How Does AI Choose Which Brands to Recommend in Its Answers?

Where does AI get its brand knowledge: model memory or live search?

AI brand knowledge comes from two sources. Model memory holds what the system learned from text gathered before its training cutoff. Live search adds current pages at the moment you ask. Which source dominates depends on whether the assistant searched for that particular question.

A quick way to tell them apart: look for citations. An answer with source links, or a Sources button in ChatGPT, used live search. An answer with neither came from memory.

Memory answers favour brands that were written about often and consistently for years. A two-year-old company may simply be absent. No change to its website will alter that until a later model is trained.

Search answers are more open. OpenAI’s help article on searching the web with ChatGPT says ChatGPT may search automatically when a question needs current information.

You can test memory on its own. Ask an assistant about your brand with search switched off, or in a tool without web access. Whatever it says, right or wrong, is roughly what the model absorbed during training.

If it has never heard of you, that is normal for a younger brand. If it repeats an old address or a service you dropped, those outdated pages are still out there, and worth finding.

For brand questions such as “best accounting software for a small firm”, assistants often search. That is good news for newer brands, because a strong page published last month can be found.

Our large language model entry explains training cutoffs, and the AI citation entry covers what a source link proves.

How does a brand get onto the AI shortlist?

A brand gets onto the AI shortlist when the pages a search returns name it clearly for the exact need in the question. The assistant first turns the prompt into search queries, then reads what ranks, then lists the brands those sources agree on. Each step can include or drop you.

1. Query rewriting. OpenAI says ChatGPT rewrites a prompt into one or more targeted queries for its search partners. Its example turns “good restaurants near me” into “top restaurants San Francisco”.
2. Retrieval. The search partner returns pages. OpenAI’s article links Microsoft’s and Shopify’s privacy policies as partners.
3. Reading. The model reads those pages and pulls out brand names, features, prices and opinions.
4. Agreement. Brands named by several sources, in matching terms, are easier to recommend.
5. Writing. The answer lists a few brands, often with a reason each, and may cite the pages it used.

Google’s AI features follow a similar path. Its guide on optimizing for generative AI features describes query fan-out, where several related searches run at once and each one adds supporting pages.

The agreement step explains a common complaint. A brand with one excellent page but no outside coverage often loses to a rival that three review sites mention. The assistant has more independent sources saying the same thing about the rival, so naming it feels safer.

So a brand needs to appear in the results for the rewritten queries, which may be narrower than the buyer’s words. “Best CRM” becomes “best CRM for real estate agents India” quickly.

Our query fan-out entry shows how one question splits into many searches.

Why can the same question return different brands?

The same question can return different brands because AI assistants rewrite each prompt using context such as location and memory, and because the model writes a fresh answer each time. Two buyers asking identical words can see different shortlists, and one buyer can see a new list on a rerun.

OpenAI names two of these inputs directly. ChatGPT may use an approximate location from your IP address, and with memory on it may use saved details when rewriting a query.

Its own example is a vegan user in San Francisco. A request for nearby restaurants becomes a search for vegan restaurants in that city.

Three other causes are less visible:

• Model sampling. Language models do not produce identical text every time, so the list order and length shift.
• Search results move. Rankings change daily, and so do the pages an assistant reads.
• Different products. ChatGPT, Gemini, Perplexity, Claude and Microsoft Copilot use different models and search sources.

This is why Digital Magma never judges brand visibility on one screenshot. We run each prompt several times, from a clean session, and count how often a brand appears.

Brand mention, citation and recommendation: three different outcomes

A brand mention, a citation and a recommendation are three separate outcomes in an AI answer. A mention names your brand. A citation links to your page as a source. A recommendation tells the user to choose you. Track each one apart, because they move for different reasons.

A brand can be cited without being recommended. A pricing page might be the source for a fact while a rival gets the endorsement.

The reverse happens too. A brand is recommended from a review site, and its own website is never cited.

The wording of the recommendation matters as well. An assistant might say a brand is “good for small budgets” or “best for enterprise teams”. That label usually echoes how reviewers and lists describe the brand.

If the label is wrong for you, the fix is upstream. Change how your own pages describe your fit, and ask reviewers and publishers to describe it accurately.

For buyers, the recommendation matters most. For reporting, all three are useful, and our AI share of voice entry shows how to count them.

Which brand signals can you influence, and which can you not?

You can influence the brand signals that live on the open web today: your own pages, your business listings, your reviews and what credible publishers write about you. You cannot directly change model training data already collected, how a model samples its words, or a user’s location and memory.

Signals a brand can influence:

• Answer pages. One clear page per service or product, with the use case, place and price range in plain text.
• Business listings. Google’s guide says Business Profile and Merchant Center data can help products and services appear in AI responses.
• Reviews. Real customer reviews on Google and trade sites, answered by the business.
• Third-party coverage. Trade press, association lists, podcasts and comparison articles.
• Consistency. The same brand name, services and contact details everywhere, which our entity SEO entry covers.

Signals a brand cannot influence directly:

• Text already used to train current models
• The model’s own sampling and wording
• Each user’s location, memory and chat history
• Which search partner a given assistant uses

Some shortcuts do not work. Google’s guide says chasing inauthentic mentions is less helpful than it seems, because its spam systems filter them.

Paid placement is separate too. OpenAI’s Ads in ChatGPT article says ads run on separate systems and do not influence answers.

What Digital Magma has seen about how AI picks brands

Digital Magma’s client work points to one steady pattern: AI assistants name brands that already have a strong search presence and plenty of outside evidence. We have not seen a brand earn AI recommendations while staying invisible in normal search for the same need.

A Canadian welding manufacturer is the clearest example in our accounts. In a three-year engagement it has held 40+ keywords in Google’s top 10, and the brand is visible in ChatGPT and Google AI Overviews.

That visibility rests on years of product pages, technical fixes and earned mentions in its trade. In one three-month window the site drew about 30K clicks from about 100K impressions.

A Toronto ENT clinic, with us for four years, earned about 20K clicks in three months. For a local clinic, reviews and the Business Profile carry much of the evidence AI tools can find.

Our own brand is a small example of review evidence. Digital Magma is rated 5.0 from 100+ Google reviews. We cannot prove that rating causes any AI mention, and we do not claim it does.

What Digital Magma can say is that reviews are among the facts AI answers quote about local brands, so they are worth earning honestly.

What to do next: a brand-evidence plan for AI answers

A brand-evidence plan gives AI assistants more reliable pages to read about you. Digital Magma builds it in five steps: test, fix your own pages, fix listings, earn outside coverage, and retest. Each step adds evidence that a search can find and that agrees with the rest.

1. Test. Write 15 buyer prompts across ChatGPT, Gemini, Perplexity, Claude and Copilot. Run each several times from clean sessions. Record mentions, citations and recommendations.
2. Own pages. Make sure each core need has one page that names it, with proof and a direct answer near the top.
3. Listings. Update Google Business Profile. For product brands, keep Merchant Center feeds accurate.
4. Outside coverage. List the third-party pages cited for your prompts. Ask to be included where you fit, and pitch trade press with real data.
5. Retest. Repeat the same prompts each month. Compare counts, not single screenshots.

Be patient with memory-based answers. They change only when models are retrained, and vendors do not publish dates for that.

Expect uneven progress across tools. One assistant may start naming you within a month while another takes far longer, because each reads a different mix of sources.

Live-search answers move faster. A new page that ranks for the rewritten query can appear within weeks.

Digital Magma’s GEO agency work follows this plan, and the generative engine optimization entry gives the background.

How AI chooses brands: frequently asked questions

These are the brand questions Digital Magma hears most from founders who found a competitor named by AI instead of them. Each short answer draws on what OpenAI and Google have documented, and says plainly where nothing is documented.

Related brand visibility questions, sources and author

These pages continue the brand visibility topics covered here. Digital Magma keeps each knowledge-base term in a dated entry of its own, and answers not yet published are marked as planned.

• AI citation: what a source link in an AI answer means
• Query fan-out: how one question becomes many searches
• Generative engine optimization: work aimed at visibility in AI answers
• LLM optimization, entity SEO and Knowledge Graph (planned entries)
• How to get cited by ChatGPT and Perplexity, How to measure AI visibility and Why a website is not in ChatGPT

Sources: OpenAI’s help articles on searching the web with ChatGPT and on ads in ChatGPT, and Google Search Central’s guide to generative AI features. All were checked on 22 October 2026.