How to Rank in ChatGPT Answers: A Step-by-Step GEO Framework

A software company we know spent four years building the best-ranked comparison page in its category. It sits first on Google for the head term and has since 2022. In February, their head of marketing asked ChatGPT the same question a buyer would ask, watched it name three competitors, and did not see her own company mentioned once. She checked her analytics. Traffic was fine. Rankings were fine. The answer that thousands of buyers now see first simply did not contain her.

There is nothing mysterious about why. ChatGPT is not running a slightly different version of Google’s algorithm. It runs a retrieval pipeline with its own crawler, its own partner index, its own query rewriting layer, and its own selection logic for which passages get quoted, and it only runs that pipeline on some questions. This is a step-by-step framework for how to rank in ChatGPT by working through each of those stages in order, built on what OpenAI publishes about its own systems and what the published research actually measured.

What actually happens between a prompt and a ChatGPT answer?

Before you can rank in ChatGPT, it helps to know which of two very different machines is answering the question.

Path one is live retrieval. OpenAI’s own documentation describes it plainly: ChatGPT may search the web automatically when a question would benefit from current information, and when it does, it “typically rewrites your query into one or more targeted queries” that it sends to partner search providers, naming Microsoft and Shopify among them. The pages that come back get read, and the ones that answer cleanly get quoted with a clickable citation.

Path two is model memory. When ChatGPT does not search, the answer is assembled from patterns the model already absorbed during training, which means your brand either exists in that pattern or it does not. No amount of on-page work changes a memory-only answer in the short term, because there is no page being fetched to influence.

That single number reframes the whole exercise. A framework that only addresses crawlability and page structure is optimizing for about a third of the conversations. The other two-thirds are won off your own website, through the mentions, citations, and corroboration that shape what a model believes about your category. Both halves belong in the same program, which is how we scope Generative Engine Optimization engagements built for answer engines rather than results pages, and the steps below run in the order that dependencies actually fall.

Step one: confirm OpenAI’s crawlers can actually reach you

Everything else in this framework is wasted effort if this step fails. OpenAI runs four separate bots, and the single most expensive mistake made by companies trying to rank in ChatGPT is treating them as one switch. A developer who read a headline about AI training in 2023 and blocked everything with “GPT” or “OAI” in the user agent may have quietly removed the company from ChatGPT search answers permanently. OpenAI’s documentation states the rule directly: each setting is independent of the others.

The distinction most teams miss is the one between GPTBot and OAI-SearchBot. A publisher who wants to be quoted and linked in ChatGPT answers but does not want their archive absorbed into a training set can disallow GPTBot and allow OAI-SearchBot, and OpenAI supports exactly that combination. Deciding those two lines separately, in writing, is a five-minute conversation that most companies have never actually had.

“Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers.”

Robots.txt is only half the check. OpenAI also publishes the IP ranges its searchbot crawls from and tells site owners to confirm their infrastructure accepts that traffic, which matters because the block is frequently not in robots.txt at all. It sits in a firewall rule, a bot-management setting at the CDN, or a rate limiter that treats an unfamiliar user agent as an attack. We pull server logs before touching content in the audit-first methodology we run at the start of every engagement, because a site that has never logged a single OAI-SearchBot request has a plumbing problem, not a content problem, and every hour spent rewriting headings before fixing it is wasted.

Step two: get into the index ChatGPT borrows from

ChatGPT does not answer from a single proprietary index of the whole web. OpenAI’s help documentation says it sends its rewritten queries to third-party search providers, and it names Microsoft as one of them. That has a practical consequence almost nobody acts on when they set out to rank in ChatGPT: a page that Google has indexed beautifully and Microsoft’s index has never seen is invisible to the retrieval path, no matter how strong its Google performance looks.

Verifying coverage in a second index takes an afternoon. Register the property in Bing Webmaster Tools, submit the same sitemap already given to Google, and compare the indexed page count against the sitemap count. Gaps show up fast on sites with heavy JavaScript rendering, aggressive parameter handling, or a robots file written years ago for a single crawler. IndexNow, the open submission protocol Microsoft supports, pushes new and updated URLs on publish instead of waiting to be discovered, which shortens the lag between hitting publish and being retrievable.

The gap widens outside the United States. Search engine share varies enormously by country, and a brand whose entire technical setup assumes one crawler’s behavior tends to discover the problem only when it looks at a market where that assumption never held. That is a standing check in the global SEO programs we run across multiple markets and languages, and it is the cheapest coverage win available to most companies trying to rank in AI search results for the first time.

Step three: optimize for the rewritten query, not the prompt

This is the step that separates a real framework from a checklist, and it follows from one line in OpenAI’s documentation. ChatGPT does not search for what the user typed. It rewrites the prompt into one or more targeted queries first, and OpenAI’s own worked example turns a conversational question about a class of drugs into a compressed, dated, technical string containing the mechanism and the year.

So the page is never competing against the prompt. It is competing against the machine’s reformulation of the prompt, which is shorter, more entity-dense, and more specific than anything a person would type. Semrush’s clickstream analysis captured both ends of that: search-enabled ChatGPT prompts averaged 8.7 words in early 2026, up from 4.7 a year earlier, while prompts that did not trigger a search averaged 13.5 words. Between 65 and 85 percent of the prompts in that dataset matched no keyword at all in a 27 billion keyword database, which is a blunt way of saying most of this demand is invisible to a conventional keyword tool.

The practical response is to write pages that cover the whole cluster of reformulations rather than one phrase. Name the category the way a specification would, name the qualifiers buyers attach to it such as price band, use case, company size, and integration, and state the timeframe explicitly where recency matters. A page that says “pricing for teams under 50 seats, updated for 2026” catches a rewritten query that a page saying “affordable plans” never will.

Vocabulary is where this gets specific to your market. Every industry has terms buyers use, terms vendors use, and terms regulators use, and a fan-out will reach for whichever set carries the most signal. Working through that gap is a recurring exercise across the industries we build search programs around, because the language a company uses internally is very often not the language its buyers or its answer engines use.

Step four: build pages where a single passage can stand alone

Retrieval systems do not quote whole pages. They pull passages, usually a heading and the text beneath it, and evaluate each one on its own merits against a query. That changes what it means to optimize content for ChatGPT, because the unit of optimization is no longer the document. It is the section.

A passage that survives extraction has three properties. Its heading states the question in the words someone would ask it. Its first sentence answers that question completely, without depending on a paragraph three screens up for context. And it names its own subject rather than leaning on pronouns, because a passage that opens with “it typically takes about six weeks” is useless once separated from the page it came from, while one that opens with “a standard commercial roof inspection typically takes about six weeks” travels perfectly well.

Structured data is the reinforcement layer, not the mechanism. Article, FAQPage, Organization, and Product schema give a machine an unambiguous version of what the prose already says, which reduces the odds of a misparse, but no schema type has ever forced a citation. If the visible answer is buried, marking it up does not unbury it. Schema also carries a second job here: Organization markup with a consistent name, location, and identifier set is part of how a brand becomes a recognizable entity rather than a string of text.

One structural warning applies to modern sites in particular. If the substance of a page only exists after JavaScript executes, a retrieval fetch that reads the served HTML sees an empty shell. Server-side rendering is not a preference at that point; it is the difference between having content and appearing to have none. We treat that as a build requirement rather than an optimization, and it is one reason the articles we publish on our own blog are written with self-contained sections and served as plain HTML rather than assembled in the browser.

Step five: add the three elements the research actually measured

Most advice about how to get cited by ChatGPT is untested intuition, repeated confidently because nobody checks it. There is one widely cited exception. Researchers from Princeton University and IIT Delhi, with independent collaborators, built GEO-bench, a benchmark of 10,000 queries drawn from nine datasets across 25 domains, and tested nine different content modifications to see which ones changed a source’s visibility inside a generated answer.

Three modifications carried the results. Adding relevant quotations from credible sources produced the largest gain, up to 40 percent on the study’s position-adjusted visibility metric. Adding statistics and adding citations to supporting sources each delivered gains in the same neighborhood, roughly 30 to 35 percent. The pattern underneath all three is the same: a passage that carries verifiable, attributable evidence is easier for a generative system to justify quoting than a passage making the same claim on its own authority.

The finding that should interest smaller companies most is what happened by rank. Sources sitting in fifth position gained the most from these changes, with the citation tactic producing a 115.1 percent visibility increase for fifth-ranked sources, while already dominant sources saw their visibility slip. In a ranked list, position five is a consolation prize. In a synthesized answer, a well-evidenced fifth-place source can be quoted ahead of a vague first-place one, which is the single most encouraging structural fact about this channel for a challenger brand.

Applying this is unglamorous editorial work. Replace “studies show” with the study, the year, and the number. Attribute the quote to the person who said it. Link the source you drew the figure from instead of absorbing it silently. That editorial standard is visible in the GEO engagements documented in our portfolio, and it is worth noting that the same discipline is what Google’s quality guidelines have rewarded for years. This is not a separate content strategy bolted onto the old one.

Step six: build the corroboration that memory-only answers run on

Return to the 34.5 percent figure, because it sets the ceiling on how far on-page work alone can take you. Roughly two out of every three ChatGPT answers are written without fetching anything, which means the brands named in them are the brands the model already associates with the category. Nothing on your website participates in that decision.

What does participate is everything written about you elsewhere: trade publication coverage, industry roundups, comparison articles, review platforms, forum threads, conference programs, podcast show notes, and directory listings. Unlike link building, this work counts mentions that carry no link at all, because the model is absorbing text, not following hyperlinks. A trade magazine paragraph naming your company alongside the category and the city is doing real work even though a backlink audit would score it as nothing.

Consistency multiplies this. The same company name, the same category description, and the same location details repeated identically everywhere is what turns scattered mentions into one recognizable entity instead of three ambiguous ones. This is where smaller companies routinely beat larger ones, because a focused brand with sixty consistent mentions in one narrow category reads more clearly than a conglomerate with six thousand scattered across forty. It is the same asymmetry we build around in the search programs we run for small businesses, where narrowness is the advantage rather than the limitation.

There is a timing reality to accept here. Retrieval-path fixes can change an answer within days. Memory-path presence changes on the training cadence of the models themselves, which no publisher controls. Both are worth doing; only one of them pays quickly, and any vendor promising fast results on the second is describing something they cannot deliver.

Step seven: measure citations, not just clicks

There is no Search Console for ChatGPT, so anyone working out whether they rank in ChatGPT has to assemble the answer from three imperfect sources, and confusing them produces bad decisions.

The first is prompt testing. Maintain a fixed set of the questions your buyers actually ask, run them on a set schedule, and log three states for each: cited with a link, mentioned without a citation, or absent. Tedious, and still the only direct read on the outcome that matters. Vary the phrasing deliberately, since the fan-out means two wordings of the same question can retrieve entirely different sources.

The second is referral analytics. Traffic arriving from chatgpt.com is real and countable, and Semrush measured 206 percent year-over-year growth in ChatGPT referrals between January 2025 and January 2026. It is also concentrated: the top ten receiving domains took 30 percent of all referrals in that dataset, so a new entrant should expect small absolute numbers long before the trend line means anything.

The third is server logs, and this is the one people misread. Crawler hits are not visits. Cloudflare’s Radar measurement of crawl-to-refer ratios found AI platforms requesting enormous volumes of HTML for every referral they send back, with Anthropic’s crawler at roughly 70,900 page requests per referral in a June 2025 sample. Logs tell you whether OAI-SearchBot can reach the page and which pages it favors, which is a diagnostic, not a performance metric. Keeping those three streams separate in a report is a discipline we hold to in enterprise SEO programs where results roll up to a board, because a chart mixing crawls and sessions will overstate performance by orders of magnitude.

What does not work, and why it keeps getting sold

Two tactics dominate the current sales pitch around how to rank in ChatGPT, and neither survives contact with evidence.

Keyword stuffing is worse than useless. The GEO study tested it directly and found it underperformed the unmodified baseline, with the authors noting the technique offers little to no improvement despite being widely used in traditional SEO. Repeating a phrase does not make a passage more quotable; it makes it less readable, and readability is what the extraction step is grading.

The llms.txt file has no consumer. The proposal, a plain-text file summarizing a site for language models, has been circulating for years, and as of June 2026 Google’s John Mueller described its value as “purely speculative for now,” pointing out that the file has existed for years and none of the AI systems use it. No major AI platform has committed to reading it. Publishing one is harmless and costs an hour; treating it as an optimization strategy is not, because the hour would have paid for itself in the crawler-access audit from step one.

The reason both keep selling is that they are easy to deliver and easy to invoice. Confirming that a firewall is not blocking OAI-SearchBot, rewriting sixty section openings so each one can stand alone, and earning trade coverage that has no link in it are slower, less demonstrable, and considerably more effective. That is the unglamorous shape of the work, and it is the whole of what it currently takes to rank in ChatGPT.

Frequently Asked Questions

Sources