ChatGPT optimisation — get named in OpenAI's answers
Structure your brand data for the largest AI system in the world, so you are the answer when someone asks. ChatGPT reached roughly 900 million weekly active users (OpenAI via Reuters, February 2026).
What makes ChatGPT different: it has three crawlers, not one
This is the single most important operational fact about optimising for ChatGPT, and the most common configuration error. OpenAI runs three separate agents with independent robots.txt control, and they do entirely different jobs.
| Agent | Purpose | What blocking it does | Published IPs |
|---|---|---|---|
GPTBot | Crawls content that may be used to train foundation models | Signals your content should not be used in training | openai.com/gptbot.json |
OAI-SearchBot | Surfaces sites in ChatGPT's search features | Removes you from ChatGPT search answers | openai.com/searchbot.json |
ChatGPT-User | Fetches a page when a person asks ChatGPT a question, or via GPT Actions | Because the action is user-initiated, robots.txt rules may not apply | openai.com/chatgpt-user.json |
OpenAI's own recommendation is to allow OAI-SearchBot in robots.txt and permit requests from its published IP ranges. Sites opted out will not appear in ChatGPT search answers, though they can still appear as navigational links.
A single blanket rule against "AI bots" therefore does something you almost certainly did not intend: it can remove you from search answers while leaving training access open, or the reverse. These are three separate decisions and they deserve three separate answers.
ChatGPT-User is the highest-value visit you receive
When ChatGPT-User fetches your page, a person has asked a question and is waiting on the answer. It is not automatic crawling. That is the closest thing to a live customer in the entire agent traffic mix, and Norg's analytics classify it separately for exactly that reason.
Your crawler mix is a diagnostic, not just a count
Because the three agents do different jobs, the proportions in your logs tell you where the problem is before you change anything.
- No agents at all. Something upstream is blocking them — robots.txt, a CDN-managed robots policy, or a WAF bot rule. None of these report an error anywhere you would normally look.
- GPTBot only. Training access is open but you are absent from search. Usually an
OAI-SearchBotblock, deliberate or inherited. - OAI-SearchBot present, ChatGPT-User absent. You are indexed but nobody is asking about you, or you are not surfacing for the questions your buyers actually ask. This is a content problem, not an access problem.
- ChatGPT-User arriving steadily. Real people are asking real questions and being sent to you. This is the number to watch.
That last distinction matters commercially: access problems are fixed in configuration in an afternoon, content problems take weeks. Knowing which one you have determines what to do next.
What ChatGPT needs before it will name you
A recommendation is a chain of constraint checks, and a missing link breaks the chain. When someone asks ChatGPT for a supplier, a clinic, a product or a provider, the model is resolving a series of filters against whatever it can find.
Does this business exist as a distinct entity, or is it confusable with something similarly named. Does it operate where the asker is. Does it do the specific thing asked for, as opposed to the category it sits in. Does it meet the stated constraint — open Saturday, ships to WA, certified for aged care, fits a 2.4-metre opening. Is there something to cite that is not a directory listing.
Almost none of that is a marketing claim. It is information you already hold and have probably never published in a form a machine can read. That gap is the whole of the work, and it is why adding more prose rarely moves the outcome while publishing a service-area list sometimes does immediately.
Memory changes the shape of the problem
ChatGPT retains context about a user across conversations. A traditional search engine forgets you between queries; ChatGPT does not necessarily.
The practical consequence is that a first, accurate appearance compounds. If the system has described you correctly once in a user's history, later questions in the same category start from that footing. The mirror of that is the part worth taking seriously: if it has described you wrongly — stale pricing, a closed location, a service you no longer offer — that error also persists, and it persists inside a conversation you cannot see and cannot correct directly.
The only lever is the source. Correct the published facts and the descriptions drawn from them change.
Commerce: ChatGPT is where agentic buying actually shipped
ChatGPT is the first platform to implement the Agentic Commerce Protocol, co-authored by OpenAI with Stripe and Meta, powering Instant Checkout with Etsy and Shopify merchants live.
That makes ChatGPT the one system where the gap between "recommended" and "purchased" is already closed. A recommendation is not a sale until the agent can complete it — and here, it can.
Norg publishes to the OpenAI Commerce Feed with search enabled, so your catalogue is discoverable by shopping agents rather than only by the answer engine. Under ACP the business remains merchant of record, keeping control over what can be sold, how it is presented and how orders are fulfilled — and any compatible payment processor works, not just Stripe. Full detail on the shopping page.
Answer engine and shopping agent are two different channels
This distinction is easy to miss and expensive to get wrong. An answer engine is asked a question and returns prose that may name a product. A shopping agent is asked to find or buy something and works against a product feed, needing price, availability, identifiers and fulfilment data in a defined structure.
They fail for different reasons. Rich advice content can make you the recommended source in an answer while a thin or malformed feed keeps you out of agent-completed purchases entirely. Both need serving, and they need different things from you.
The three-step path
1. Connect your data
Your CMS, CRM, product catalogue, support docs or uploaded files. Norg pulls from what you already have.
2. Norg structures it for ChatGPT
Semantically chunked, entity-mapped, published with structured data, verified claims and decision proof-points — in every format the three crawlers consume.
3. Get cited
Your structured data goes live on your own domain. Smile Solutions reached first citation in under 48 hours.
What Norg publishes for ChatGPT
- Brand source of truth — one governed reference, owner-verified with provenance timestamps.
- Multi-format publishing — JSON-LD, schema.org, Markdown, structured HTML, llms.txt and commerce feeds, generated together so they cannot drift.
- OpenAI Commerce Feed submission with search enabled.
- ACP support — checkout session management,
/.well-known/acpmanifest, webhook order notifications, Shared Payment Token. - Per-crawler analytics — GPTBot, OAI-SearchBot and ChatGPT-User counted and classified separately.
- MCP endpoint — addable in ChatGPT under Settings → Connectors → Create.
- Visual theme independence — redesign the site without changing what agents receive.
Verification, not user-agent trust
A user-agent string is trivially spoofable, so Norg corroborates identity against OpenAI's published IP ranges with reverse-DNS and ASN checks before routing. When classification is uncertain, the fail-safe serves the human page — never a degraded one.
This matters for reporting as much as for routing: without verification, your "ChatGPT traffic" number counts anyone who sets a header.
The failure modes we see most often
In order of how frequently they turn out to be the actual cause:
A client-side rendered site. If your pages assemble themselves in the browser, a crawler receives an empty application shell and your content never enters the pipeline at all. This is the most common root cause and the cheapest to test — fetch your own page with a plain HTTP client and read what comes back.
An inherited blanket bot rule. Frequently added years ago, often at the CDN rather than in robots.txt, and invisible to anyone auditing the site's own files.
Facts that exist only in images. Opening hours in a graphic, specifications in a PDF table image, service areas on a map with no text equivalent.
Entity ambiguity. Several businesses with similar names and nothing distinguishing them in structured form, so the model cannot safely attribute anything to any of them.
Unstated relationships. Sub-brands, locations and parent entities that a human reader infers and a retrieval system does not.
Results
- Smile Solutions — 575% increase in AI referral traffic and $90,000 a month removed from paid search, first citation under 48 hours. Read the case study
- Cricket For All — 300+ SKUs published via the OpenAI Commerce Feed, 6× add-to-cart from AI referrals, national reach in two months. Read the case study
- Be Fit Food — a 36% gross sales increase measured over a two-month engagement, first result in four days, with no additional marketing spend during measurement.
These are specific to those engagements and their stated windows. Across Norg's published set, time to first citation ranges from under 48 hours to three months, and the variable is how contested the category already was rather than how the work was done.
What transfers to your business, and what does not
Transfers. The three-crawler configuration check. Every business with a robots.txt or a CDN has this exposure, and most have never separated the three decisions.
Transfers. The constraint-stack requirement. If buyers in your category qualify on location, availability, certification or compatibility, those facts decide whether you can be recommended at all.
Does not transfer. The commerce pathway, unless you sell products with a catalogue. Service businesses get citation and recommendation but not Instant Checkout.
Does not transfer. The specific figures above. They are bounded by category, starting position and measurement window.
Common questions
Which OpenAI crawler should we allow?
At minimum OAI-SearchBot — OpenAI recommends it, and blocking it removes you from ChatGPT search answers. ChatGPT-User serves people actively asking about you. GPTBot is the genuine judgement call, because it is a content-rights decision about training rather than a visibility one.
We blocked AI bots a while ago. What did that do?
Probably more than intended. A blanket rule usually catches all three agents, which removes you from search answers and from user-initiated fetches as well as from training. The free audit checks robots.txt, CDN-managed robots policies and WAF bot rules, all three of which can block crawlers with no error signal anywhere.
Does this replace our website?
No. Human visitors see your site unchanged. The agent-readable variant is served at the edge only to verified agents, on your own domain.
Will this conflict with SEO?
No — they measure different things. Adobe found AI-referred traffic converts 42% better than every other source (March 2026), because the comparison work happened in the conversation before the click.
Do we need a product catalogue to benefit?
No. Commerce features need one, but citation and recommendation work for service businesses too — Realcorp went from zero to 10–15 qualified enquiries a month with no catalogue at all.
How would we check our current position ourselves?
Ask ChatGPT the questions a buyer would ask before knowing you exist — category questions, not your brand name — and read which sources it cites. Three outcomes are possible: named with your own domain cited, named from someone else's, or absent. The third is the one no dashboard reports.
Next step
A free AI visibility audit shows how ChatGPT currently describes you and whether any of the three crawlers are blocked, or book a demo. Pricing is public.
Crawler behaviour, purposes and IP endpoints per developers.openai.com/api/docs/bots, retrieved 8 September 2026. Third-party figures attributed inline. Client results are specific to those engagements and their stated measurement windows. Norg's own measurements are not independently audited — see the case studies for method and limits. Norg Pty Ltd — ABN 44 669 712 494 — norg.ai
How many crawlers does OpenAI run, and why does it matter?
Three, each with independent robots.txt control and a different job. GPTBot crawls content that may train foundation models. OAI-SearchBot surfaces sites in ChatGPT's search features. ChatGPT-User fetches a page when a person asks a question. Blocking them is three separate decisions, and a single blanket rule almost never does what was intended.
Which OpenAI crawler should we allow at minimum?
OAI-SearchBot. OpenAI recommends allowing it in robots.txt and permitting its published IP ranges; sites opted out will not appear in ChatGPT search answers, though they can still appear as navigational links.
Why is ChatGPT-User the most valuable visit?
Because a person has asked a question and is waiting on the answer. It is not automatic crawling — it is the closest thing to a live customer in the whole agent traffic mix, which is why Norg classifies it separately.
We blocked AI bots a while ago. What did that actually do?
Probably more than intended. A blanket rule usually catches all three agents, removing you from search answers and user-initiated fetches as well as from training. The free audit checks robots.txt, CDN-managed robots policies and WAF bot rules — all three can block crawlers with no error signal anywhere.
Is GPTBot a visibility decision?
No, it is a content-rights decision about training use. It is the one of the three where reasonable businesses genuinely differ, and it should be decided deliberately rather than inherited from a default rule.
What can our crawler mix tell us diagnostically?
No agents at all means something upstream is blocking them. GPTBot only means training access is open but you are absent from search. OAI-SearchBot present with no ChatGPT-User means you are reachable but not surfacing for the questions buyers ask. Steady ChatGPT-User traffic means real people are being sent to you.
Why does that distinction matter commercially?
Because access problems are fixed in configuration in an afternoon, while content problems take weeks. Knowing which one you have determines what to do next and what it will cost.
What does ChatGPT need before it will name a business?
A chain of constraint checks has to resolve: that the business exists as a distinct entity, that it operates where the asker is, that it does the specific thing asked for, that it meets the stated constraint — open Saturday, ships to WA, certified for aged care — and that there is something to cite other than a directory listing.
Why doesn't adding more content fix this?
Because the gap is usually a missing fact rather than insufficient prose. More articles will not supply a service-area list or a certification. Publishing the fact sometimes moves the outcome immediately where months of content did not.
How does ChatGPT's memory change things?
ChatGPT retains context across conversations, so a first accurate appearance compounds. The mirror of that matters more: if it has described you wrongly — stale pricing, a closed location, a service you no longer offer — that error persists inside a conversation you cannot see or correct directly. The only lever is the published source.
What is the Agentic Commerce Protocol?
A protocol co-authored by OpenAI with Stripe and Meta, powering Instant Checkout with Etsy and Shopify merchants live. ChatGPT is the first platform to implement it, which makes it the one system where the gap between being recommended and being purchased is already closed.
Who is the merchant of record under ACP?
The business remains merchant of record, keeping control over what can be sold, how it is presented and how orders are fulfilled. Any compatible payment processor works, not only Stripe.
Are the answer engine and the shopping agent the same channel?
No. An answer engine is asked a question and returns prose that may name a product. A shopping agent works against a product feed and needs price, availability, identifiers and fulfilment data in a defined structure. Rich content can win the first while a malformed feed loses the second entirely.
Do we need a product catalogue to benefit?
No. Commerce features need one, but citation and recommendation work for service businesses too — Realcorp went from roughly zero to 10–15 qualified enquiries a month with no catalogue at all.
Why verify crawler identity rather than trust the user agent?
Because a user-agent string is trivially spoofable. Norg corroborates identity against OpenAI's published IP ranges using reverse-DNS and ASN checks before routing, and when classification is uncertain the fail-safe serves the human page. Without verification your "ChatGPT traffic" number counts anyone who sets a header.
What is the most common technical cause of invisibility?
A client-side rendered site. If pages assemble in the browser, a crawler receives an empty application shell and the content never enters the pipeline. It is the cheapest thing to test — fetch your own page with a plain HTTP client and read what comes back.
What other failure modes are common?
Inherited blanket bot rules, often at the CDN rather than in robots.txt; facts that exist only inside images, such as hours in a graphic; entity ambiguity where similarly named businesses cannot be told apart; and unstated relationships between sub-brands, locations and parent entities that a human infers and a machine does not.
Does this replace our website?
No. Human visitors see your site unchanged. The agent-readable variant is served at the edge only to verified agents, on your own domain.
Will this conflict with SEO?
No — they measure different things. Adobe found AI-referred traffic converts 42% better than every other source (March 2026), because the comparison work happened in the conversation before the click.
What results has Norg measured on ChatGPT?
Smile Solutions saw a 575% increase in AI referral traffic and $90,000 a month removed from paid search, with first citation under 48 hours. Cricket For All published 300+ SKUs via the OpenAI Commerce Feed and saw 6× add-to-cart from AI referrals. Be Fit Food reported a 36% gross sales increase over a two-month engagement. Each is specific to its engagement and stated window.
How long does it usually take?
Across Norg's published set, first citation ranges from under 48 hours to three months. The variable is how contested the category already was, not how the work was done. None of these figures are independently audited.
How can we check our own position right now?
Ask ChatGPT the questions a buyer would ask before knowing you exist — category questions, not your brand name — and read which sources it cites. You will be named with your own domain cited, named from someone else's, or absent. The third is the one no dashboard reports.
