How ChatGPT decides what brands to recommend

When you ask ChatGPT "What are the best running shoes for flat feet?", you see a single, coherent answer appear in seconds. What you don't see is what happens between your question and that response.

ChatGPT doesn't simply "know" the answer. Behind every response is a research pipeline: your single question gets expanded into multiple high-intent searches, information gets retrieved from dozens of sources, results get ranked and filtered, insights get synthesized, and citations get attributed.

For anyone trying to appear in AI-generated answers, understanding this pipeline matters. If you're optimizing for what users ask but ignoring what the AI actually searches for, you're invisible to the system.

This post reveals what happens between prompt and response.

Stage 1: Intent Analysis

Before ChatGPT generates any searches, it needs to understand what you're really asking for. This interpretation stage happens nearly instantaneously, but it's doing substantial work.

The model analyzes your natural language question to identify several key elements. First, it determines your primary intent. Are you looking for information to learn something new (informational)? Are you comparing options before making a purchase (commercial)? Are you ready to buy and need to know where or how (transactional)? Or are you asking the AI to create something for you (generative)?

Second, it extracts the key entities embedded in your question: brands, product categories, features, constraints. When you mention "flat feet," the system recognizes this as a contextual constraint that will shape everything downstream.

Third, it identifies implicit context that wasn't explicitly stated but is clearly relevant. "Flat feet" signals overpronation, arch support needs, and stability requirements. The system understands these implications without you spelling them out.

Finally, it determines what response format will best serve your needs. A comparison table? A ranked list with explanations? A step-by-step guide? The structure of the answer gets mapped before any research begins.

Take our example question: "What are the best running shoes for flat feet?" The system interprets this as a commercial intent question requiring a comparison or ranking, with flat feet as the critical context, and determines that ranked recommendations with justification will best satisfy the request.

This interpretation shapes everything that follows. Get the intent wrong, and the entire research process optimizes for the wrong outcome.

Stage 2: Query Fanout

That single question you asked? ChatGPT transforms it into multiple high-intent search queries designed to gather information.

On average, a prompt generates three to seven search queries. Simple informational questions might produce two or three queries. Complex commercial questions, especially those involving comparisons, specific use cases, or multiple constraints, can generate eight to twelve separate searches.

Let's see this in action with our running shoes example. The user asked one question, but ChatGPT internally generates something like:

"best running shoes flat feet 2025"
"stability running shoes overpronation"
"motion control shoes reviews"
"running shoes arch support comparison"
"best shoes for flat feet runners"

Notice what the AI is doing here. It's adding temporal qualifiers like "2025" to ensure current information. It's inserting quality signals ("best," "top," "review") to surface evaluative content rather than product descriptions. It's converting your general question into more specific category searches. And it's exploring synonyms and related concepts: "flat feet" becomes "overpronation" and "motion control."

This transformation reveals how the AI interprets category boundaries. It's not just searching for what you literally typed. It's searching for what it thinks you need to get a complete, accurate answer.

If your content ranks well for "running shoes" but doesn't appear for "stability running shoes overpronation" or "motion control shoes reviews," you're invisible to this research process. Even though you're in the category being researched, the AI will never see your content because it doesn't show up in the actual queries being executed.

The fanout isn't random. Analysis of millions of ChatGPT searches reveals consistent patterns in what gets added. The most common additions include "best," "top," "reviews," and the current year. The system transforms vague questions into high-intent, commercial searches designed to surface evaluative content.

Stage 3: Parallel Retrieval

With queries generated, ChatGPT executes them all simultaneously. This parallel retrieval happens fast, typically one to three seconds for the entire batch.

Each of those three to seven queries returns results. The system examines the top five to ten search results per query, which means a single user question triggers the retrieval of twenty to fifty web pages. That's twenty to fifty opportunities for your content to enter the consideration set, or twenty to fifty points where you could be absent.

What gets pulled back isn't just URLs. The system retrieves full metadata: page titles, URLs, snippets, publication dates, and signals about domain authority. It often pulls the full text of the page or excerpts, not just the snippet that would appear in traditional search results. When available, it also grabs structured data like ratings, pricing information, and specifications that can be incorporated into comparisons.

The retrieval uses integrated search APIs. For ChatGPT, that's primarily Bing's search index, though the specific implementation may vary. The critical point: the AI sees what search engines surface for these queries. If your content doesn't rank in traditional search for the fanout queries the AI generates, it won't even make it into the retrieval pool.

This is why traditional SEO and Answer Engine Optimization aren't competing strategies. They're complementary. The AI's research process is built on top of search engine results. You can't shortcut past search rankings to optimize directly for AI. The AI finds you through search, or it doesn't find you at all.

Stage 4: Filtering and Ranking

Retrieving twenty to fifty sources is just the beginning. Most of that content won't make it into the final answer. The filtering and ranking stage is where the AI decides what's actually relevant and trustworthy.

Multiple criteria come into play. Relevance to the original prompt matters most—not just relevance to the individual query that retrieved the source, but relevance to what the user actually asked. A page might rank well for "running shoes 2025" but if it doesn't address flat feet or overpronation, it gets deprioritized or filtered out entirely.

Source credibility acts as a major filter. The system favors established publishers, expert content, and authoritative domains. A blog post and an analysis from Runner's World don't enter the synthesis stage on equal footing.

Information quality matters: depth, specificity, substance. Thin pages that provide surface-level information get filtered. Pages with concrete details, specific comparisons, and analysis move forward.

Recency becomes critical for certain query types. When the AI added "2025" to a search query, it's signaling that current information matters. A review from 2022 loses to a comparison from 2025, even if the older content is more comprehensive.

The system also checks for consistency across sources. Claims that appear in multiple authoritative sources get weighted more heavily. Contradictory information triggers different handling—sometimes the AI filters out the outlier, sometimes it explicitly notes the disagreement in the answer.

Watch what happens to the pool of sources at this stage:

Retrieved: 20-50 sources
Evaluated: 15-30 sources
Actually used in synthesis: 8-15 sources
Explicitly cited: 4-8 sources

Getting retrieved is necessary but not sufficient. The content must be substantive, current, authoritative, and consistent with other credible sources to survive this stage.

This is where brand reputation compounds. Sources that consistently provide accurate information build implicit trust that improves their ranking in future pipelines. Conversely, thin content, contradictory claims, or outdated information gets filtered before synthesis begins.

Stage 5: Synthesis

The synthesis stage is where ChatGPT transforms information from multiple sources into a coherent narrative. This isn't simple copy-paste. The model is combining insights, resolving contradictions, and structuring information to match what you actually asked for.

First, it identifies common themes across the sources that made it through filtering. If six different sources mention that arch support is critical for flat feet runners, that theme becomes central to the answer. If only one source mentions a particular feature, it gets less emphasis or might not appear at all.

The system extracts specific claims and recommendations from each source. Not just "this shoe is good," but "this shoe offers 12mm heel-to-toe drop, dual-density midsole foam, and medial post for stability." Specificity from sources translates to specificity in answers.

When sources disagree, the AI has to resolve the contradiction. The approach favors consensus. If five sources say one thing and one source says another, the majority view usually wins. Sometimes contradictions get noted: "While most reviewers praise the mobile app, some users report syncing issues."

The structure of the answer gets tailored to match user intent. Someone asking "what's the best" gets a ranked list with justifications. Someone asking "how do I choose" gets decision criteria. Someone asking for a comparison gets a structured breakdown of differences.

Throughout synthesis, the model maintains factual accuracy by staying close to source material. It's not inventing claims or extrapolating beyond what sources actually said. The synthesis is a recombination and reorganization of information from the filtered sources, not creative generation.

What determines which brands or products get placement in the synthesized answer? Several factors compound:

Frequency of mentions matters most. If Brooks appears in seven sources and ASICS appears in four, Brooks typically gets more coverage, assuming other quality signals are comparable.

Position of information in source content creates an advantage. Claims in headlines, subheadings, and opening paragraphs get weighted more heavily than details buried mid-article.

Specificity trumps generic claims. "Medial TPU post reduces overpronation by 15 degrees" beats "great stability" every time.

Direct relevance to the specific question matters. Generic category information loses to content addressing the user's constraints or use case.

Source authority creates a multiplier effect. A mention in an authoritative source counts for more than a mention in a marginal one.

The result is a coherent answer that sounds like it came from a single expert source but actually represents a synthesis of information from eight to fifteen different pages, weighted by all these factors.

Stage 6: Citation

The final stage determines which sources get explicit credit. This matters because citations are visible proof of influence and the primary driver of click-through when users want to verify information or make a purchase.

A typical answer includes four to eight explicit citations. These appear as inline citations linked to specific claims, footnote-style references at the end of sections, or "according to" attribution statements woven into the narrative.

What earns a citation? Direct quotes always get cited. Specific claims—statistics, technical specifications, pricing details—typically earn attribution. Unique information that appeared in only one source almost always gets cited when it's included. Expert opinions or authoritative statements get explicit attribution. And direct product recommendations usually cite the source making that recommendation.

What doesn't earn a citation? Generic information synthesized across many sources gets included anonymously. Common knowledge within a category doesn't need attribution. Information that was derived from but not directly stated in sources often appears without citation. Background context built from multiple sources typically goes uncited.

Citations show clear bias toward certain source types. Authoritative domains (established publishers, expert sites, official brand websites) get cited more often than blogs or forums. Sources with clear attribution (author credentials, recent publication dates, their own references) earn more citations. Recent sources get preference, especially for product recommendations and reviews. And content that directly addresses the user's question gets cited over tangentially related content.

For brands, citations represent the visible evidence of influence. Even if the AI learned from your content during synthesis, users only see what gets cited. Being mentioned anonymously in the synthesis might build the AI's understanding of your category, but citations drive awareness and traffic.

This is why shallow, promotional content rarely gets cited even when it ranks well enough to be retrieved. The AI might read it, might even extract a fact or two, but authoritative citations go to substantive sources that provide concrete, verifiable information.

What the User Sees vs. What Actually Happened

From the user's perspective, the experience is simple:

One question asked
One answer received
Three to eight seconds elapsed
Four to eight sources cited

Behind that simple experience, a complex research pipeline executed:

Three to seven search queries generated from the single prompt
Twenty to fifty sources retrieved across those queries
Fifteen to thirty sources evaluated for relevance and quality
Eight to fifteen sources synthesized into the answer
Four to eight sources cited explicitly
Multiple ranking and filtering stages throughout

The entire process, from intent analysis through query fanout, parallel retrieval, filtering, synthesis, and citation, happens faster than you can read this sentence. The user experiences it as instant expertise. Brands experience it as the new battleground for visibility.

Beyond Web Search: Product Database Access

The pipeline described above covers how ChatGPT searches the web. For commerce queries, there's an additional component: direct access to product databases.

When you ask about running shoes or winter jackets, ChatGPT can display product cards with images, prices, ratings, and specifications pulled from its internal product catalog. These product cards appear alongside web citations and count as additional sources in the synthesis.

This means the total citations in a shopping answer often exceed the web results retrieved. A response might cite 60+ sources while only retrieving 30 web pages because product cards contribute the remaining citations.

This product search component operates in parallel with web search and will be covered in detail in a separate post.

The Visibility Gap

The invisible research happening between question and answer determines which brands appear, how they're described, and whether they get cited.

When someone asks about "running shoes for flat feet," the AI searches for "stability running shoes overpronation," "motion control shoes reviews," and "running shoes arch support comparison." Content that ranks for the category term but not for the specific, high-intent queries the AI executes never makes it past the retrieval stage.

Getting retrieved puts you in the pool of fifty potential sources. But you need to be in the final eight to fifteen that get synthesized, and ideally in the four to eight that get cited. The filtering stage eliminates thin content, contradictory claims, and outdated information before synthesis begins.

Citations go to sources that provide concrete information: specifications, test results, comparative data, analysis. Generic marketing content gets synthesized anonymously at best, ignored at worst.

The gap between what users ask and what the AI actually searches is where visibility gets won or lost.