ChatGPT Search: How It Works and How to Show Up

ChatGPT Search is the answer-plus-sources web search built into ChatGPT. Per OpenAI, it runs on a fine-tuned version of GPT-4o that calls third-party search providers and partner content feeds, then cites a curated set of pages in a sources panel under each answer (source). It is not a separate search engine sitting next to ChatGPT. It is ChatGPT deciding, mid-conversation, that your question needs the live web, going to fetch it, and showing its work.

And the single highest-impact move to show up in it is almost insultingly simple: let OAI-SearchBot crawl you. That is the crawler that builds the index ChatGPT Search reads from, and OpenAI states plainly that sites opted out of it will not be shown in ChatGPT search answers (source). Block that bot and you are invisible, full stop. No amount of clever content fixes a closed door.

The rest is doing three things in order. Get crawl access. Get into the underlying index. Write answer-first so the model actually picks you. Let me walk through how the machine works, then how to win inside it.

Key takeaways

• ChatGPT Search runs on a fine-tuned GPT-4o that calls third-party search providers plus partner feeds, then cites sources under the answer. That is OpenAI's own description (source), so the whole game is being one of those cited sources.
• OAI-SearchBot is the gate. Per OpenAI's bot docs, sites opted out of OAI-SearchBot are not shown in ChatGPT search, robots.txt changes take roughly 24 hours to reflect, and GPTBot is a separate training-only crawler.
• Getting retrieved is not getting cited. An AirOps study of 548,534 retrieved pages found only 15% were cited in the final answer. The other 85% got read and tossed.
• ChatGPT fans out your question into multiple sub-searches on 89.6% of searches (source), so you can get cited for adjacent sub-topics you never explicitly targeted.
• ChatGPT has roughly 900 million weekly active users (source), and the brand-level citation rate sits around 0.59% (source), so the audience is enormous and the slots are scarce.

The short answer: how ChatGPT Search works, and the one move that matters

Here is the whole thing in a paragraph. You ask ChatGPT something. A fine-tuned GPT-4o model decides whether your question needs the live web (a lot of questions don't, so it just answers from memory). If it does need the web, the model rewrites your question into one or more search queries, fires them at third-party search providers and partner feeds, gets back a pile of candidate pages, reads them, ranks them, writes an answer, and stamps a sources panel underneath with the handful of pages it actually used. That sources panel is the prize. That is what we are fighting over.

Now the move that matters. Out of all the levers you could pull, exactly one is binary: crawl access. Every other lever (schema, structure, freshness) moves a probability. Crawl access is a gate. If OAI-SearchBot can't fetch your pages, you are not eligible to be cited, period, no matter how perfect the page is. So before you touch anything else, confirm OAI-SearchBot is allowed in your robots.txt and not getting a 403 from your CDN. We wrote a whole field guide on AI crawlers and robots.txt if you want the copy-paste configs, but the one-line version is: never block OAI-SearchBot.

Everything past this point assumes you have opened that door. If you haven't, the clever stuff is wasted effort. Get the door open, then optimize.

How ChatGPT Search actually works

Let me pop the hood. ChatGPT Search is a few distinct steps glued together, and understanding each one tells you exactly where you can intervene.

Step one: a fine-tuned GPT-4o decides when to search

ChatGPT Search isn't a search box. It's a model that sometimes chooses to search. OpenAI describes the feature as a fine-tuned version of GPT-4o, post-trained with techniques including ones from their o1 work, that knows when a question benefits from live web data (source). Ask it to write a poem, no search. Ask it "what's the cheapest flight to Lisbon next Tuesday," it searches. Ask it "best project management tool for a 5-person agency," it usually searches, because that answer changes and the model knows its training data is stale.

This matters for you because not every query triggers retrieval. The ones that do tend to be factual, fresh, local, comparative, or commercial. Those are exactly the queries where brands want to show up. So when you think about where ChatGPT can cite you, think about the questions that force it to go look something up.

Step two: query rewriting and fan-out

Here is the part most people miss. ChatGPT rarely searches for your exact words. It rewrites your question into cleaner search queries, and it usually generates several of them. The AirOps study found ChatGPT produced two or more fan-out queries on 89.6% of searches (source). Fan-out means your one question ("how do I stop my sourdough from being dense") quietly becomes three or four sub-searches ("sourdough proofing time," "sourdough hydration ratio," "underproofed vs overproofed bread," and so on).

Why should you care? Because 32.9% of cited pages in that study showed up only for a fan-out query, never the original prompt (source). Translation: you don't have to be the perfect match for the headline question. If you cover the adjacent sub-topics well, you can sneak into the citations through a fan-out query you didn't even know existed. This is why topical depth beats a single shiny money page. If you want the mechanics, the fan-out query glossary entry breaks it down further.

Step three: third-party search providers and partner feeds

The model takes those rewritten queries and sends them to search providers. OpenAI's language is "third-party search providers" plus partner content from publishers it has agreements with (source). The providers return a ranked list of candidate web pages. The partner feeds layer in licensed content from news and reference publishers. So the candidate pool ChatGPT considers is a blend of open-web search results and licensed partner material.

This is the retrieval step, and it's pure RAG under the hood. The model isn't recalling your page from training. It's fetching it live, in the moment, to ground its answer in real sources. If you've never met the pattern, our retrieval-augmented generation glossary entry explains why "grounded in retrieved documents" is the entire reason citations exist at all.

Step four: ranking, answering, and the sources panel

Now the model has a pile of candidate pages. It reads them, judges which ones best answer the question, writes a synthesized answer in plain language, and attaches a sources panel listing the pages it actually used. Most candidates never make the panel. The model retrieved them, skimmed them, and decided they weren't worth citing.

That sources panel is the visible output of a brutal selection process. And selection, not ranking, is the whole game in ChatGPT Search. In classic SEO you fight for position. Here you fight to be one of the few pages the model chooses to quote out of dozens it retrieved. More on that selection step later, because it's where most of your optimization actually lands.

The Bing question: what OpenAI says versus what everyone infers

Let me be honest here, because this is where a lot of guides overstate their case. OpenAI's official language is "third-party search providers." That's it. OpenAI does not publish a sentence that says "ChatGPT Search uses Bing and only Bing." So anyone telling you OpenAI confirmed Bing is putting words in their mouth.

What we can say honestly: ChatGPT's web search is widely understood to lean heavily on Microsoft's Bing index. Microsoft is OpenAI's largest investor and infrastructure partner, OpenAI's earlier Bing-powered browsing used Bing directly, and OpenAI's own bot documentation discusses Bingbot in the context of ChatGPT's web search (source). The practical, observable behavior matches a Bing-heavy retrieval layer. But "third-party search providers" is deliberately plural and vendor-neutral, and OpenAI may blend or swap providers over time. So I'll tell you what I actually believe: Bing is the dominant retrieval source today, and treating it that way is the correct bet, while remembering it's an inference, not an OpenAI quote.

What does that mean for you, concretely? Get your pages into Bing's index. Sign up for Bing Webmaster Tools, submit your sitemap, and confirm coverage. If your pages aren't in Bing, ChatGPT's retrieval step may never surface them, and you're cut before the audition. But don't abandon Google either, because strong Google rankings still correlate with AI citations as a quality signal. Both indexes, different roles. We go deeper on the source-selection picture in how AI engines choose sources.

OAI-SearchBot vs GPTBot vs ChatGPT-User

OpenAI runs three crawlers, they do completely different jobs, and people mix them up constantly. Get this table tattooed on the inside of your eyelids, because confusing these three is the number one reason brands go missing from ChatGPT Search without knowing why.

A few honest notes the table can't hold. Per OpenAI's bot documentation, OAI-SearchBot is the one that builds the ChatGPT search index, and OpenAI says sites opted out of it will not be shown in ChatGPT search answers. GPTBot is training only, so you can block it and still get cited (a perfectly coherent "cite me, don't train on me" stance). ChatGPT-User is user-triggered and not used for search ranking, so it powers the live "fetch this URL" experience but doesn't build the persistent index.

Here's the trap that catches everyone. Back in 2023, tons of sites added a blanket block on GPTBot because they didn't want their content training a model for free. Reasonable. But a lot of those rules were too broad, or got OAI-SearchBot bolted on later out of habit, and now those sites are locked out of ChatGPT Search and have no idea. Same vendor, opposite outcomes. If your brand vanished and you can't figure out why, start with this table, then read why your brand is not showing up in ChatGPT for the rest of the checklist. The AI crawler glossary entry is a quick reference too.

One more gotcha: robots.txt is not the only gate. Per OpenAI's docs, changes take roughly 24 hours to propagate, and a WAF or CDN (Cloudflare's bot rules, say) can return a 403 to OAI-SearchBot before robots.txt ever applies. So check your edge layer and your server logs, not just the file in your editor. A clean robots.txt means nothing if your firewall is dropping the bot at the door.

How to show up in ChatGPT Search

Crawl access gets you eligible. Now you want to actually win the citation. Here's the stack, in priority order.

1. Crawl access (the gate)

Already covered, but it's first for a reason. Allow OAI-SearchBot. Confirm it gets a 200, not a 403, in your server logs. Wait the ~24 hours for propagation before you panic. If you do nothing else on this list, do this, because it's the difference between "impossible to cite" and "in the running." The llms.txt guide covers an optional curation layer that sits on top of access, but access comes first.

2. Index presence

Get into the index ChatGPT Search actually queries. In practice that means Bing. Submit your sitemap to Bing Webmaster Tools, check that your important pages are indexed (not just submitted), and turn on IndexNow if your CMS supports it so fresh content gets picked up fast. If your pages aren't in the index the retrieval step reads from, you're invisible to it even with a wide-open robots.txt. Crawlable but unindexed is still uncited.

3. Answer-first content

ChatGPT reads a page top to bottom and weighs the opening most heavily. So put a self-contained answer of 40 to 80 words at the very top of every page, directly answering the question in the title. No "in this article we'll explore." No 600-word brand origin story. Just the answer, stated plainly, like you'd tell a friend. That capsule is the chunk the model lifts and quotes. Burying your answer is the most common self-inflicted wound I see.

4. Fact density

Pages thick with concrete, attributable facts get cited far more than pages full of fluff. Give ChatGPT a number it can quote and attribute to you. "Most users prefer it" gets ignored. "73% of 1,200 surveyed users preferred it" gets cited. Add original statistics, cite your sources, and write the way Wikipedia writes: dense, neutral, factual. The model trusts content that reads like an answer, not an ad. Our AI content optimization guide goes deeper on the fact-density playbook.

5. Schema and structure

Wrap your content in JSON-LD so the model can parse it cleanly. FAQPage schema turns each question into a tidy, self-contained answer capsule. Article schema with a named author and visible dates is your E-E-A-T signal. One idea per heading, short paragraphs, comparison tables (the model loves quoting a clean table). Schema doesn't make weak content get cited, but it removes guesswork and makes your good content easier to extract. The schema markup for AI search piece has the templates.

6. Freshness

ChatGPT Search exists specifically to give current answers, so it favors recently updated content for time-sensitive questions. Show visible published and updated dates. Refresh your important pages on a real cadence, not once a year. A page dated last week beats an identical page dated 2022 for any query where recency matters, and a surprising number of queries quietly care about recency.

For the engine-specific version of this whole stack, how to get cited by ChatGPT walks through each step with configs and a timeline.

How citation selection works (only a small slice gets cited)

This is the part that reframes everything, so let me hammer it. ChatGPT retrieves way more pages than it cites. The AirOps study looked at 548,534 pages ChatGPT retrieved across thousands of prompts and found only 15% got cited in the final answer (source). The other 85% got pulled in, read, judged, and dropped. The model looked at them and decided they weren't worth quoting.

Sit with that. Getting retrieved feels like winning. It isn't. It just means you made the audition. Five out of six pages that make the audition get cut. So the real question isn't "can ChatGPT find me," it's "out of the dozens of pages it grabs for this question, does it pick mine to quote." That's citation selection, and it's a different sport from ranking.

What tips selection your way? Pages that most clearly and credibly answer the specific question. That means the answer is near the top (not buried), the facts are concrete and attributable (not vague), the authorship and dates are visible (trust signals), and the structure is clean enough to extract a quotable chunk (headings, tables, short paragraphs). It's the same answer-first, fact-dense, well-structured content from the last section, but now you understand why it matters: it's literally the difference between being one of the 15% and being one of the 85%.

And zoom out for scale. ChatGPT has roughly 900 million weekly active users (source), but the brand-level citation rate sits around 0.59% (source). Massive audience, tiny number of citation slots per answer. That's exactly why selection is everything and why small structural wins compound into real visibility. And because the slots are scarce, tracking your slice of them matters: the AI share of voice breakdown shows how to measure how much of the cited real estate you own versus your competitors.

How to measure your ChatGPT Search visibility

You can do every step above and still have no idea whether it worked, because ChatGPT Search is non-deterministic. Ask the same question twice and you can get different citations. So you can't eyeball it once and call it a win. You need a before-and-after loop with enough runs to be statistically honest.

Here's the loop, the same one I'd run by hand:

1. Pick your target prompts. The actual questions where you want to show up. "Best [your category] for small business," "how to [problem you solve]," "[your brand] vs [competitor]." Start with 20 to 50.
2. Measure your baseline. Run each prompt multiple times (repetition matters because of the non-determinism) and record how often your brand appears in the sources panel. That's your citation rate, ideally with a confidence interval, not one lucky run.
3. Make one change. Unblock OAI-SearchBot. Just that. Don't change five things at once or you'll never know which one moved the needle.
4. Wait for the lag. Remember the ~24 hour robots.txt propagation plus Bing re-indexing time. Give it days, not minutes.
5. Re-measure. Same prompts, same number of runs. Compare citation rate before and after.
6. Repeat per fix. Then add schema and measure. Then push index coverage and measure. One variable at a time.

Doing this by hand is doable but genuinely brutal. You're running hundreds of prompts, logging which sources appeared, and doing the stats. That's exactly what AI Citation Monitor automates: it tracks whether ChatGPT (plus Perplexity, Gemini, and Google AI Overviews) cites your brand, gives you citation rates with confidence intervals, shows competitor share of voice, surfaces which sources are winning, and tells you the next fix to make. The point is to turn "I think that helped?" into "citation rate went from 8% to 23%, here's the chart." There's a free instant check if you want to see your current standing before you change a thing.

Whether you use a tool or a spreadsheet, measure. The conceptual groundwork is in AI citation tracking and the citation rate glossary entry. A ChatGPT Search strategy you can't measure is just vibes with a robots.txt file.

How ChatGPT Search stacks up against the other engines

ChatGPT Search isn't the only answer engine you need to think about, and they don't all behave the same way. Perplexity is more aggressively citation-first and shows sources inline as it writes. Gemini is wired into Google's index and ecosystem. Google AI Overviews sit on top of classic search results. The retrieval plumbing, the crawlers, and the citation behavior differ enough that optimizing for one doesn't automatically win you the others.

If you're deciding where to focus, two head-to-heads help: ChatGPT vs Perplexity breaks down how each picks and shows sources, and ChatGPT vs Gemini covers the index and ecosystem differences. The short version: the answer-first, fact-dense, well-structured content that wins in ChatGPT Search tends to travel well to the others, but the crawler access and index presence are engine-specific, so you can't skip those per-engine. Good content is portable. Plumbing is not.

Putting it together

ChatGPT Search is a fine-tuned GPT-4o that decides when to hit the web, rewrites and fans out your question, pulls candidate pages from third-party search providers and partner feeds, and cites a small curated set in a sources panel (source). To show up, the priority order is fixed: allow OAI-SearchBot so you're eligible at all, get into the underlying index (read: Bing), write answer-first with high fact density so the model picks you out of the 85% it discards, add schema and freshness signals, then measure with enough runs to know a result is real.

None of this is magic, and I won't pretend the brackets always move overnight. The robots.txt fix is fast. Index and content changes compound over weeks. But the mental model is clean: access, then index, then selection, then measurement. Skip access and the rest is wasted. Nail access and the rest compounds. Get the door open first, then make the model want to quote you.

FAQ

What is ChatGPT Search?

ChatGPT Search is the web search feature built into ChatGPT that answers your question in plain language and then shows the sources it used in a panel under the answer. Per OpenAI, it runs on a fine-tuned version of GPT-4o that calls third-party search providers plus partner content feeds. So it is not a separate search engine, it is ChatGPT deciding to go look things up on the live web when your question needs current or factual information.

Does ChatGPT Search use Google or Bing?

OpenAI describes ChatGPT Search as using third-party search providers, and it never officially says Bing only. In practice the web search is widely understood to lean heavily on the Bing index, which is why getting your pages into Bing matters for showing up. Strong Google rankings still correlate with citations, so the honest answer is that Bing is the gate you must pass and Google is a quality signal that helps once you are inside.

How do I show up in ChatGPT Search?

Start by allowing OAI-SearchBot in your robots.txt, because that crawler builds the index ChatGPT Search reads from, and OpenAI says sites opted out of it will not appear in ChatGPT search answers. Then get your pages into the underlying index, write answer-first content with high fact density, and add schema. Access first, content second, measurement third.

What is the difference between OAI-SearchBot and GPTBot?

OAI-SearchBot builds the index that ChatGPT Search uses to find and cite live web pages. GPTBot is a separate crawler that collects content to train future OpenAI models. They do different jobs under different user agents, so blocking GPTBot only opts you out of training while blocking OAI-SearchBot removes you from ChatGPT Search entirely. Mixing them up is the most common robots.txt mistake of the last two years.

Why does ChatGPT cite some pages and not others?

ChatGPT retrieves far more pages than it cites. An AirOps study of 548,534 retrieved pages found only 15% were cited in the final answer, meaning 85% were read and discarded. The model picks the pages that most clearly and credibly answer the specific question, so answer-first structure, fact density, and trust signals are what tip the selection your way.

How is ChatGPT Search different from regular ChatGPT?

Regular ChatGPT answers from what the model already knows, which can be stale or made up. ChatGPT Search adds a live retrieval step, so the model goes out to the web, pulls current pages, and cites them under the answer. The model decides when to trigger search, usually for questions that need fresh, local, or factual information rather than general reasoning.