ChatGPT is the answer engine with the most users and the least transparency about how it picks sources. Perplexity shows you every source it used; ChatGPT shows you one to three, sometimes none, and never explains the ranking. That opacity is why most “ChatGPT SEO” advice in 2026 is guesswork dressed up as method.

Here is the contrarian part up front — getting cited by ChatGPT is not the same job as getting cited by Perplexity, and the agencies selling you one “GEO retainer” that treats every engine identically are leaving citations on the table. ChatGPT rewards entity authority. Perplexity rewards structure. You need both, but the last 20% of the work is platform-specific, and this post is that 20% for ChatGPT.

How ChatGPT actually retrieves and cites

ChatGPT cites from two distinct surfaces, and confusing them is the first mistake.

The first is ChatGPT Search — OpenAI’s own search product, backed by its own index and a long-running data partnership with Bing. When a user runs a search-shaped query, ChatGPT queries that index, pulls candidate pages, and synthesises an answer with one to three citations. This surface behaves like a search engine — it has an index, it has a crawler, and getting into the index is a prerequisite for being cited.

The second is in-chat browse — the live web fetch that happens mid-conversation when ChatGPT decides it needs fresh information. This is not the index. It’s a real-time HTTP request to specific URLs, triggered by the model, using a different user agent. Browse can cite a page that isn’t in the Search index at all, if the model navigates to it directly.

Most pages get cited through the Search index. Browse matters for fresh or niche queries where the index is thin. You optimise for both, but the index is the volume play.

Why ChatGPT differs from Perplexity in citation behaviour

Both are answer engines. They are not the same animal.

Perplexity is structure-first. It shows 4–8 sources per answer, ranks them on domain authority plus recency plus structural extraction match, and rewards a clean rewrite inside 7–30 days. We covered that mechanic in detail in Perplexity citations tactics.

ChatGPT is authority-first. It shows fewer sources — one to three in 2026, down from three to five a year ago — and it weighs who you are more heavily than how your page is shaped. A page with a credentialed named author, brand mentions in publications the model already trusts, and a coherent entity footprint will beat a structurally cleaner page from an unknown domain. Schema still helps — but it helps ChatGPT attribute a passage, not decide whether to retrieve it.

The practical consequence — ChatGPT is slower to reward you and harder to game. A structural rewrite shows up in Perplexity in a week. The same rewrite shows up in ChatGPT in two to four weeks, and only if your entity authority clears the bar. That’s frustrating, and it’s also why ChatGPT citations are stickier once you earn them.

The content structure ChatGPT extracts from

ChatGPT still extracts from chunks — so the four-layer recipe still applies — but it weighs the layers differently.

What ChatGPT pulls most reliably:

  • Definition passages — a clean one-to-two-sentence answer to “what is X” directly under a question-shaped H2. ChatGPT loves to open an answer with a definition and it wants to lift that definition verbatim.
  • Comparison tables — row-level facts ChatGPT can quote and attribute. Comparison queries are a large share of buying-stage prompts.
  • Named-expert answers — a direct answer with a visible byline and a credential. This is where ChatGPT’s authority bias and your content structure meet.
  • Numbered procedures — “how to” prompts pull ordered steps; ChatGPT reproduces the steps and cites the source.

What ChatGPT skips — wall-of-text pages with no H3 breaks, hero pages with no depth, generic “ultimate guide” content that summarises instead of answering, and anonymous content on anything resembling a YMYL topic. The full mechanic is in the four-layer extraction recipe — for ChatGPT specifically, lead with the direct answer in ≤30 words and put the author’s name where the model can see it.

robots.txt: OAI-SearchBot vs GPTBot vs ChatGPT-User

This is the single most expensive mistake we find in audits, so be precise here. OpenAI runs three crawlers and they do three different jobs.

OAI-SearchBot indexes pages for ChatGPT Search. If you block it, you are not in the Search index, and the Search surface cannot cite you. Allow it.

ChatGPT-User is the in-chat browse agent — the real-time fetch triggered when ChatGPT navigates to a URL during a conversation. Block it and ChatGPT cannot read your page when a user asks it to. Allow it.

GPTBot crawls pages for model training. It has nothing to do with citations. Blocking GPTBot keeps your content out of future training runs but does not remove you from ChatGPT Search or browse. This is the one you can legitimately block if you have a content-licensing or IP reason — and it will not cost you a single citation.

So the rule is blunt — allow OAI-SearchBot and ChatGPT-User, decide GPTBot on policy grounds alone. We see sites that blanket-blocked “OpenAI” in 2024 over training fears and accidentally deleted themselves from ChatGPT Search. Check your file today. The full crawler map across every AI engine is in the AI crawler access policy.

Schema and llms.txt signals ChatGPT browse uses

Schema matters less for ChatGPT than for Perplexity or Google AI Overviews — but “less” is not “zero.”

When ChatGPT browse fetches a page, valid JSON-LD tells it who wrote the page, what organisation stands behind it, and what the page is about. That’s attribution. Without it, ChatGPT still reads your prose, but it tends to cite the bare domain or — worse — misattribute the claim. Article, FAQPage, Person and Organization markup, all validated, is the baseline. The schema half of the recipe is in the schema stack for AI citation.

llms.txt is the newer signal. It’s a plain-text file at your root that tells AI systems which pages are your canonical, high-value content — a curated map rather than a crawl. ChatGPT browse can use it as a navigation hint when it’s deciding which URL on your domain to fetch. It is not a ranking factor and it will not manufacture citations on its own. But on a large site it helps the browse agent land on your best page instead of a thin one. The spec and a generator pattern are in the llms.txt spec for 2026.

Brand-entity recognition in ChatGPT

This is the lever that moves ChatGPT more than any on-page tweak — and it’s the one most teams ignore because it lives off your own site.

ChatGPT decides whether to cite you partly on whether it recognises you as an entity. An entity is a thing the model has a coherent internal representation of — a company with a known name, a known category, a known set of facts. If ChatGPT has a clear entity for your brand, it cites you with confidence. If your brand is ambiguous or invisible, it hedges and cites a competitor it knows better.

What builds the entity — a Wikidata item with a Q-number, a consistent name and description everywhere the model crawls, schema.org Organization markup with sameAs pointing at verifiable profiles, and mentions in publications ChatGPT already trusts. Named authors with real LinkedIn profiles and credentials feed the same machine. None of this is fast. All of it compounds.

The contrarian implication — a smaller brand with a tight, well-defined entity beats a bigger brand with a messy one. We have watched a 12-person SaaS get cited over a category incumbent because the incumbent’s name collided with two other companies and ChatGPT couldn’t tell them apart. Disambiguation is half the battle, and we wrote it up separately in brand-name disambiguation.

How to measure your ChatGPT citation rate

You cannot improve what you don’t measure, and “I asked ChatGPT once and it mentioned us” is not measurement.

Build a fixed prompt set — 15 to 30 prompts your actual buyer types into ChatGPT during the buying cycle. Not vanity informational queries; decision-stage prompts. Run that set weekly. For each prompt, record three things — were you cited at all, what position were you cited in (first, second, third), and how did ChatGPT describe you in the surrounding text.

Citation rate is the share of the prompt set where you appear. Baseline it in week one. A clean B2B niche moves from 0% to roughly 20–30% on ChatGPT over 90 days — slower than Perplexity, because the entity layer takes time.

Position matters more on ChatGPT than anywhere else, precisely because it shows so few sources. Being cited fourth on Perplexity is a weak placement. Being “cited fourth” on ChatGPT means you weren’t cited — there is no fourth slot. The deeper framework, including the three tracking tools worth paying for, is in measuring AI citations.

5 concrete tactics with expected impact

Ranked by impact per hour of effort. Ship them in this order.

  1. Fix robots.txt — allow OAI-SearchBot and ChatGPT-User. Impact: binary. If you are blocking them you are invisible to ChatGPT; unblocking is the difference between a 0% ceiling and a real number. Effort: one hour. Do this today.

  2. Build the entity layer — Wikidata item, consistent Organization schema, named authors. Impact: high and compounding — this is the lever that lifts your ceiling. Effort: weeks, mostly off-site. Start now because it’s the slowest to mature.

  3. Rewrite your top five buyer-prompt pages to lead with a ≤30-word direct answer under a question H2. Impact: medium-high — gets you extracted once the entity layer lets you in. Effort: a few days per page. This is the four-layer recipe applied to specific prompts.

  4. Add comparison tables to commercial pages. Impact: medium — comparison prompts are a big slice of buying-stage queries and ChatGPT quotes table rows cleanly. Effort: half a day per page.

  5. Earn brand mentions in three publications ChatGPT already trusts. Impact: medium and slow — guest contributions, original research someone cites, a credible directory listing. Effort: ongoing. This feeds the entity layer from the outside.

Tactics 1 and 3 are quick wins you can finish this month. Tactics 2 and 5 are the programme. Skipping the programme means you plateau around month three.

Common mistakes

The failure modes we see in every ChatGPT audit, ranked by frequency.

  • Blocking OpenAI crawlers by accident. A 2024 training-fear robots.txt edit that nuked OAI-SearchBot. Most common, easiest to fix, most expensive to leave.
  • Treating ChatGPT like Perplexity. Pouring effort into structure and ignoring the entity layer. Structure alone gets you into Perplexity; it does not clear ChatGPT’s authority bar.
  • Anonymous content. No named author, no credentials, “by the Team.” ChatGPT cannot attribute it and quietly down-weights it on anything YMYL-adjacent.
  • Chasing search volume instead of prompt frequency. Optimising pages for keywords nobody types into ChatGPT. Mine prompt research, not keyword research.
  • No measurement. Shipping content, never running the prompt set, unable to say after a quarter whether anything worked.

A contrarian closing opinion

Here’s the take I’ll put my name on — ChatGPT citations are the hardest of the major engines to earn and the most worth earning, and that ratio is going to get worse, not better.

ChatGPT is compressing its citation count. One to three sources today; I expect one to two on most answers by mid-2027. As the slots shrink, the entity layer becomes everything — the model will cite the brands it already has a confident representation of and ignore the long tail entirely. The window where a structural rewrite alone could sneak you in is closing.

So the move is not to optimise harder on-page. It’s to start the entity work now — Wikidata, named experts, brand mentions — because it takes two to three quarters to mature and the brands that started in 2025 are already uncatchable in their niches. The on-page recipe is table stakes. The entity is the moat.

If you disagree, that’s a productive argument — bring the counterexample. And if you want a baseline before you commit, run the free AI visibility audit — you get a real ChatGPT citation report against your buyer prompts and a punch-list ranked by citation potential per hour. Or read the service packages and decide whether this ships in-house or with us. Either way the answer is a programme, not a one-off rewrite.