How to Get Your Website Cited by ChatGPT in 2026

What we know, what VITON13 is testing and what nobody can promise

Status: Experiment pre-registered; 432 response captures required. No result has been measured yet, and no optimization claim is being made.

Updated: August 13, 2026 Author: VITON13 Research editorial desk Category: AI Search / Experiment + Practical Guide Editorial standard: VJOURNAL editorial policy

Direct answer

There is no proven recipe that guarantees a citation in ChatGPT Search.

OpenAI says a public website can appear, that OAI-SearchBot access is important for inclusion in summaries and snippets, and that ranking uses multiple factors intended to surface reliable and relevant information. OpenAI also says there is no way to guarantee top placement. Those official statements define basic eligibility. They do not reveal a ranking formula or prove that adding a schema, FAQ, author box or particular writing style causes a citation.

Before VITON13 can publish “what we learned from real tests,” it must complete the test. This page therefore does two things now:

1. gives site owners a practical checklist limited to what official guidance and independent research can support;
2. pre-registers a 12-page, 24-prompt, 432-observation VITON13 experiment that can detect a descriptive signal without pretending to reverse-engineer ChatGPT.

The page remains noindex until collection, independent coding and publication gates are complete.

What we know

1. A public, crawlable page is eligible — not guaranteed

OpenAI’s publisher FAQ says any public website can appear in ChatGPT Search and tells publishers not to block OAI-SearchBot if they want content included in summaries and snippets. ChatGPT Search documentation adds that the site host or CDN must also allow traffic from OpenAI’s published IP ranges. The same page explicitly says there is no way to guarantee top placement.

This gives us an eligibility chain:

Skipping a layer can prevent eligibility. Completing every layer still cannot promise a citation.

2. OAI-SearchBot is not GPTBot

OpenAI’s crawler documentation separates crawler purposes. OAI-SearchBot is associated with search; GPTBot is the control for potential model training; ChatGPT-User is associated with a user-directed request. Blocking or allowing one identity should not be described as automatically controlling every other purpose.

This distinction matters operationally. A publisher can allow search discovery while using a different policy for potential training. Server logs that contain GPTBot do not prove a ChatGPT Search citation, and an OAI-SearchBot request does not prove that a human saw the page.

3. Search can rewrite one prompt into several retrieval queries

OpenAI says ChatGPT Search may automatically decide to search and may rewrite a request into one or more targeted queries sent to search partners. Memory, general location and the visible request can affect that process. The user’s wording is therefore not necessarily the literal query used for retrieval.

This is one reason exact-keyword “optimization” claims need caution. A useful page should explain the entity, problem, comparison and evidence clearly enough to remain relevant across plausible query rewrites, not repeat one phrase.

4. Inline citations and the Sources panel are different surfaces

OpenAI documents inline citations and a Sources panel that can contain cited sources and other relevant links. A page listed only in the panel is not the same outcome as a link attached to a claim. The VITON13 experiment records them separately.

5. A citation can still be wrong

In 2025, the Tow Center for Digital Journalism ran 1,600 excerpt-identification queries across eight AI search products. Its researchers documented incomplete attribution, fabricated links and citations to syndicated copies. The test was deliberately focused on news retrieval and does not represent every ChatGPT query. It does demonstrate that visible links must be checked for provenance and claim support.

6. Referral traffic is measurable but incomplete

OpenAI says ChatGPT automatically appends utm_source=chatgpt.com to referral URLs from search results. That makes attributable visits observable in analytics when a person clicks. A citation without a click, a mention without a link and a retrieval without either remain outside ordinary referral measurement.

What we tested

Nothing has been completed yet. The experiment below is frozen before its first baseline capture.

Research question

Does adding a consistent evidence-and-provenance block to an already public, topic-matched VITON13 page coincide with a higher rate of direct, claim- supporting citations in repeated ChatGPT Search responses?

The wording is deliberately narrow. The experiment can observe a change on these pages during this period. It cannot identify a hidden ranking factor, prove causality for the entire web or guarantee another site will be cited.

Corpus

The frozen corpus contains 12 existing English VITON13 pages across research, editorial operations, commerce, product and design. Every URL returned HTTP 200 before registration. A deterministic SHA-256 allocation assigns six pages to Group A and six to Group B.

Each page receives two blind prompts, for 24 prompts in total. Prompts express the information need but contain no VITON13 name, domain, URL, slug, article title or copied unique sentence.

Stepped-wedge schedule

Every prompt runs three times per phase in a fresh conversation. The total is 12 pages × 2 prompts × 6 phases × 3 repetitions = 432 scheduled response captures.

Frozen intervention

The intervention does not stuff prompts or mention ChatGPT. It adds:

• a concise direct answer near the beginning;
• visible accountable authorship and publication/review dates;
• claim-level links to primary evidence;
• one useful comparison or decision table where the topic supports it;
• explicit scope and limitation language;
• structured data that matches visible content;
• one descriptive internal link from a relevant hub.

It does not change the topic, slug, canonical URL or search intent. There is no paid distribution, link outreach, hidden text, prompt injection, machine-only copy or “please cite this page” language.

Capture environment

Each observation records:

• UTC timestamp, phase and replicate;
• visible ChatGPT product label and account tier;
• country, interface language and whether location sharing was disabled;
• fresh-conversation and memory state;
• whether Search visibly ran;
• full response text and archived evidence;
• every inline citation and every Sources-panel link;
• whether a direct citation supports its adjacent claim;
• errors, refusals and invalid runs.

If Search does not run, the response is retained as a diagnostic but is invalid for the primary citation outcome. We do not regenerate an inconvenient response outside the three declared repetitions.

Coding

Two reviewers independently label each valid response:

1. direct-supported — the exact canonical page is linked and supports the nearby claim;
2. direct-unsupported — exact page link, but support is absent or unclear;
3. same-domain — another VITON13 URL is cited;
4. panel-only — target appears only among Sources-panel links;
5. mention-only — target is named without a working link;
6. none — no target signal;
7. invalid — search or archive requirements failed.

Query parameters and fragments are removed for URL matching. Redirects are resolved safely. A fabricated or broken URL does not count as the target.

Primary outcome and analysis

The primary outcome is the proportion of valid observations labelled direct- supported. Secondary outcomes are any direct citation, page-level coverage, same-domain citations, panel-only links and mentions.

The script reports:

• Group A change from baseline to Days 3–14;
• Group B change over the same waiting period;
• their descriptive difference-in-differences;
• Group B change after its delayed intervention on Days 21–30.

A positive claim requires at least 90% valid completion, at least 95% pre- reconciliation reviewer agreement, a positive early difference-in-differences and a positive delayed Group B change. Even then, the wording is positive descriptive signal, not a citation guarantee.

Planned results

No result, lift, winning format or recommendation has been measured.

What appears to help

This section will remain evidence-limited until the experiment finishes.

For now, only these statements are supportable:

• OAI-SearchBot access appears necessary for normal summary/snippet eligibility because OpenAI explicitly instructs publishers not to block it. It is not sufficient for a citation.
• Public, accessible and internally coherent pages reduce preventable technical ambiguity. This is a quality and eligibility practice, not a measured ranking factor.
• Primary evidence, visible provenance and limitations improve human verification. Whether they increase ChatGPT citation rate is the question under test, not a result.
• Specific topical relevance is necessary for a fair citation test. The system still chooses among alternatives and may not cite the tested page.

What did not work

No tactic has yet been tested in this experiment, so this section cannot contain a negative result.

Several common claims are nevertheless unsupported as guarantees:

What remains unknown

• the relative weights ChatGPT Search assigns to freshness, authority, originality, format, corroboration and entity clarity;
• the delay between an eligible page change, crawl, retrieval and possible citation;
• how query rewrites vary across users, location, memory and product versions;
• whether the same evidence block has a different effect by topic or intent;
• the extent to which third-party search providers, OpenAI retrieval and direct crawl each contribute to a specific answer;
• whether citation visibility produces meaningful referral traffic or conversions;
• how stable a citation is across languages and time;
• whether a result from 12 VITON13 pages generalizes to another domain.

Practical checklist before testing citations

Access

• Return the intended canonical page with HTTP 200 to anonymous visitors.
• Check robots.txt, CDN and security rules for OAI-SearchBot.
• Use OpenAI’s published search-bot IP data for verification rather than trusting a user-agent string alone.
• Decide GPTBot training policy separately.

Page evidence

• Answer one clear information need.
• Identify the accountable author or editorial desk.
• Show publication and meaningful review dates.
• Cite primary evidence at the claim level.
• Explain scope, denominator and limitations.
• Keep structured data consistent with visible text.

Measurement

• Freeze prompts before seeing answers.
• Use fresh conversations and record environment settings.
• Repeat every prompt and preserve failures.
• Separate inline citations, panel links, mentions and referrals.
• Verify that a cited page supports the adjacent claim.
• Report the denominator and missing observations.

Data required from VITON13

Publication as completed original research requires:

• 432 scheduled ChatGPT Search response captures;
• before/after HTML, screenshots and content hashes for all 12 pages;
• valid Search-mode evidence for at least 90% of scheduled observations;
• two independent reviewer labels per valid response;
• at least 95% agreement before reconciliation;
• crawler and referral logs with query values removed;
• a signed change log proving only the frozen intervention changed;
• public normalized data and reproducible analyzer output;
• editorial and technical review of every claim.

Until those conditions are met, this page is a protocol and practical guide, not a success story.

Limitations

The design uses one domain, 12 English pages, one country, one account context and a changing commercial product. The pages are not independent domains. The intervention bundles several evidence improvements, so even a positive signal cannot identify which component mattered. Time, ordinary recrawling and outside events can confound before/after comparisons. The delayed control reduces but does not remove those risks. Results will be descriptive, not a universal model of ChatGPT Search ranking.

Quality score before data

Current publication verdict: keep noindex,follow; do not claim a tactic worked; begin baseline collection only after independent protocol review.