How ChatGPT Search works, from index to answer

A language model on its own is a snapshot. Everything it knows was fixed at training time, which makes it excellent at explaining concepts and unreliable at anything with a date on it: prices, product launches, reviews, availability, who currently leads a category. For a product with roughly 900 million weekly active users in early 2026, that gap was untenable, because a huge share of real questions are exactly the dated kind. Search grafts a live web connection onto the model so that answers about the present can be grounded in the present.

The scale of that graft is easy to underestimate. ChatGPT Search processes an estimated 250 to 500 million queries per week, which puts it in the same conversation as mid-sized traditional search engines, and it reached that volume in under two years. For comparison, Perplexity, the most search-native of the AI assistants, handles around 50 million weekly queries. Whatever share of your buyers use ChatGPT, a meaningful slice of their category questions now route through this pipeline instead of a results page.

The habit change around it is broad, and it is measured. McKinsey's October 2025 research found 50 percent of consumers already use AI-powered search intentionally as a primary way to find information and make buying decisions, and the Aeolyft 2026 U.S. Search Trends Report found 58 percent of Americans use AI weekly. In B2B, G2's 2026 research shows 51 percent of software buyers starting research with an AI chatbot more often than with Google, up from 36 percent seven months earlier, with ChatGPT the dominant chatbot for software research at 63 percent. When people say search behavior is shifting, this pipeline is the specific thing it is shifting into.

Understanding the pipeline matters because each stage is a separate gate your content can pass or fail. A page can be missing from the index, present but never retrieved, retrieved but never cited, or cited without your brand being named in the answer. Those are four different problems with four different fixes, and most teams cannot tell which one they have. The rest of this article walks the stages in order.

ChatGPT Search launched on Bing's index. That was the pragmatic choice: building a web-scale index from scratch takes years, and Microsoft's partnership made Bing's available immediately. But OpenAI has been steadily building its own eyes. It operates a dedicated crawler, OAI-SearchBot, documented in OpenAI's developer pages, and a Botify analysis found OpenAI roughly tripled its web crawl since August 2025. The trajectory points one direction: ChatGPT Search increasingly reads the web through OpenAI's own crawl rather than through a partner's.

For site owners this changes the checklist. OAI-SearchBot is now a first-class visitor in your server logs, and your robots.txt decisions about it have direct visibility consequences. Blocking it removes your pages from the candidate set ChatGPT reads, which means answers about your category get assembled from other people's pages. Some publishers block AI crawlers over licensing concerns, and that is a legitimate business decision, but a brand that wants to appear in answers should let the crawler in and make its pages easy to fetch. We profile the full cast of AI crawlers, and who feeds what, in the AI crawler boom, measured.

The transitional reality is that both indexes matter. Bing remains part of the supply chain, so Bing Webmaster Tools is no longer an afterthought: submitting pages there is submitting them, indirectly, to ChatGPT. Google Search Console matters too, both because Google indexing correlates with general crawlability and because Google's own AI features cite from Google's index. The boring plumbing of indexing has quietly become AI visibility work.

Not every ChatGPT answer involves search. The model first judges whether a question needs current information, and that judgment shapes everything a brand should care about. Conceptual questions, definitions and how-to explanations are usually answered straight from training data, with no citations and no live pages involved. Questions with recency signals, prices, comparisons, "best X in 2026" framing, or explicit requests for sources trigger retrieval. Buying questions, the ones that decide shortlists, sit squarely in the retrieval-triggering category most of the time.

That is quietly good news for marketers. The questions that matter commercially are disproportionately the ones grounded in live pages, and live pages are the part of the system you can change. When retrieval fires, ChatGPT typically reformulates the user's question into one or more search queries, which may differ from the literal words typed. A question about "tools to see if AI mentions my company" might become queries about AI visibility platforms, brand monitoring in chatbots, or specific tool comparisons.

The practical implication is to stop optimizing for single keyword strings and start covering question territory. A page that answers one buying question thoroughly, with the phrasings a model might reformulate into, beats ten thin pages chasing keyword variants. This is a real break from classic SEO instinct, and it is measurable: the Princeton and Georgia Tech GEO study found keyword stuffing performed near the bottom of all techniques for generative engines, while evidence density performed at the top.

Once queries are issued, the index returns candidates, and the system selects a small set of pages to actually read. This selection is the steepest funnel in the pipeline: millions of indexed pages, a handful read per answer. Search-style ranking gets a page into the candidates, but the final cut favors pages the model can extract a confident answer from. Studies of what survives the cut keep finding the same profile: pages that state answers directly, carry data and named sources, and are structured so a machine can lift a self-contained passage.

The domain-level data confirms how unlike a media plan the winners are. According to 5W Research's 2026 analysis, Wikipedia accounts for 13.15 percent of ChatGPT's U.S. citations and Reddit for 11.97 percent, more than a quarter combined, while the Wall Street Journal, New York Times and Bloomberg are absent from the top 20. Beyond the giants, Otterly's 2026 AI Citations Report, built on over a million data points, found a long fragmented tail of niche pages that win individual questions. The tail is where brands get in. We break down the full source data in what ChatGPT actually cites.

One structural note belongs here because the evidence conflicts. SE Ranking found roughly 71 percent of pages cited by ChatGPT carry structured data, which suggests schema correlates with citability. But Ahrefs tested the causal claim in May 2026 on 1,885 pages and found adding JSON-LD schema produced no measurable citation lift on pages that were already cited. The reconciliation is probably that schema helps parsing and discovery without boosting established pages. Ship it cheaply, expect nothing magical.

The final stage is generation: the model writes an answer grounded in the pages it read, attaching citations to specific claims. Two properties of this stage matter enormously for brands. First, being read does not guarantee being cited, and being cited does not guarantee being named. The model synthesizes; it may take a fact from your page while attributing the framing to another source, or cite your comparison article while recommending the competitor it describes. Mentions and citations are separate currencies and deserve separate tracking.

Second, the output is probabilistic. SparkToro measured under a 1 percent chance that two identical ChatGPT runs return the same brand list. Retrieval can pull different page sets run to run, and generation samples from probability distributions, so the same question produces different answers on different days, and sometimes different minutes. Any single check of what ChatGPT says about your category is a coin flip wearing the costume of a fact. Honest measurement means repeated runs and trend lines, a point we develop in how ChatGPT decides which brands to recommend.

This is precisely why Reachroller stores the complete answer text for every run it makes through the official ChatGPT API. When the number says you appear in 40 percent of runs for a question, you can open the runs and read the answers, including which sources were drawn on. A score you cannot audit against raw answers is a score you cannot trust, in this category more than most.

Citations carry links, and links carry visitors, so ChatGPT Search has become a measurable referral channel. The growth numbers are steep from a small base: WebFX's analysis of 2.3 billion sessions from January 2024 through December 2025 measured AI traffic growth of 796 percent, ChatGPT referral traffic grew about 12.8 times over 19 months, and SE Ranking's May 2026 study of 101,574 websites recorded a 36.7 percent single-month jump to an all-time high. ChatGPT alone accounts for about 92 percent of trackable LLM referral traffic.

The conversion story is where the studies genuinely diverge, and honesty requires showing the spread rather than cherry-picking the biggest number.

Ranges reflect different panels, industries and definitions of conversion; the consistent direction is that AI-referred visitors convert better than organic.

The mechanism behind the spread is intuitive: a visitor arriving from a ChatGPT citation has already read a synthesized comparison, had objections addressed and seen your brand endorsed in context. They arrive later in their decision. Keep the volume caveat in view, since AI referrals are still around 0.18 percent of sessions in some panels and much of the traffic hides in "direct", but understand that the channel's value is concentrated in intent, and in the recommendations that never produce a click at all.

The pipeline suggests its own checklist. For the index stage: keep pages crawlable for OAI-SearchBot and Bingbot, submit new URLs through Bing Webmaster Tools and Google Search Console, and fix anything that blocks rendering. For the retrieval stage: publish pages organized around specific questions buyers actually ask, with the answer stated completely in the opening paragraph. For the selection and citation stages: load pages with the evidence the GEO research validated, quotations, statistics and named sources, which boosted visibility in generative responses by up to roughly 40 percent in the Princeton and Georgia Tech experiments.

Then verify, because every stage is probabilistic. Run the question, wait, run it again, and track whether your appearance rate moves. The full working sequence, from picking questions to confirming the flip, is laid out in our playbook for getting mentioned by ChatGPT, and the honest warning is that doing it manually is a part-time job: dozens of runs per week, answers filed somewhere queryable, drafts written, indexing chased, rechecks scheduled.

Reachroller exists because that job is automatable. It runs your buyers' questions through the official ChatGPT API on a schedule, never scraping, scores mentions only when your brand name literally appears in stored answer text, and generates a publish-ready fix page for each question you lose, with the URL slug, title tag, meta description, schema markup and indexing steps included. After you publish, a recheck shows whether the answer flipped. The full loop is described on how it works, and Starter pricing is $29 per month with a three-day, 50-credit, no-card trial.

Three trends are worth watching. First, index independence: the tripling of OpenAI's crawl since August 2025 suggests Bing's role will keep shrinking, which makes your direct relationship with OAI-SearchBot more important every quarter. Second, scale: with weekly active users roughly doubling year over year and search queries in the hundreds of millions weekly, the surface is growing into the default research starting point G2's data already shows it becoming for software buyers.

Third, and most important for planning: the fundamentals of the pipeline are stable even as the details shift. Some index will be read, some pages will be selected, evidence-dense answers will be preferred, and outputs will remain probabilistic. Strategies built on those fundamentals, citable pages, third-party presence, indexing discipline and repeated measurement, survive model updates. Strategies built on quirks of the current version do not.

Our own position is easy to state. ChatGPT Search is the single most valuable AI surface for most brands today, which is why it is the first engine Reachroller supports in production, with Claude, Gemini, Perplexity and Grok built and rolling out behind it. Learn the pipeline, measure your place in it, and treat every lost answer as a page waiting to be written.