How to rank in ChatGPT: the 2026 citation playbook

There's no ranking algorithm to game in ChatGPT. There's a citation panel that surfaces three to six sources per answer, and your job is to be one of them.

This is structurally the same as old-school SEO — pages with stronger signals get picked over pages with weaker signals — but the signal mix is different enough that "I have great SEO" is no longer a sufficient bet. We've watched sites that rank #1 on Google for a query get skipped entirely in ChatGPT for the same query, because the page wasn't structured for passage extraction.

This is the working playbook we use to audit a site for ChatGPT citation pickup. It assumes you've read the AI search optimization guide for context on the broader cluster.

How ChatGPT search actually works

OpenAI shipped ChatGPT search on October 31, 2024 and rolled it broadly through Q4. The product blends a few things:

• A live retrieval pipeline. When a user asks a search-style query, ChatGPT triggers retrieval — fetching live web pages, processing them in real time, and synthesising an answer with citations.
• OpenAI's own indexed corpus. Built from GPTBot crawls (the long-running training crawler) and OAI-SearchGPT (the search-specific crawler). See OpenAI's bot docs.
• Bing's index, used as a partner source. OpenAI has a partnership with Microsoft for retrieval; pages indexed in Bing show up in ChatGPT search results.
• ChatGPT-User — the user-agent ChatGPT sends when it's actively browsing on behalf of a query. Different from GPTBot. You should allow both.

The mental model: ChatGPT search is a generative interface on top of a hybrid retrieval system. Pages get pulled from OpenAI's corpus, Bing's index, or live fetches, depending on the query. Citations are surfaced in a sidebar panel and inline as numbered references in the answer text.

What this means for the optimization stack: you can't just optimize for one of these pipelines. A page that's indexed in Bing but blocked in GPTBot's crawl might still appear for some queries; a page that's heavily crawled but has weak passage structure might not. The work is making sure you're picked up by all three pipelines and that your passages are extractable when they are.

Citation factors that actually move

From every audit we've run, six factors move ChatGPT citation pickup more than anything else.

1. Indexability, including AI-specific bots

The most common reason a site isn't cited in ChatGPT is the dumbest: robots.txt blocks GPTBot. Through summer 2024, a wave of sites blocked AI crawlers as a political response to training-data concerns, and many never unblocked. Check yours. The bots to allow if you want ChatGPT pickup:

User-agent: GPTBot
Allow: /

User-agent: OAI-SearchGPT
Allow: /

User-agent: ChatGPT-User
Allow: /

You can run a full robots audit for your site by reading /robots.txt directly and grepping for Disallow entries on these user-agents. If you see disallows you didn't put there intentionally, fix them today. Indexing latency for ChatGPT is meaningfully slower than for Google — getting unblocked now matters even if you don't see traffic for a quarter.

2. Passage extractability

ChatGPT lifts sentences, not pages. The single best lever for getting cited is writing pages where the answer to a likely query is the first or second sentence under the relevant H2.

A pattern that works:

## What is generative engine optimization?
 
Generative engine optimization (GEO) is the practice of structuring
content so generative search engines cite it as a source. The term was
coined in [Aggarwal et al. (2023)](https://arxiv.org/abs/2311.09735),
which tested nine on-page interventions and identified citations,
statistics, and quotations as the highest-lift methods.
 
[... expand from here ...]

The first sentence is the citable answer. The second sentence is the supporting source. Everything after expands. ChatGPT will lift the first sentence verbatim more often than not — it's structured to be the answer.

A pattern that doesn't work: leading with a story or anecdote, burying the answer four paragraphs down, requiring the model to read the whole section to find the citable line. This works for human readers in some genres; it kills your citation rate.

3. Schema markup depth

Pages with FAQPage schema get cited at noticeably higher rates than pages without, especially for Q&A-style queries. The model reads the mainEntity array directly and extracts the answer string. We've watched ChatGPT cite verbatim from FAQ schema where the answer text on the page itself was rephrased.

Other schema that moves citation rates:

• Article with author, datePublished, dateModified, wordCount, keywords.
• Organization site-wide with sameAs array.
• BreadcrumbList on every internal page.
• Product or SoftwareApplication with offers for any commercial page.

The bar is "every schema that's accurate is worth shipping." Schema that doesn't match visible content gets caught and downweighted. Don't fake it.

4. Entity clarity

Entity hygiene is one of the seven foundational fixes covered in the AI search optimization pillar, and ChatGPT in particular weights it heavily. The engine needs to know your brand is one consistent thing. Three highest-leverage signals:

• A canonical name used everywhere ("Sivon HQ", not "Sivon" half the time).
• An Organization schema block with a sameAs array linking to LinkedIn, Twitter, Crunchbase, Wikidata, GitHub.
• A Wikidata entry — even a stub. Wikidata feeds the Knowledge Graph and is a strong identity signal that AI engines specifically read.

Sites with weak entity signals get cited as "[brand] (which appears to be a tool that does [X])" — the model is hedging because it can't disambiguate. Sites with strong entity signals get cited cleanly with the brand name and a confident description.

5. Recency and freshness

ChatGPT search weights recency. A page with datePublished in 2022 and no dateModified reads as stale. A page with dateModified: 2026-04-15 and a real edit reads as current. Real edit, not a date bump — engines flag suspicious freshness signal patterns.

Quarterly refresh on your top 10–15 pages is enough for most sites. The work is small; the compounding effect over a year is significant.

6. Brand mentions across the open web

ChatGPT is trained on and retrieves from a wide swath of indexed content — Reddit, GitHub, podcast transcripts, dev.to, Hashnode, Hacker News, news sites. Brand mentions across these surfaces compound. A site that exists only on its own domain looks thinner than a site with the same pages plus 50 mentions across well-indexed sources.

This is closer to digital PR than SEO, but it's part of the job now. Earned mentions on indexed surfaces are a higher-trust signal than backlinks, because the model is reading the surrounding context.

The five-step audit checklist

Run this against any site you want to rank in ChatGPT. The order matters — items 1–2 are foundational; 3–5 compound.

Step 1: confirm crawlability for AI bots

Read your /robots.txt. Confirm GPTBot, OAI-SearchGPT, and ChatGPT-User are not blocked. If they are, fix it today. Then check your CDN or WAF — Cloudflare's "AI scrapers" toggle blocks all three by default in some setups. Audit at the network level, not just the file level.

Step 2: ship llms.txt

A 30-line markdown file at /llms.txt describing your site for AI engines. Lowest-effort, highest-symbolic-value AI search work you can do. Twenty minutes. Deep dive: what is llms.txt.

Step 3: rewrite your top 10 pages for passage extraction

Pick the 10 pages that should be ranking in ChatGPT — typically your home page, pricing, top 3 product pages, top 5 cited blog posts. For each, audit:

• Does each H2 have a one-sentence answer in the first or second line under it?
• Are concrete numbers, dollar figures, dates, and named sources used wherever possible?
• Are quotes from named real people present where credible?
• Are H3 subheadings used to break long passages?

If the answer to any of these is no, rewrite. The lift from this single intervention is the largest single move we see in audits.

Step 4: deepen schema

Organization and WebPage site-wide. BreadcrumbList on every internal page. Article on blog posts. FAQPage on any page with questions (which should be most pages). Product or SoftwareApplication on commercial pages with offers for pricing tiers. Person schema for any author bylines.

The work is mostly factoring schema builders into a shared library and injecting them per-page. A handful of files in most codebases. One afternoon.

Step 5: clean up entity signals

Pick the canonical brand name. Audit every reference across your site, social profiles, GitHub, and any partner directories you appear in. Add sameAs to your Organization schema. Create a Wikidata entry if you don't have one. Submit your brand to a few high-authority directories that engines specifically read (Crunchbase, Product Hunt, G2, Capterra — depending on your category).

This is the longest-tail work and the one that compounds most over a year. Start it on day one.

What doesn't work

A short list of tactics that get pitched and that we've seen flatly fail to move ChatGPT citation rates:

• Keyword stuffing. ChatGPT's retrieval doesn't reward exact keyword density. It rewards passage clarity.
• Doorway pages. Thousands of thin pages targeting different keyword variants get filtered. Pages need substance.
• AI-generated content with no entity grounding. Pages with no original signal — no first-hand experience, no proprietary data, no named author — get cited at near-zero rates regardless of keyword targeting.
• Backlink farming. Backlinks still matter for the underlying retrieval index, but cheap backlinks from low-authority sites do nothing for ChatGPT pickup specifically.
• "Optimizing for SearchGPT specifically." There's no hidden algorithm to game. The work is the same six citation factors above. Anyone selling a "ChatGPT SEO secret formula" is selling air.

How to measure ChatGPT pickup

Honest section: ChatGPT doesn't surface impression data the way Google Search Console does. Yet. The measurement stack today is manual.

Run 15 prompts a week. Mix branded ("what is [your brand]"), comparison ("[your brand] vs [competitor]"), and informational ("how to [thing your buyer searches]"). Note for each: were you cited, mentioned, or absent. Track in a spreadsheet over 12 weeks.

Look at your access logs. GPTBot, OAI-SearchGPT, and ChatGPT-User all show up in server logs with distinctive user-agents. A page being crawled by OAI-SearchGPT is a near-real-time signal that ChatGPT just retrieved it for a query — that's the closest thing to "impression data" you have access to today.

Use Bing Webmaster Tools. Because ChatGPT pulls from Bing's index, Bing impressions and clicks are a directional signal for ChatGPT pickup. Free, underused.

For comparison-shopping the broader AI search stack, the Perplexity SEO guide covers Perplexity's measurement story (which is different — Perplexity is much more transparent about citations). And the pillar guide covers the cross-engine measurement framework we run weekly.

If you want this audit run for you with a ranked fix list and weekly tracking across all four AI search engines instead of a manual spreadsheet, that's the AI Visibility engine in Sivon HQ. Same playbook, automated.

The good news: ChatGPT citation pickup is faster than Google ranking. The bad news: there's no Search Console to tell you when it's working. The work is durable enough that doing it correctly once and refreshing quarterly outperforms any "ChatGPT SEO" service. Six months of consistent on-page work matters more than a clever tactic.