pages and measured how much of each AI answer drew on them. An unchanged page scored 19.5. Adding quotations from relevant sources raised it to 27.8 and adding statistics to 25.9, while keyword stuffing lowered it to 17.8. The tests used the researchers' own setup, not ChatGPT specifically, and results varied by subject, so treat them as direction rather than guarantee.

What can a South African brand do about it?

1. Let OAI-SearchBot in. Check robots.txt and any firewall or CDN bot rules. Decide on GPTBot separately: it is a training decision, not a search one.
2. Make your pages readable without JavaScript. If your content appears only after scripts run, a crawler that reads the raw HTML sees almost nothing. Our own site, built as a single-page app, showed crawlers 66 words on its homepage; after prerendering, GPTBot sees about 2,090. Check yours with our guide on whether ChatGPT can read your website.
3. Put the facts on the page, in text. State what you sell, who it is for, where you operate and what it costs in rand, with dates. Answers quote specifics they can find.
4. Be described the same way everywhere. The memory route learns from the wider web, so your name, category and locations should match across your site, social profiles, directories and press coverage.
5. Appear on the pages that get cited. Find which comparison sites, publishers and directories ChatGPT's answers cite for your category, and earn a place on them.
6. Measure, then change one thing at a time. Put the same unbranded buyer questions to ChatGPT on a schedule and count presence, first position and AI share of voice. One screenshot proves nothing; a trend on a fixed question set does.