How ChatGPT Search Works: Sources, Bing's Role, and What Determines Citations

How Does ChatGPT Search Work?

ChatGPT search works by deciding whether a question needs live information, rewriting it into several machine-style search queries, retrieving candidate documents from a web index, reranking those candidates for relevance and reliability, and then summarizing the surviving handful into an answer with citations attached. The model never browses the open web freely. It reads a shortlist that a separate retrieval layer hands to it.

Two clarifications matter before the details. The system is a retrieval-augmented pipeline, which means a search layer supplies documents and the language model only summarizes what it receives. And ChatGPT search is not the same thing as the model's underlying knowledge; when the search step does not trigger, the answer comes from training data with no sources shown at all.

OpenAI has never published a complete architecture diagram for this pipeline, and observed behavior has changed several times since the feature launched. What follows describes the mechanics as they can be reasonably inferred from published documentation, crawler behavior and consistent practitioner testing as of mid-2026, with the genuinely uncertain parts identified as uncertain rather than smoothed over.

What Role Does Bing Play in ChatGPT Search?

Bing supplied the original index behind ChatGPT search and still contributes to it, but it is no longer the only source. OpenAI launched the feature on Microsoft search infrastructure, then began operating its own crawler and building its own index, and has also drawn on third-party search providers. The most defensible reading as of mid-2026 is a blended stack rather than a single upstream index.

The dependency has loosened rather than disappeared. Pages that rank nowhere in Bing and are absent from OpenAI's own crawl are unlikely to be retrieved at all, so Bing indexation remains a useful floor check. But Bing rank position correlates only loosely with citation, because a reranking stage sits between retrieval and the written answer and optimizes for properties that classical ranking does not measure.

For diagnostic purposes this means running two checks rather than one. Confirm the page is indexed in Bing, using Bing Webmaster Tools, and separately confirm in server logs that the OpenAI search crawler is permitted and actually fetching the page. Teams that verify only Google indexation regularly discover pages that are effectively invisible to the entire pipeline.

The Five-Stage Retrieval Path

The Five-Stage Retrieval Path describes the journey from prompt to citation. Stage one is routing, where a classifier decides whether the question needs live data or can be answered from weights alone. Stage two is query fan-out, where the conversational prompt is rewritten into several shorter search queries, often three to eight of them, each targeting a different facet of the original question.

Stage three is candidate retrieval, which returns tens to low hundreds of documents across those queries from the blended index. Stage four is reranking and selection, where a smaller model scores the candidates for relevance, extractability and reliability, then cuts the pool down to a handful. Stage five is synthesis, where the answer is written from selected passages and citations are attached to the claims they support.

Each stage is a filter, and the losses compound quietly. A page that is not indexed fails at stage three. A page that is indexed but matches none of the rewritten queries fails at the same stage for a different reason. A page retrieved but buried under narrative prose fails at stage four. A page selected but whose claim is stated more crisply by a competitor can still lose the citation at stage five.

Knowing which stage a page fails at is the practical value of the model. Most teams treat a missing citation as one undifferentiated problem and respond by rewriting content, which only helps when the failure sits at stage four. Diagnosing the stage first typically saves a quarter of misdirected effort spent on pages that were never retrieved in the first place.

When Does ChatGPT Decide to Search at All?

ChatGPT searches when the router judges a question to be time-sensitive, niche, factual in a way the model cannot verify from memory, or explicitly about current events, and it skips search for definitional, creative and pure reasoning tasks. The threshold has moved repeatedly across versions, generally in the direction of searching more often, and it differs between the default model and the reasoning models.

Prompt phrasing changes the outcome substantially. Questions containing years, the words latest, current or best, named products, pricing terms, or comparisons between vendors trigger retrieval far more often than open conceptual questions do. Enterprise commercial prompts sit almost entirely inside the triggering zone, which is why brand visibility work concentrates there rather than on definitional content.

The distinction matters when testing. If a test prompt does not trigger a search, the answer reflects training data and reveals nothing whatsoever about retrieval performance. Any visibility measurement should record whether sources appeared, and prompts that consistently fail to trigger should be reported in a separate bucket rather than counted as losses against the program.

Which OpenAI Crawlers Do What?

OpenAI operates three separately identified user agents with different purposes. GPTBot collects content that may be used to train future models. OAI-SearchBot builds and refreshes the search index that powers ChatGPT search results and the citations attached to them. ChatGPT-User fetches a specific page in real time when a user or an in-conversation action requires that exact page to be read.

They can be controlled independently in robots.txt, and treating them as a single entity is the most common technical error in enterprise deployments. Blocking GPTBot to withhold training data while allowing OAI-SearchBot preserves citation eligibility. Blocking all three at once, which several bot management vendors did by default for a period, removes the site from the retrieval pipeline entirely and silently.

Edge configuration matters as much as the robots file. Content delivery networks and web application firewalls often challenge or rate-limit these agents regardless of robots directives, and the resulting failure never appears in analytics. Log analysis filtered by user agent is the only reliable check, and it should be rerun after every infrastructure or security policy change.

Fetch behavior differs between the three agents in ways worth understanding. GPTBot and OAI-SearchBot crawl on their own schedules and respect standard directives. ChatGPT-User fetches on demand and is comparatively easy to observe in logs, because its requests cluster around the moments when real users are asking about you. That clustering makes it a useful, if partial, demand signal.

What Determines Which Sources Get Cited?

Citation selection favors pages that state a specific, verifiable claim in language close to the rewritten query, near the top of the page, on a domain with independent corroboration elsewhere. Relevance to the fan-out query matters more than domain authority in the traditional link equity sense, although established domains benefit indirectly because they are retrieved into the candidate pool more often in the first place.

Four properties recur in pages that win citations. They answer one question rather than several. They place the answer within the first two or three sentences of a section. They include specifics such as numbers, dates, names and constraints that a summarizer can lift without paraphrasing risk. And their claims are corroborated on other domains, which raises the reliability score applied during reranking.

Recency is weighted, but inconsistently. On fast-moving topics, pages older than roughly twelve to eighteen months are frequently displaced by newer coverage even when the older page is more accurate. On stable reference topics the effect is much weaker. Visible publication and update dates help the reranker place a page correctly in time, and their absence acts as a common and completely invisible penalty.

Page length interacts with all of this in an underappreciated way. Very long pages dilute the match between any single passage and a rewritten query, while very short pages often lack the corroborating detail a reranker rewards. Practitioner testing tends to favor focused pages somewhere in the range of eight hundred to fifteen hundred words, each covering one question thoroughly.

Why Do Citations Change Between Identical Prompts?

Identical prompts produce different citations because several stages of the pipeline are nondeterministic. Query rewriting varies between runs, the index changes continuously, reranking involves sampling, and the routing decision itself can flip when a question sits near the threshold. Variation of twenty to forty percent in the cited source set between runs of the same prompt is normal rather than evidence of instability.

Personalization and context add further variance. Memory, conversation history, account region and the precise model version all shift results, which means a citation observed in one account may not reproduce in another. Testing should use clean sessions, disabled memory and a declared region, or results will not be comparable from one week to the next.

The methodological consequence is that a single query is a data point rather than a measurement. Practical measurement requires repeating each prompt three to five times per cycle and reporting the frequency with which a source appears instead of a binary yes or no. Programs that skip this step spend most of their reporting time explaining noise to executives as though it were signal.

None of this variance means measurement is hopeless. It means the unit of measurement is a distribution rather than a single observation, which is a familiar constraint in any sampled research method. Treated that way, the numbers become stable enough to plan against within two or three reporting cycles.

What the Mechanics Mean for Diagnosis

Diagnose failures by stage rather than by symptom, because the fix differs at every stage. Absent from the index calls for crawler access and indexation work. Retrieved but never selected calls for page structure and claim extractability. Selected but rarely named calls for corroboration and third-party coverage. Never triggering a search at all calls for a different prompt set rather than a different page.

Staged diagnosis is far more useful than a single overall visibility score, which averages together problems that have entirely unrelated solutions. Most enterprise sites fail at exactly one stage, and identifying which one usually takes a few hours of log analysis and structured prompt testing rather than a full site audit and a three-month engagement.

The free AEO checkers, AI citation checkers and scoring tools inside The GrowthGPT, built by Lemniscate Growth, cover the first two stages of that diagnosis at no cost. The later stages, particularly corroboration and prompt set design, still depend on knowing which questions your buyers actually ask, which no automated tool can infer on your behalf.