How to get cited by ChatGPT: the operator's playbook for 2026.

About 18% of ChatGPT conversations trigger a web search, and only ~15% of the pages it pulls in get cited.

That is the surface. The operational problem has two parts: enter the retrieval set, then give the answer system a reliable reason to select your page as a source. A cited answer can transfer trust even without a click, but visibility, referral traffic, and revenue must still be measured separately.

This playbook covers the parts a site owner can control. It does not promise a citation deadline because the source-selection system, model, index, prompt, and competing pages all change the result. For the cross-engine technical framework, start with LLM SEO.

What ChatGPT actually reads

ChatGPT can answer from model knowledge and can use live web retrieval:

1. Model knowledge. Information learned during training is not a live copy of the web, and it does not provide a controllable path to a current citation.
2. Live web retrieval. ChatGPT search can surface and link to current web pages. OpenAI documents OAI-SearchBot for this discovery layer and ChatGPT-User for user-triggered visits.

For publishers, live retrieval is the actionable layer: it can discover a current canonical page, attach a link, and send a referral. Training permission is a separate policy choice, not a shortcut into search results.

The three things that lift citation rate

After running this retrofit on 5 client sites, the same three patterns moved the needle every time.

1. The answer-first paragraph

Every page that targets a citable query needs a 60 to 120 word direct answer in the first paragraph after the H1. Not a hook. Not context. The answer.

Use the literal question as an H2 immediately under the H1. The pattern an LLM extractor recognizes is: H2 question → 80-word paragraph that answers it → optional follow-up nuance.

Bad opening:

“In today’s fast-moving digital landscape, the question of X has become increasingly important. In this post, we’ll cover the key concepts and explore why they matter.”

Good opening:

“X is [definition in one sentence]. The two reasons it matters in 2026 are [reason 1] and [reason 2]. The standard implementation looks like [3-bullet sketch].”

The second one extracts cleanly. The first one extracts into nothing.

2. Named authorship and a real bio

LLMs disproportionately cite content with a clear author. “by Niko Alho” beats “Posted by admin” by a wide margin. The author needs:

• A name in the byline and a <link rel="author"> or Person schema
• A bio block with credentials, named clients or employers, and a portrait photo
• A consistent author URL (/about or /team/[slug]) linked from the article

The signal you are sending is: there is a human behind this claim. LLMs cite humans more confidently than they cite domains. See E-E-A-T guidance for the broader story.

3. Inline citations inside the answer paragraph

A paragraph that opens with “According to [study from publisher 2025], X is true” is way more citable than the same paragraph with no source. The LLM is doing exactly what you do — looking for evidence under the claim — and your inline citation makes its job easy.

Three rules:

• Cite real, reachable URLs. Broken links hurt.
• Cite primary sources, not “as reported by another blog.”
• Cite at the sentence level, not in a bibliography at the bottom. Citations adjacent to the claim get pulled.

The crawl layer

Before any of this work matters, the search crawler has to fetch the page. Check three things.

robots.txt. Decide each policy intentionally. Allow OAI-SearchBot if you want OpenAI to surface the site in ChatGPT search. Decide separately whether GPTBot may crawl for potential training use. Other platforms publish their own user agents and controls.

WAF and CDN. A permissive robots.txt does not help if bot protection blocks the request. Verify access against the official user-agent and published IP guidance, then inspect server logs for successful fetches.

Rendered content. Serve the core answer in the initial HTML. If the article only appears after a client-side interaction, you are adding a failure mode that static HTML does not have. See hydration and Next.js SEO for the technical pattern.

Schema markup: what helps, what does not

Structured data should clarify what is already visible.

• Article schema can identify the author, dates, headline, and canonical page.
• Person schema can connect the author page and consistent identity facts.
• FAQPage schema is appropriate only when genuine questions and answers are visible.
• HowTo schema is appropriate only for a visible procedural page that meets the type’s requirements.

No supported schema type guarantees a ChatGPT citation. Organization, BreadcrumbList, and WebSite markup still help describe the site, but citation impact must be measured rather than assumed. See schema markup for the implementation guide.

Page length: a non-factor

Conventional SEO wisdom says longer pages win. LLM citation breaks that pattern.

A 600-word post by a named expert with one strong direct-answer paragraph gets cited more often than a 3,000-word generic listicle. The LLM is not weighting depth; it is weighting extractability and source credibility.

This does not mean go short. Aim for the length the topic deserves — usually 1,200 to 2,500 words. Just stop padding to hit a word count. Every section past “the answer” should add a real distinction, not warm air.

The measurement loop

You cannot improve what you cannot see. ChatGPT does not push citation data to Search Console. You need at least one of:

• Profound. GEO-native platform. Tracks share-of-voice across LLMs at the query level. Best onboarding. About $499/mo entry.
• Ahrefs Brand Radar. Mentions across ChatGPT, Perplexity, Gemini, Copilot. Bundled with Ahrefs Enterprise; add-on otherwise.
• DataForSEO LLM mentions API. Programmatic citation pulls. For teams building custom dashboards. Usage-based.
• Manual scrapes. Run your 30 to 100 target queries through ChatGPT each week. Log the citations. Tedious; works.

Whichever you pick, build a weekly review where you look at three things: which queries cite you, which queries cite competitors but not you, and which sentences are being lifted from your content. The third one is the most actionable signal in the loop.

The retrofit, week by week

A realistic 4-week sprint to push citation rate on an existing site.

Week 1. Audit. Identify the top 30 to 50 buying-intent queries. Run them through ChatGPT. Log which ones return citations, which ones cite competitors, which ones return sourceless answers. This is your baseline.

Week 2. Rewrite. Pick the 10 highest-value queries. For each, find or create the target page. Rewrite the first 200 words to follow the answer-first pattern. Add the H2 question. Add inline citations.

Week 3. Schema and authorship. Ship Article, Person, and FAQPage schema across the 10 pages. Add real bio blocks. Verify all schema in the Rich Results Test.

Week 4. Crawl and tracking. Open robots.txt to all LLM bots. Submit Bing sitemap. Set up Profound or a manual tracking sheet. Bookmark the queries.

Then wait. Citation rates lift starting week 4 to 6, with the bulk of movement happening between week 8 and week 12. Some queries never lift — that usually means the answer is fundamentally not differentiated, and the work is editorial, not structural.

Sentence-level iteration

Once you have a few citations, the highest-use work is sentence-level rewrites.

Example. A client’s page got cited for “what is X” by Perplexity but never by ChatGPT. The Perplexity-cited sentence was a definition with a precise number. The same page in ChatGPT was getting outcited by a competitor whose first sentence framed the answer as “X is the new Y” — a contrast frame rather than a definition.

We rewrote our opening sentence as a contrast frame, kept the precise number in the next sentence. Three weeks later, ChatGPT started citing the page.

The lesson: each LLM has a slightly different extraction style, and the cited sentence is the unit you optimize, not the page. See intent classification with AI for the broader pattern of querying LLMs about their own preferences.

What gets oversold

A few things that sound like GEO best practices but do not earn their cost.

Stuffing content with question-pattern H2s. Three good Q-and-A H2s outperform fifteen forced ones. The model penalizes thin extraction; if every H2 is “What is X?” with two sentences under it, the page reads as low-information.

Filing for inclusion in OpenAI’s training data. OpenAI does not have an opt-in submission process. The way in is the open web.

Writing content “for ChatGPT” without thinking about humans. A page that reads as if it was written for an LLM gets pruned by the LLM. The model is trying to cite content humans found useful; if your page reads as machine-bait, you go down the ranking.

What to do tomorrow

If you read nothing else, do this:

1. Pick your top 10 buying-intent queries. Open ChatGPT. Ask each. Log who gets cited.
2. For the queries where competitors are cited, open the competitor’s page. Look at the first 200 words. Compare to yours.
3. Rewrite your first 200 words. Ship today. Check back in 4 weeks.

There is no clever shortcut. The work is structural, repeatable, and rewards patience. The teams that start it in 2026 will own their categories in 2027. The ones that wait will be invisible in the layer of search that is replacing the SERP.