ChatGPT Search SEO: how to rank in SearchGPT

How ChatGPT Search actually works

Since late 2024, ChatGPT routes search-style queries through a retrieval pipeline rather than answering only from training. The public confirmations and the third-party reverse engineering both point to the same stack: Microsoft Bing's search index supplies the candidate pool, OpenAI's own crawler (OAI-SearchBot) feeds an internal cache, a GPT-class re-ranker picks the passages, and ChatGPT-User may then fetch a live copy of a few sources to ground the answer. Three different user-agents, three different jobs.

The practical consequence: if you're not in Bing's index, you are essentially invisible to ChatGPT Search. And if your robots.txt blocks OAI-SearchBot or ChatGPT-User, the pipeline can find you through Bing but can't verify or re-fetch you, which tanks your citation rate even when relevance is high.

The five signals that actually move the needle

1. Bing indexability (the unglamorous prerequisite)

You cannot rank in a search engine that doesn't know you exist. Submit your sitemap through Bing Webmaster Tools, verify the property, and use the IndexNow ping for new and updated URLs. Sites that have lived in Google's shadow for years often discover their Bing index coverage is half what they expected. Fix this first.

2. Passage quality, written for extraction

The re-ranker picks passages, not pages. A short, declarative answer at the top of an H2 — followed by the explanation — is the single biggest content lever. The same pattern that wins AI Overviews wins here: answer-first, citation-shaped, one claim per paragraph. We unpack the writing pattern in get cited by ChatGPT, Claude and Perplexity.

3. Recency, dated honestly

ChatGPT Search heavily prefers recent passages for anything time-sensitive — pricing, APIs, statistics, regulation, tooling. Real dateModified in your ArticleJSON-LD plus a visible “Last updated” line beats any amount of evergreen framing. Updating the date without updating the content stops working the moment the re-ranker compares the page against an older cached copy.

4. Entity coherence (sameAs and the open-web graph)

ChatGPT disambiguates entities — your brand, your authors — using the broader signal graph it inherits from the training corpus and the live index. A consistent author bio with a sameAsgraph that points to a real LinkedIn, GitHub, Wikipedia or Crunchbase page measurably raises citation rate for that author's pages. Conflicting or sparse entity signals are one of the most common reasons we see qualified sites get skipped.

5. Letting the right user-agents fetch you

Three OpenAI user-agents matter, and they are not interchangeable:

• GPTBot — trains future models. Blocking it does not affect retrieval at all.
• OAI-SearchBot — populates the ChatGPT Search index/cache. Blocking it makes you invisible to the search product.
• ChatGPT-User — fetches live URLs on behalf of a specific chat. Blocking it means ChatGPT cannot ground or verify against your current page.

Most accidental blocks come from CDN bot-protection rules that classify all three under one bucket. The full breakdown is in robots.txt for AI crawlers.

Three myths to ignore

• “ChatGPT only uses its training data.” True until late 2024, comprehensively false since. Search queries route through live retrieval; even conversational queries with a current-events flavour trigger a fetch.
• “You can submit your site to ChatGPT.”There is no submission endpoint, no console, no “index now” for ChatGPT itself. You submit to Bing, and to your own crawler-discoverable sitemap. That's it.
• “Blocking GPTBot improves your ranking.” It does nothing to ranking — it just removes you from training. If anything, blocking the wrong agent (the search or live-fetch ones) hurts citations.

Verifying you're cited

ChatGPT Search sends a referrer for some click-throughs and surfaces source chips inline for most answers — so you can both inspect server logs for the OpenAI user-agent fetches and prompt ChatGPT itself with your most important branded and unbranded queries to see whether you appear. The combined-signal approach is the only one that survives — a single check is always misleading. Full method in how to measure if AI assistants cite your site.

A 5-step playbook

1. Get fully indexed in Bing. Webmaster Tools, sitemap, IndexNow on every publish. Audit existing coverage before chasing anything else.
2. Allow the three OpenAI user-agents. Or pick deliberately — block GPTBot for training, allow OAI-SearchBot and ChatGPT-User. Verify on a CDN-level bot rule too, not just robots.txt.
3. Rewrite the lede of each ranking H2 as a single self-contained answer sentence. The rest of the section explains. That sentence is the candidate citation.
4. Date and re-date honestly. Real dateModified inArticleJSON-LD, visible “Last updated.” Edit the content when you bump the date.
5. Build the entity graph. Author bios with sameAs to LinkedIn, GitHub, Crunchbase, Wikipedia where applicable. A coherent organisation Organization node anchored across your sitemap.

FAQ

Want to see exactly which of the five signals your site is missing? Run a free audit, or read the terminology primer in our 2026 AEO/GEO/AI SEO glossary.