You blocked GPTBot
in robots.txt
to keep your content out of training runs — and then wondered why ChatGPT Search never cites your docs. Those are different systems. OpenAI’s own crawler docs say each bot is independent: allowing OAI-SearchBot
keeps you eligible for ChatGPT search answers while disallowing GPTBot
opts you out of foundation-model training.
This article is a Next.js App Router playbook for Generative Engine Optimization (GEO): how answer engines discover pages, which user-agents actually matter, how to configure robots.ts
and sitemaps, how to keep HTML crawlable, and what llms.txt
does — and does not — guarantee.
Treat “appearing in an LLM” as three separate pipelines:
| Pipeline | What it does | Typical bots / tokens |
|---|---|---|
| Training crawl | ||
| Collects public pages that may enter future model training | ||
GPTBot , ClaudeBot , Google-Extended (token), Common Crawl’s CCBot |
||
| Search / answer index | ||
| Builds or refreshes retrieval so answers can cite your URLs | ||
OAI-SearchBot , Claude-SearchBot , PerplexityBot , classic Googlebot / Bingbot (and partners) |
||
| User-triggered fetch | ||
| Downloads a specific URL because a human asked for it (or pasted a link) | ||
ChatGPT-User , Claude-User , Perplexity-User |
||
Blocking the training bot does not automatically block the search bot. OpenAI states this explicitly for GPTBot
vs OAI-SearchBot
. Anthropic documents the same split for ClaudeBot
, Claude-SearchBot
, and Claude-User
. Perplexity documents PerplexityBot
for search indexing and Perplexity-User
for live fetches.
ChatGPT Search can also partner with third-party search providers. OpenAI’s help center documents that rewritten queries may be sent to partners such as Bing (and others listed in that article). Independently, OpenAI recommends allowing ** OAI-SearchBot** if you want to appear in ChatGPT search answers. Practical implication: keep search bots allowed
From OpenAI’s crawler overview:
Used to surface websites in ChatGPT’s search features. Sites opted out of OAI-SearchBot
will not be shown in ChatGPT search answers, though they can still appear as plain navigational links. OpenAI recommends allowing it in robots.txt
and permitting its published IP ranges. Changes can take about 24 hours to propagate.
Crawls content that may be used to train generative foundation models. Disallowing GPTBot
signals that content should not be used for that training. It does not control ChatGPT Search eligibility.
Used when ChatGPT or Custom GPTs fetch a page because of a user action. OpenAI notes it is not used for automatic web crawling, not used to decide Search inclusion, and that robots.txt rules may not apply because the fetch is user-initiated. Manage Search with OAI-SearchBot
; treat ChatGPT-User
as a separate live-fetch channel.
Only visits pages submitted as ads on ChatGPT; not used to train foundation models. Relevant if you run ChatGPT ads — ignore it for organic GEO.
A common, defensible policy for content sites that want citations but not training:
User-agent: OAI-SearchBot
Allow: /
User-agent: GPTBot
Disallow: /
Anthropic’s help center (updated April 2026) defines three bots:
ClaudeBot
Claude-SearchBot
Claude-User
Anthropic honors robots.txt
(including non-standard Crawl-delay
) and warns that IP-blocking alone is unreliable because it can prevent the bot from reading your robots.txt
.
Perplexity’s crawler docs recommend allowing ** PerplexityBot** so your site can appear in Perplexity search results.
Perplexity-User
perplexitybot.json
/ perplexity-user.json
) — robots.txt alone is not enough when the edge drops the request. Google-Extended is a
Important trade-off: for OpenAI you can allow search and disallow training separately. For Google, Google-Extended
covers both Gemini training and Gemini grounding. Allow it if you want Gemini apps to ground on your content; disallow it if you want to opt out of those Gemini uses (Search itself stays separate via Googlebot
).
app/robots.ts
App Router can generate /robots.txt
from a typed file. Official docs: robots.txt file convention.
Example that keeps search/citation bots allowed, optionally opts out of training, fences private routes, and advertises the sitemap:
// app/robots.ts
import type { MetadataRoute } from 'next'
const SITE = 'https://example.com'
export default function robots(): MetadataRoute.Robots {
return {
rules: [
{
userAgent: '*',
allow: '/',
disallow: ['/api/', '/admin/', '/drafts/'],
},
// ChatGPT Search + live fetch
{ userAgent: 'OAI-SearchBot', allow: '/' },
{ userAgent: 'ChatGPT-User', allow: '/' },
// Training opt-out (optional — remove this rule to allow training)
{ userAgent: 'GPTBot', disallow: '/' },
// Claude
{ userAgent: 'Claude-SearchBot', allow: '/' },
{ userAgent: 'Claude-User', allow: '/' },
{ userAgent: 'ClaudeBot', disallow: '/' },
// Perplexity
{ userAgent: 'PerplexityBot', allow: '/' },
{ userAgent: 'Perplexity-User', allow: '/' },
// Gemini grounding/training token (allow if you want Gemini apps to use you)
{ userAgent: 'Google-Extended', allow: '/' },
],
sitemap: `${SITE}/sitemap.xml`,
host: SITE,
}
}
After deploy, open https://your-domain/robots.txt
and confirm the groups look right. Spoofed user-agents exist — for verification, OpenAI and Perplexity publish IP range JSON files; Anthropic currently points publishers at robots.txt rather than relying on IP blocks.
Crawlers need URLs to discover. Next.js can generate /sitemap.xml
from app/sitemap.ts
(docs):
// app/sitemap.ts
import type { MetadataRoute } from 'next'
export default async function sitemap(): Promise<MetadataRoute.Sitemap> {
const posts = await getPublishedPosts() // your data layer
return [
{ url: 'https://example.com', lastModified: new Date(), priority: 1 },
...posts.map((p) => ({
url: `https://example.com/blog/${p.slug}`,
lastModified: p.updatedAt ?? p.publishedAt,
changeFrequency: 'monthly' as const,
priority: 0.7,
})),
]
}
Then:
If ChatGPT Search partners with Bing for some queries, a healthy Bing index is still cheap insurance. A healthy Google index still matters for Google Search, AI Overviews / AI Mode, and anything that grounds on Google’s index.
Answer engines and classic crawlers are more reliable when the important prose is in the first HTML response. App Router Server Components and SSR help; shipping an empty shell that only fills after client JavaScript is a classic way to look invisible.
Practical checklist for Next.js:
<title>
, meta description, and Open Graph tags via the Metadata API — they travel with the document.alternates.languages
in the sitemap when you ship locales.200
responses for public pages — timeouts and soft 404s waste crawl budget.Structured data (JSON-LD for Article
, FAQPage
, Organization
, etc.) does not replace good prose, but clear headings, short definitional paragraphs, tables, and FAQ sections match how Bing’s own GEO guidance describes content that is easier to cite accurately.
Technical access is necessary but not sufficient. Pages that get cited tend to:
##
/ ###
headings that match how people ask questionsYou do not need a new CMS. You need pages that are the best extractable answer for a specific intent.
llms.txt
: useful map, not a ranking switch
Jeremy Howard’s llms.txt proposal suggests a Markdown file at /llms.txt
: site name as #
heading, a short >
summary, then ##
sections with curated absolute links. Optional companion files (for example full concatenated context) exist in the ecosystem. Docs platforms (and many agent workflows) already use this pattern.
What it is not: a formal standard enforced by OpenAI, Google, or Anthropic for citation ranking. It does not replace robots.txt
. It cannot block crawlers. Treat it as a curated table of contents for agents and humans who fetch /llms.txt
on demand — especially documentation sites — not as a guaranteed ChatGPT ranking lever.
In Next.js you can start with public/llms.txt
, or generate it from a Route Handler:
// app/llms.txt/route.ts
export function GET() {
const body = `# Acme Docs
> Official documentation for the Acme API and SDKs.
## Docs
- [Quick start](https://example.com/docs/quickstart.md): Install and make your first request
- [Auth](https://example.com/docs/auth.md): API keys and OAuth
## Optional
- [Changelog](https://example.com/changelog): Release history
`
return new Response(body, {
headers: {
'Content-Type': 'text/plain; charset=utf-8',
'Cache-Control': 'public, max-age=3600',
},
})
}
If you also serve Markdown mirrors of key pages (the proposal’s .md
suffix idea), agents get cleaner context than parsing your marketing HTML.
GPTBot
and assuming you left ChatGPT SearchOAI-SearchBot
allowed for Search answers.robots.txt
, hostile WAFcurl
does not show the answer text, many crawlers will not either.llms.txt
as access controlGoogle-Extended
while expecting Gemini groundingInvest in GEO when you publish public expertise (docs, tutorials, comparisons, research) and want referral traffic or brand citation from ChatGPT, Claude, Perplexity, Copilot, or Gemini.
De-prioritize (or fully opt out) when the product is private, paywalled, legally sensitive, or you deliberately do not want model/training or answer-engine reuse — then disallow the relevant bots and accept lower AI visibility.
No. Per OpenAI, Search visibility is governed by ** OAI-SearchBot**.
GPTBot
is the training-oriented crawler. You can disallow GPTBot
and still allow OAI-SearchBot
.OpenAI documents third-party search partners (including Bing) and its own OAI-SearchBot
. Do both: allow OAI-SearchBot
and keep a healthy presence in major indexes. Do not rely on a single secondary blog’s claim about exclusive Bing or Google dependency.
llms.txt
make ChatGPT cite me?
There is no public commitment from major AI labs that llms.txt
controls ChatGPT citation ranking. It is still worth shipping for docs and agent UX. Citations still depend on crawl access, indexability, and content quality.
Not automatically. Decide per pipeline: search/citation vs training vs user fetch. Many content sites allow search bots, allow or disallow training consciously, and keep sensitive paths disallowed for everyone.
No — Server Components and the Metadata / sitemap / robots file conventions are well suited to crawlable HTML. Problems come from client-only rendering patterns, not from App Router itself.
To show up in ChatGPT Search and other LLM answers: allow the search crawlers, do not let your WAF undo robots.txt, ship crawlable HTML from Next.js, submit sitemaps (and IndexNow where it helps), and write pages that answer specific questions clearly. Optionally add llms.txt
as a curated map. Optionally disallow training bots if that is your policy — without confusing them for search bots.
Ship the OAI-SearchBot
allow rule this week if it is missing. Everything else compounds on top of being fetchable.