ChatGPT Search visibility now depends on web.run and fan-out queries

ChatGPT Search visibility is getting tighter, and the reason is not just citations. After the default experience switched to GPT-5.3 Instant on March 4, 2026, a reverse-engineering study found the average answer cited 15 unique domains instead of 19, while unique URLs fell from 24 to 19. The bigger shift sits underneath that drop: ChatGPT often rewrites one prompt into several fan-out queries, sends them through web.run, and only then decides which pages to fetch and cite. If your site is not present across those hidden retrieval paths, being relevant is no longer enough.

For SEO and GEO teams, this changes the job. You are not only optimizing for a visible keyword anymore. You are optimizing for the sub-questions the model generates, the crawler that retrieves your pages, and the authority signals that make your brand a candidate before search even starts.

What changed in ChatGPT Search on March 4, 2026?

The immediate change was citation concentration. The study tracked 400 daily prompts over 14 weeks and found that after the move from GPT-4o/5.2 to GPT-5.3 Instant, the average response cited fewer domains and fewer URLs. In simple terms, the same answer surface was shared by fewer publishers.

That matters because a drop from 19 to 15 unique domains is not a rounding error. It means more competition for each citation slot. The study also notes that the URL-to-domain ratio stayed roughly stable at 1.26, which suggests ChatGPT did not suddenly start reading much deeper within each site. It simply pulled from a narrower set of sites overall.

This does not mean classic SEO stopped mattering. It means AI search is concentrating reward more aggressively. If your site makes the shortlist, you may gain disproportionate visibility. If it misses, you can disappear fast after a model update.

Why do fan-out queries matter more than one visible prompt?

Fan-out queries are the hidden searches generated from a single user question. Instead of sending one literal prompt to the web, ChatGPT can split that request into several narrower searches, then merge the results into one answer.

The study found that GPT-5.4 can chain 5 to more than 10 search rounds per response, while GPT-5.3 Instant usually runs 2 or 3. That difference alone helps explain why the same prompt can return different citations across model versions. One model explores more branches. Another stops earlier.

A product query makes the shift easier to see. For a prompt like best 3D printer to buy in 2026, the system first appears to build a candidate list, then launch separate shopping retrieval paths for each product to collect specs, reviews, and pricing. That is a different retrieval logic from old-school search, where one query leads to one result page. In AI search, one query can become many retrieval chances, and your content has to win in the branches, not just the headline term.

What does web.run reveal about the ranking system?

web.run is the internal browsing layer that powers live retrieval. According to the study, it changed meaningfully after GPT-5.3. Earlier versions used compact text commands. Newer versions use structured JSON with typed parameters, which points to a more explicit and more flexible search workflow.

The tool reportedly expanded from 4 operations to 12, including actions like search_query, open, find, click, screenshot, and specialized widgets for areas like sports, finance, and weather. That sounds technical, but the business implication is simple: ChatGPT is not just pulling a few blue links. It can search, inspect, refine, and revisit sources in multiple rounds before it writes the answer.

The study also surfaced signs that third-party search providers sit upstream in the process, which lines up with OpenAI's own documentation that ChatGPT Search may rewrite prompts into one or more targeted queries and send them to partner providers. For brands, that means AI visibility depends on more than copy alone. Retrieval architecture now shapes who gets seen.

Why do crawlability and brand memory both matter?

ChatGPT visibility now appears to have two layers. First, the model needs a reason to consider your brand or page as a candidate. Second, it needs to fetch and understand the content in real time.

The study describes the first layer as parametric visibility. Parametric visibility is what the model already "knows" from its training data and post-training signals, such as press coverage, Wikipedia presence, and mentions on high-authority sites. The second layer is dynamic visibility, which is what ChatGPT can retrieve live when search is enabled.

This framework is useful because it explains a common frustration in GEO. A brand can publish a strong page and still struggle if it has weak upstream authority signals. The opposite is also true: a known brand can be considered more often, yet lose live citations if the relevant page is blocked, poorly structured, or missing the sub-answer ChatGPT needs.

There is also a crawler nuance. OpenAI says sites need to allow OAI-SearchBot and the published IP ranges to be included in ChatGPT Search. But the study's honeypot experiment found that when ChatGPT selected a page during a live conversation, ChatGPT-User fetched the actual page content. In practical terms, discovery and live retrieval are not necessarily the same event, and both can fail for different reasons.

BotRank's Take

The biggest mistake brands can make here is treating AI visibility like a single ranking report. That model no longer fits the way ChatGPT Search works. If one prompt can branch into several hidden queries, and different model versions can produce different citation sets, then you need measurement that works at the prompt, model, and source level.

This is exactly where BotRank AI Visibility becomes useful. It lets teams run reusable prompts across multiple LLMs, track visibility over time, compare model-specific outcomes, and inspect which sources and pages are actually being cited. That matters because a citation is not automatically a win. Sometimes the cited page barely mentions the brand, or the answer frames the brand poorly. Measuring citations, entities, and sentiment together gives teams a much clearer view of real AI search presence than a yes-or-no ranking check.

The broader lesson is simple: GEO is no longer just about publishing content. It is about understanding retrieval behavior, then closing the gaps that retrieval exposes.

What should SEO and GEO teams do now?

The immediate play is not to chase every hidden query individually. It is to make your most important pages easier for ChatGPT to select, fetch, and reuse across adjacent questions.

• Audit crawl access. Make sure OAI-SearchBot is allowed, your CDN is not blocking the published IP ranges, and your core pages load cleanly when crawled.
• Design pages for answer coverage. One page should often resolve several related sub-questions, not just one keyword. Strong comparison pages, category explainers, and clear product or service pages tend to travel better through fan-out retrieval.
• Improve extractability. Use precise headings, direct definitions, structured lists, transparent specs, and clear evidence. If the model can identify the answer block fast, your odds improve.
• Build authority outside your site. Coverage on trusted publications, strong entity consistency, and reputable citations help strengthen the parametric layer that influences whether ChatGPT considers you at all.
• Test by model, not just by prompt. The study showed GPT-5.2, 5.3, and 5.4 can cite different sources even within the same model family. Visibility that looks solid in one experience can be weak in another.

One important nuance: this approach works best for topics where users ask layered, research-like questions. It is less decisive for purely navigational intent, where the user already knows the destination. But for commercial research, product comparisons, local discovery, and complex B2B questions, fan-out behavior is now central to the visibility game.

FAQ

What is a fan-out query in ChatGPT Search?

A fan-out query is when ChatGPT turns one user prompt into several targeted searches, then combines the returned information into one response. This gives the model more ways to find specific evidence, but it also means brands compete across multiple hidden sub-queries instead of one visible keyword.

Why did citation diversity drop after GPT-5.3 Instant?

The study points to a more concentrated retrieval pattern after the March 4, 2026 default-model switch. Fewer domains and URLs were cited per answer, which suggests a narrower winner set even if the overall answer format stayed similar.

What is the difference between OAI-SearchBot and ChatGPT-User?

OpenAI says OAI-SearchBot should be allowed if you want inclusion in ChatGPT Search. The study found that ChatGPT-User fetched the selected page content during live browsing, which suggests indexing and live retrieval can involve different agents.

Does this mean brands should stop doing traditional SEO?

No. Traditional SEO still supplies many of the signals AI systems rely on, especially authority, crawlability, and clear page structure. What changes is the unit of optimization: not just rankings for one keyword, but visibility across the sub-questions AI systems generate on the way to an answer.

ChatGPT Search is becoming easier to study and harder to game. That is good news for teams willing to measure what the model actually retrieves, not what they hope it sees. If you want to know which prompts, models, and cited pages are shaping your brand's AI visibility, BotRank gives you a practical place to start.