How to Rank in ChatGPT: Three Retrieval Paths, and What Actually Moves Each One

There is no ranking in ChatGPT. No position eleven, no index you can be number one in, no list your page sits on. There are three distinct ways a website ends up cited in a ChatGPT answer, and OpenAI documents them itself, as separate crawlers with separate purposes and separate robots.txt settings:

Almost every guide on this query treats those three as one thing and issues one list of tips. That is why the advice reads as unfalsifiable: create quality content, add structured data, get mentions. Those instructions target different mechanisms, on different timescales, with different odds of working — and one of them, below, has been tested and did not survive.

Worth disclosing before you read further: we are a GEO agency, and one of the three paths below is one nobody can sell you a fix for. We say so anyway.

How does ChatGPT rank websites?

It doesn’t rank them. It retrieves them, or it recalls them, or it reads them off a feed. Everything below follows from which of those three happened.

The starting fact is in OpenAI’s crawler documentation: four user agents with independent robots.txt settings. GPTBot crawls “content that may be used in training our generative AI foundation models.” OAI-SearchBot is “used to surface websites in search results in ChatGPT’s search features.” ChatGPT-User fires when a user asks ChatGPT to open a specific page, and OpenAI notes that “because these actions are initiated by a user, robots.txt rules may not apply.” OAI-AdsBot validates pages submitted as ads.

Blocking one does not block the others. A site that disallowed GPTBot to stay out of training remains fully eligible for citation through OAI-SearchBot — and a site that blocked everything with one broad rule is invisible on every path at once. That distinction resolves more confusion than any tip list on this topic.

Path 1: live retrieval, which is search engine optimization wearing a new hat

ChatGPT does not search the web on every prompt. Semrush’s clickstream analysis of over one billion lines of U.S. panel data across 17 months, published April 2026, found web search enabled on 34.5% of ChatGPT queries as of February 2026 — down from roughly 46% in late 2024. Two-thirds of answers involve no live retrieval at all, so two-thirds of the time this path is not in play.

When retrieval does fire, the underlying ranking is not yours to control directly. OpenAI’s help documentation states that ChatGPT “may share disassociated search queries with third-party search providers such as Bing.” A conventional search engine picks the candidate set. Your leverage is the leverage you have always had: technical crawlability and indexation, and pages that answer the query better than the alternatives.

The 87%-versus-8% problem, and why both studies are right

Two credible studies asked how often ChatGPT’s citations match conventional search rankings and got answers an order of magnitude apart.

Neither is wrong. They measured different things. Seer matched ChatGPT’s simplified query against the SERP for that query, and flagged the limitation themselves — a “limited” 100-query sample, plus Wikipedia citations that matched no SERP at all. Ahrefs matched against the original prompt on deliberately long-tail queries, and offered the mechanism that reconciles the two: query fan-out. ChatGPT decomposes a prompt into several sub-queries, retrieves for each, then merges the sets with techniques like reciprocal rank fusion. A page ranking sixth for four sub-queries beats a page ranking first for the prompt nobody actually issued.

The practical consequence is not “write quality content.” It is that the unit of optimization is the sub-query, not the head term. That changes the work in two concrete ways: the keywords you research now have to include the conversational, natural language variants a model would fan out to, and headings have to be structured as the questions themselves rather than as topic labels. Pages built that way enter more candidate sets — a content architecture decision, and what topical coverage in SEO content has always been for. Ahrefs flagged its own expiry date too: data collected in early July 2025, before further changes to ChatGPT’s search stack.

Path 2: training data, which you cannot move this quarter

GPTBot collects text for foundation model training, and training runs are periodic. Whatever you publish this month cannot reach the model’s parameters until a subsequent run, and publishing velocity does not change that. Any guide promising to influence what ChatGPT “knows” about your brand within 30 days is describing Path 1 or Path 3 and mislabelling it.

What does accumulate is the shape of your brand across the corpus. The largest public dataset on this is Ahrefs’ May 2025 analysis of 75,000 brands, about 74% of which had at least one AI Overview mention. Ranked by Spearman correlation with AI visibility:

• Branded web mentions — 0.664
• Branded anchors — 0.527
• Branded search volume — 0.392
• Domain Rating — 0.326
• Referring domains — 0.295
• Number of backlinks — 0.218

Two caveats the study states and most citations of it drop. The authors write plainly that “correlation ≠ causation” and that all factors showed “moderate to very weak correlations.” And, more important: this measured Google AI Overviews, not ChatGPT. Applying it to ChatGPT is an inference, not a finding.

Read with those caveats intact, the ranking still says something useful. Unlinked brand mentions correlate roughly three times as strongly as raw backlink counts, which reorders the work: being named on authoritative third-party sites — roundups, listicles, forums, industry press, review platforms — feeds a different machine than link acquisition feeds, even though earned links and earned mentions come from the same outreach motion. Reputation, reviews and credibility are not soft signals here; they are the trace the corpus keeps of your brand.

Neil Patel’s widely cited six-factor list — brand mentions, reviews, relevancy, age, recommendations, authority — comes from asking ChatGPT 100+ questions and having an analyst test 82 candidate factors. It is directionally consistent with the Ahrefs correlations, and directional is how to read it: the query set, coding scheme and correlation values were not published. He also reports 27.41% of responses in his sample were simply wrong, including recommending software companies as service providers.

Path 3: third-party indexes, feeds and the few domains ChatGPT actually cites

Semrush analyzed over 230,000 prompts and more than 100 million AI citations across 13 weekly snapshots between July and October 2025. Reddit and Wikipedia remained ChatGPT’s two most-cited domains throughout, with Medium, Forbes and LinkedIn close behind. Citation share moved violently inside the window — Wikipedia fell from roughly 55% of responses to under 20% in mid-September, Reddit from around 60% to near 10% — and the authors note the data “doesn’t tell us exactly why these shifts happened.”

That volatility is the finding. A strategy anchored to one platform’s share of voice is anchored to a number that halved inside a quarter.

For products, this path is documented rather than inferred. OpenAI’s March 2026 announcement on product discovery states that through the Agentic Commerce Protocol, “merchants share product feeds and promotions so their catalogs are fully represented in ChatGPT,” delivered directly or “through third-party providers like Salesforce and Stripe.” Shopify merchants are included automatically via Shopify Catalog, with “no additional work required from individual merchants.” OpenAI’s help centre adds that merchants are “ranked based on factors like availability, price, quality, and whether they are the maker or primary seller.”

So for an ecommerce catalogue, feed accuracy is now a visibility surface with a published specification. For a service business, the equivalent surface is the directories, review platforms and community threads retrieved on your behalf — largely the same footprint local search visibility already depends on.

What the evidence supports, and what it does not

Three pieces of standard advice have now been tested. One held, one did not, and one has never been implemented by anybody.

Structured data did not survive its best test. Ahrefs (Linehan, Guan, May 2026) ran a matched difference-in-differences study on 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 control pages, across four separate statistical tests. Google AI Overviews showed a statistically significant decline of roughly 12 daily citations per page. AI Mode and ChatGPT showed effects close enough to zero to be noise. Their own reading: “if you’re already doing the rest of the SEO work well, JSON-LD isn’t going to be the unlock.” Their caveats are real — the sample was already heavily cited pages, the window was 30 days, JavaScript-injected schema was excluded — and an earlier Search/Atlas analysis reached the same null. As Search Engine Land summarized the field, there are no peer-reviewed studies on schema’s impact on AI search visibility. Ship schema because it earns rich results in classic search, not as a ChatGPT tactic.

llms.txt has no reader. John Mueller, June 2026: “I don’t think anyone knows — it’s purely speculative for now (the file has existed for years, yet none of the AI systems use it — what does it mean?).” No AI platform has confirmed reading it.

On-page framing did move the needle, on a different engine. GEO: Generative Engine Optimization (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande — ACM SIGKDD 2024) tested nine content modifications against GEO-BENCH, 10,000 queries, and validated on Perplexity. Relative improvement on the position-adjusted word count metric: quotation addition ≈ +41%, statistics addition ≈ +31%, citing sources ≈ +28%, fluency optimization ≈ +28%, technical terms ≈ +18%, keyword stuffing ≈ −8%. The authors’ limitations matter: methods “may need to adapt over time as GEs evolve,” and the tests ran on Perplexity and a research engine, not ChatGPT. Read it as the strongest available evidence that verifiable quotations, named statistics and real citations raise a page’s odds of being pulled into a generated answer — and as evidence that the oldest trick in the book still fails.

Why “your ChatGPT ranking” is not a number

This is the part no competing guide on this query mentions, and the part that decides whether a tracking invoice is worth paying.

Rand Fishkin (SparkToro) and Patrick O’Donnell (Gumshoe.ai) ran 2,961 prompt executions across ChatGPT, Claude and Google AI in November–December 2025, published January 2026: 600 volunteers, each prompt repeated 60–100 times per tool at default settings.

• There is a less than 1 in 100 chance that ChatGPT or Google’s AI, asked the same question 100 times, returns the same list of brands in any two responses.
• The odds of identical ordering are roughly 1 in 1,000.
• Across 142 human-written prompts about a single product category, semantic similarity between responses averaged 0.081.

Layer personalization and memory on top and a shared, observable ranking dissolves entirely. A tool reporting that you are “position 3 in ChatGPT” is reporting one draw from a distribution and presenting it as a coordinate.

What is defensible is a visibility rate: the share of n repeated executions of a fixed prompt set in which your brand appears, with n large enough to be stable, run on a schedule, prompt set and sample size disclosed. Fishkin’s team recommends exactly this, and advises verifying that any vendor publishes a reviewable methodology.

What is verifiable is money. Ahrefs published its own numbers in June 2025: AI search accounted for 0.5% of traffic and 12.1% of signups over a 30-day window — a 23x conversion advantage, ChatGPT the majority of it. Patrick Stox attached the asterisk himself, estimating AI search users click links about 75% less than organic search users, and questioning whether the premium survives scale. Semrush adds a second check: ChatGPT’s outbound referral traffic grew 206% year over year to January 2026, but 21.6% of it goes to Google.

We build AI visibility programs around sampled visibility rates and referral revenue, because those two survive contact with a non-deterministic system.

A working order of operations

Each step names the path it serves. Steps that serve no path are not on the list.

1. Audit robots.txt for all four OpenAI agents separately. A blanket disallow written in 2023 to stay out of training still removes you from ChatGPT search today. Paths 1 and 2.
2. Fix retrieval fundamentals — render, speed, indexation, internal linking. A page a search engine cannot rank cannot be retrieved. Path 1.
3. Build content around sub-queries, not head terms. One question per page, answered in the first hundred words, adjacent questions on adjacent pages. Path 1.
4. Put verifiable quotations, named statistics and real citations in the page body. The KDD result, with its stated limits. Path 1.
5. Run an unlinked-mention program, not just a link program. Roundups, review platforms, community threads, industry press. Paths 2 and 3.
6. Audit your presence on the domains cited in your category, and on any feed specification that applies to your catalogue. Path 3.
7. Instrument measurement before buying tracking. A fixed prompt set, a sample size, a schedule, a referral revenue line. All paths.

The compounding is the point, and it shows up in conventional search long before AI referral volume is worth reporting. Anonymized, from our own portfolio: a local marketing firm went from under 10 organic clicks a month to over 300 a day inside 12 months, reaching 16.7k clicks against 987k impressions. A local contractor moved from 5 clicks a week to roughly 80 a day — 7.21k clicks, 494k impressions, 1.5% CTR at average position 9.2. And one property has held 175.8k clicks a year after the initial climb, which is the number that separates a program from a spike. More are documented in our case studies.

The mistakes that cost the most

• Blocking crawlers by accident. The most expensive line in this discipline is a robots.txt directive nobody has read since it was written.
• Treating the three paths as one. Publishing more blog posts will not change what the model was trained on. Digital PR will not get an unrenderable page retrieved.
• Buying position tracking. Two identical prompts return different brands more than 99 times out of 100. There is no position to track.
• Shipping schema as the AI unlock. 1,885 treated pages against 4,000 controls said otherwise.
• Publishing unreviewed generated text at volume. It competes for the same retrieval slots as everything else, carrying none of the quotations, statistics or citations the KDD study found actually moved visibility.

The free SEO audit covers crawler access, indexation and citation footprint, and our SEO consulting engagements start from the same diagnostic.

FAQ

How do you rank your website in ChatGPT?
You don’t rank — you become retrievable, recalled, or fed. Allow OAI-SearchBot and ChatGPT-User in robots.txt, make the page rankable in conventional search for the sub-queries a model generates from your target prompt, put verifiable quotations and named statistics in the body, and build unlinked brand mentions on the domains ChatGPT cites. Each serves a different one of the three paths.

How does ChatGPT rank websites?
It doesn’t maintain a ranking. On roughly 34.5% of queries as of February 2026 it retrieves live, using a search index partly sourced from third-party providers such as Bing, decomposing your prompt into sub-queries and merging the result sets. On the rest it answers from training data. Product results come from merchant feeds delivered through the Agentic Commerce Protocol.

How do you do SEO for ChatGPT?
The retrievable-answer half is conventional search engine optimization aimed at sub-queries rather than head terms, which is why SEO content and technical foundations still carry it. The half that is genuinely new is off-site: unlinked brand mentions correlated at 0.664 with AI visibility in Ahrefs’ 75,000-brand study, against 0.218 for backlink counts.

What is the ChatGPT ranking tool?
There is no ranking to measure, so treat any tool named that way with the SparkToro numbers in hand. What the credible platforms actually produce is a visibility rate — how often your brand appears across many repetitions of a fixed prompt set. Ask any vendor for their prompt set, their sample size per prompt, their refresh cadence, and their published methodology before you pay.

How do you rank in SearchGPT?
SearchGPT is ChatGPT’s search behaviour, which means Path 1. Seer Interactive found 87%+ of its citations matched Bing’s top organic results on a 100-query sample in early 2025; Ahrefs found 8% overlap with the top 10 on 15,000 long-tail queries in mid-2025. The reconciliation is query fan-out — so rank for the questions inside the question.

Does schema markup help you get cited by ChatGPT?
On the best available evidence, no. Ahrefs’ difference-in-differences study of 1,885 pages that added JSON-LD against 4,000 controls found effects near zero for ChatGPT and a statistically significant decline for Google AI Overviews. Implement schema for rich results in classic search, which is a documented benefit, and do not expect it to move AI citations.

Where this leaves you

Most advice on this query is unfalsifiable because it is written about a single imaginary mechanism. Separate live retrieval from training data from third-party feeds, and the tips sort themselves into three piles: “this is search engine optimization,” “this cannot work on your timeline,” and “this has been tested and did not hold.”

That sorting is the work. Our generative engine optimization program runs on sampled visibility rates and referral revenue, sits on the technical foundation every retrieval path depends on, and carries a written 90-day guarantee. We run our own properties too — one at roughly 6,000 organic clicks a month across 180 bilingual pages — which is where we test this before a client pays for it.

Talk to us about your AI visibility.