ChatGPT recommends brands in two ways: from what the model absorbed during training (its parametric memory) and, when it searches the web, from the pages it retrieves and cites at answer time. There is no public ranking formula and no paid placement in ordinary answers, so the brands that appear are those that are consistently, clearly and credibly described across the sources ChatGPT can reach. Because answers vary from run to run, the only honest way to know where you stand is to measure your presence across many repeated questions.
Key takeaways
- ChatGPT draws on two sources of knowledge: what it learned in training and what ChatGPT search retrieves live. They respond to different levers and on different timescales.
- OpenAI documents a few concrete signals for search and shopping, such as crawl access for OAI-SearchBot and structured product metadata. Most other "ranking factors" you will read about are inferences, not documented facts.
- The best controlled evidence (Aggarwal et al., KDD 2024) found that adding citations, quotations and statistics to content improved visibility in generative engine responses by up to 40% in their benchmark.
- Answers are probabilistic. A single screenshot proves nothing; a presence rate over a repeated, fixed question set does.
- Third-party coverage (reviews, comparisons, press, community discussions) often matters as much as your own site, because that is where models meet your brand in context.
Two paths to a recommendation: training memory and live retrieval
For brand recommendations, ChatGPT has two quite different mechanisms, and each responds to different levers.
Parametric memory: what the model already "knows"
A language model compresses patterns from its training data into its weights. If your brand appeared often, in consistent contexts ("X is a project management tool for agencies"), the model is more likely to associate it with the matching category and needs. This memory has a cut-off date, it updates only when a new model is trained, and it cannot be edited directly.
Live retrieval: ChatGPT search
Since OpenAI launched ChatGPT search on 31 October 2024, ChatGPT can decide by itself to search the web, or the user can trigger a search manually. According to OpenAI, search uses third-party search providers plus content that partners supply directly, and answers show the sources they used. Here the logic is closer to search engines: a page has to be crawlable, indexed by the providers, relevant to the query that ChatGPT actually sends, and useful enough to be selected and cited.
In practice the two paths blend. The model may form an initial answer from memory and then refine or confirm it with retrieved sources. For shopping queries, OpenAI's help documentation explicitly lists "earlier model responses, generated before it considers new search results" as one of the inputs.
Which recommendation signals are actually documented?
The table separates what OpenAI documents from what is plausible but unproven.
| Signal | Mechanism | Status | What you can do |
|---|---|---|---|
| OAI-SearchBot access | OpenAI says sites opted out of OAI-SearchBot are not shown in ChatGPT search answers (they can still appear as navigational links) | Documented | Check robots.txt and firewall/CDN rules |
| Structured product metadata | Shopping results use metadata such as price and description from first- and third-party providers | Documented (shopping) | Keep feeds, schema markup and marketplace listings accurate |
| User context | Query wording, Memory and custom instructions shape which products are shown | Documented (shopping) | Cover the different needs and constraints your buyers express |
| Merchant factors | Availability, price, quality and whether the merchant is the maker or primary seller | Documented (shopping) | Make your official store and stock status clear |
| Third-party mentions | Reviews, comparisons and press feed both training data and retrieval | Plausible, widely observed | Earn coverage on the sources that ChatGPT cites in your category |
| Evidence-rich content | Citations, quotations and statistics increased visibility in a controlled study | Experimental evidence | Add verifiable data and named sources to key pages |
Two points matter here. First, OpenAI states that product results "are selected independently by ChatGPT and are not ads"; ads, where they exist, are handled separately. Second, nothing in the documentation describes a numeric score you could optimise directly. Anyone selling a "ChatGPT ranking score" is offering a proxy of their own.
What does the research evidence say?
The most cited controlled study is "GEO: Generative Engine Optimization" by Aggarwal and colleagues (arXiv:2311.09735, published at KDD 2024). The authors built a benchmark of diverse queries (GEO-bench) and tested content changes on a simulated generative engine, and also on Perplexity. Methods that added citations to credible sources, relevant quotations and statistics produced the largest improvements, with visibility gains of up to 40% in their setting. Keyword stuffing, the classic SEO shortcut, performed poorly. Effects also varied by domain.
The study was not run on ChatGPT's production system, which changes frequently. Its value lies in the direction it points: generative engines appear to favour content that is easy to verify and easy to quote. If you want to apply this, our guide on how to write content that AI quotes turns those findings into an editorial checklist, and which sources AI cites shows where retrieval tends to look.
Why the same question gets different recommendations
Seeing your brand in one answer and not in the next is expected behaviour, not a bug:
- Sampling. Generation is probabilistic, so the same prompt can produce different lists.
- Search or no search. One run may trigger a web search and another may answer from memory, with different candidates.
- Context. Memory, custom instructions, conversation history, language and location all shift the answer.
- Phrasing. "Best CRM for small agencies" and "affordable CRM with WhatsApp integration" are different questions with different winners.
For this reason, recommendation visibility should be expressed as a rate, not a yes/no. The presence rate is the number of valid answers that mention your brand divided by all valid answers; failed or empty answers belong outside the denominator rather than counting as misses. The article on how to measure AI visibility explains sampling sizes and common mistakes in detail.
An evidence-based method to improve how ChatGPT recommends you
- Build a question set from real buyer intent. Collect 30 to 100 questions your customers actually ask: category questions ("best payroll software for startups in Germany"), comparison questions and problem questions. Write them in each market's language.
- Measure a baseline. Ask each question several times, in the relevant market and language, and store every answer with its sources. Compute your presence rate and the same metric for competitors on the same sample, so the comparison is fair.
- Analyse the citations. List the domains ChatGPT cites when it recommends your category. These are the pages shaping the answer; see AI citations for the distinction between a mention and a citation.
- Fix access first. Make sure OAI-SearchBot is not blocked and that key pages render their main content without heavy client-side scripts.
- Close the evidence gaps. Publish precise, verifiable pages: clear category statements, pricing, specifications, independent data and named sources. Pursue accurate coverage on the third-party sites you found in step 3.
- Re-measure on the same question set. Compare rates month to month. A change of a few points on a small sample may be noise; look for consistent trends.
A tool such as GeoDog automates steps 2 and 3 by asking real questions in a chosen market and storing every answer with its sources, but the method works with a spreadsheet too. What matters is a fixed sample and an honest denominator, as described in our methodology.
Common myths about ChatGPT recommendations
- "You can pay to be recommended." OpenAI states that shopping results are not ads and are not influenced by partnerships.
- "llms.txt makes ChatGPT recommend you." It is a proposed convention, and OpenAI has not documented using it as a ranking signal. See what llms.txt is for a balanced view.
- "One good answer means you are visible." One run is an anecdote, not a measurement.
If you are new to the field, start with what GEO is for the broader framework.
FAQ
Can I ask OpenAI to include my brand in recommendations?
No. There is no submission process for ordinary answers. You can make your site accessible to OAI-SearchBot, keep product data accurate, and earn credible third-party coverage; selection itself remains with the system.
How long does it take to see changes?
Changes that work through ChatGPT search can appear within weeks, once pages are crawled and indexed by the providers. Changes to the model's training memory depend on new model releases and cannot be scheduled.
Is being mentioned the same as being recommended?
Not quite. A mention may be neutral or negative, while a recommendation positions you as a fit for the user's need. Track both, and read the answers to understand the context.