How to Get Cited by ChatGPT (and Stay Cited)

Disclosure: I publish Irvale Studio. We sell AI search visibility work to UK SMBs and global brands through our AI Visibility pillar and our Revenue Engineering engagement. Vendor behaviour and citation overlap claims below were verified against the cited primary sources on the date of publication.

What ChatGPT Search actually is in 2026

The architecture matters because the optimisation work follows from it. ChatGPT Search is not a wrapper around the OpenAI training corpus. It is a retrieval system that fetches live web pages and reranks them with a generative model on top.

Three components are worth naming clearly.

• OAI-SearchBot. OpenAI's retrieval crawler, indexing the public web for the OpenAI search index. Documented in OpenAI's developer docs (OpenAI, 2024).
• ChatGPT-User. The user initiated agent that fetches live pages when a user pastes a URL or the model needs a fresh read. Read by ChatGPT itself during a session.
• GPTBot. The training crawler. Different beast. Used to ingest content into model training data.

ChatGPT Search defaults to OpenAI's own index, which is built on top of Bing's underlying web index per the November 2024 announcement and Microsoft Bing engineering posts. When the index lacks a current answer, it falls back to Browse with Bing for live retrieval (Bing Webmaster blog, 2024). The practical implication for UK SMBs is that Bing visibility now sits upstream of ChatGPT visibility, which makes Bing Webmaster Tools setup a precondition rather than a nice to have.

How does ChatGPT Search choose sources?

The retrieval and rerank pipeline runs roughly as follows, based on OpenAI's documentation, third party reverse engineering work and Profound's citation analysis through 2026.

1. Query intent classification. Search style queries route to retrieval; pure conversational queries skip it.
2. Candidate retrieval. Top results from OAI-SearchBot's index, expanded with live browse where needed.
3. Passage scoring. Candidate passages scored on relevance, freshness, schema verified entity alignment, and corroboration with other retrieved sources.
4. Reranking and selection. Two to six sources picked for the final answer.
5. Citation rendering. Source cards rendered inline or in the right rail depending on UI surface.

Profound's analysis through April 2026 found that ChatGPT Search skews toward encyclopaedic and professionally written sources. Wikipedia leaning extraction sat at around forty eight per cent of citations across general queries, with a long tail of news, trade press and well structured SMB content (Profound, 2026).

The pattern that gets cited is what we call EEAT signals plus passage extractability plus freshness. None of those three are optional.

ChatGPT, Perplexity and AI Overviews — citation behaviour compared

Source citation behaviour — ChatGPT, Perplexity, Google AI Overviews — 2026

The takeaway is that there is no single AI search optimisation. There is a stack. The schema and EEAT work is shared. The corroboration and freshness work needs different distribution channels per engine.

Bing Webmaster Tools setup — the precondition

The three concrete steps, in order.

1. Verify the site in Bing Webmaster Tools

Go to bing.com/webmasters, sign in with a Microsoft account, and add the site. Verification methods are the same as Google Search Console: meta tag, DNS TXT record, or XML file. The meta tag is fastest for a Next.js site. Drop it in your root layout once, deploy, click verify in Bing, done.

Once verified, import the sitemap. Bing reads /sitemap.xml by default but does not auto discover unless told. Submit it explicitly through the Sitemaps panel.

2. Set up IndexNow

IndexNow is a protocol that lets you push URL changes to Bing, Yandex, Naver and Seznam within minutes. According to Bing's own engineering blog, around twenty two per cent of Bing clicked URLs in 2025 came through IndexNow rather than the slower crawl path (Bing Webmaster blog, 2025).

The mechanical setup:

1. Generate an IndexNow API key inside Bing Webmaster Tools.
2. Host the key file at https://yourdomain.com/<key>.txt containing the key.
3. POST changed URLs to https://api.indexnow.org/IndexNow on every successful build that changes a route's lastmod.

For a Next.js site, this is one post build script that diffs the previous sitemap against the new one and submits only changed URLs. The boost is measurable inside a week.

3. Confirm the index includes your priority pages

Use the site:yourdomain.com operator inside Bing once a week for the first month. Pages missing from the index after IndexNow submission should be checked in the URL Inspection tool. The most common cause of missing indexing is unintentional noindex headers or incorrect canonicals carried over from a previous CMS.

Robots.txt — the allow list that decides citation eligibility

Reference robots.txt block for ChatGPT eligibility, plus the wider AI search ecosystem.

# Classical search
User-agent: Googlebot
Allow: /
User-agent: Bingbot
Allow: /

# OpenAI / ChatGPT — retrieval
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /

# OpenAI / ChatGPT — training
User-agent: GPTBot
Allow: /

# Perplexity
User-agent: PerplexityBot
Allow: /
User-agent: Perplexity-User
Allow: /

# Anthropic / Claude
User-agent: Claude-SearchBot
Allow: /
User-agent: Claude-User
Allow: /
User-agent: ClaudeBot
Allow: /

# Google AI training opt out
User-agent: Google-Extended
Allow: /

# Apple Intelligence
User-agent: Applebot-Extended
Allow: /

# Hard blocks for known bad actors
User-agent: Bytespider
Disallow: /

User-agent: *
Allow: /
Disallow: /api/
Disallow: /_next/

Sitemap: https://yourdomain.com/sitemap.xml

The decision matrix on GPTBot is worth treating honestly. GPTBot only affects training data, not retrieval. Blocking GPTBot does not prevent you from being cited inside ChatGPT today. It does prevent your content from being absorbed into the next model checkpoint, which weakens your brand recall in zero click answers where the model is responding from memory rather than retrieval.

For most UK SMBs, the cost benefit favours allow. Your blog content is already public. The benefit of being inside the training corpus is a stronger model prior on your brand. The risk of allowing is content that you have already published publicly being seen by the model. For a regulated client or a brand with sensitive proprietary data behind a paywall, the answer flips.

EEAT signals ChatGPT actually consumes

EEAT in 2026 is a specific set of structured and unstructured signals, not a vibe. Qwairy's 2026 study found content with verifiable author credentials earning about forty per cent more AI citations than identical content without (Qwairy, 2026). The signals worth shipping, in order of leverage divided by cost.

sameAs on Organization

The single most load bearing entity property. Inside the Organization JSON-LD at site level, list every authoritative identity for the brand:

• LinkedIn company page
• Companies House (for UK)
• Wikidata entry, even a stub
• Crunchbase profile
• X / Twitter handle
• GitHub organisation, if relevant
• Trustpilot / G2 page, if relevant

The model uses this list to resolve "is this Irvale Studio the same Irvale Studio that has a LinkedIn page" and to corroborate factual claims across them.

Author Person + ProfilePage

For every article, ship a real /about/[author] page with a ProfilePage wrapping a complete Person node. Include name, givenName, familyName, jobTitle, worksFor linked by @id to the Organization, sameAs to LinkedIn and similar, knowsAbout array of topics, and hasCredential if applicable.

Reference the author from each Article by @id, not by inline name. The model resolves the author once and carries the EEAT signal across every piece they have published.

Third party mentions

Unlinked brand mentions in 2026 correlate roughly three times more strongly with AI visibility than backlinks, based on Profound's 2026 analysis (Profound, 2026). The mechanism is corroboration. The model trusts a brand that appears across multiple independent domains saying roughly the same thing.

The work is unglamorous. Trade press placements, podcast guest spots, structured directory entries, Reddit thread participation, LinkedIn newsletter syndication. The cumulative effect over six months is the difference between being a citation candidate and being a default citation.

Branded query strategy

A non obvious lever: own the branded query. ChatGPT increasingly answers branded queries by retrieval rather than memory. If a user asks "what is Irvale Studio", you want the top three retrieval candidates to be your own owned channels. The site's About page, the LinkedIn company page, and the Crunchbase entry. Ship all three with consistent copy. Misalignment between the three is one of the easiest sources of hallucination on branded queries.

Sourced numbers worth knowing

The compounding effect of those three numbers is the case for prioritising AI search visibility now. The audience size is real. The plumbing for inclusion is already standardised. The structural work that lifts citation rates is durable across model changes.

Citing your own sources to be cited by them

The mechanic is mundane and works. Every numerical claim gets a (Source name, Year) inline citation. Every section opens with a definitional sentence. Every important answer is wrapped in <Speakable>. The model learns from extraction patterns: pages that look like clean primary sources are cited as clean primary sources. Pages that look like marketing copy are skipped.

This is also the discipline that makes content survive across model changes. Schema is volatile. Heading structure is volatile. The fundamental shape of an extractable, attributable answer block is durable.

Monitoring your share of voice — what to actually measure

The measurement stack we recommend for UK SMBs running an AI visibility programme:

• Prompt universe. A named list of one hundred to three hundred prompts that real buyers run, mapped to discovery, comparison and purchase intent.
• Engine coverage. ChatGPT, Perplexity, Gemini AI Mode, Google AI Overviews, Claude, Copilot, Meta AI. Manual sampling for any not yet supported by your tooling vendor.
• Metric set. Citation rate (per cent of prompts where you are cited), share of voice (your citation count divided by total citations on the prompt set), position (which slot you appear in), sentiment (factual versus negative versus positive), hallucination patrol (any descriptor that is wrong).
• Cadence. Weekly run. Monthly review. Quarterly model recalibration.

Profound, Athena Intelligence and similar tools automate the run. They do not automate the editorial response. The response is where the engineering work happens.

What we don't know yet — the open questions

The honest list, as of May 2026.

• Whether OpenAI applies a quality classifier in addition to Bing's. OAI-SearchBot is documented as a separate crawler. Whether it filters Bing's index or accepts it wholesale is not public.
• The exact freshness window for ChatGPT citations. Profound's analysis suggests fresh content (under thirty days) is preferred but not exclusively. The decay curve is not measured.
• How ChatGPT handles paywalled content. It can cite paywalled content. It is unclear whether it does so via OAI-SearchBot index entries, via partnerships with publishers, or via training data leakage.
• Whether llms.txt will become a real retrieval signal. OpenAI has not committed publicly. Anthropic has not committed publicly. Both ship one. Neither documents using it. Treat as zero benefit until proven otherwise.
• The effect of GPTBot blocking on retrieval citation rates. Logically blocking training should not affect retrieval. Anecdotally some practitioners report citation drops. The mechanism is not understood and the data is thin.

The pragmatic stance is the same as for Google AI Overviews: invest in durable structural work, treat engine specific tactics as quarterly review, and stay honest about what is measurement versus speculation.

What to ship this week — the seven item checklist

The ordered list, by leverage divided by cost.

1. Verify Bing Webmaster Tools and submit an accurate sitemap. This is the precondition.
2. Generate an IndexNow API key, host the key file at root, and wire a post build script to push changed URLs.
3. Audit robots.txt. Allow OAI-SearchBot, ChatGPT-User, GPTBot, PerplexityBot, ClaudeBot at minimum. Block Bytespider. Document the policy.
4. Ship Article or BlogPosting schema with a full Person author node on every blog and guide. Reference the author by @id.
5. Build the /about/[author] page with ProfilePage wrapping Person. Include sameAs, knowsAbout, credentials.
6. Audit Organization schema for accurate sameAs cluster covering LinkedIn, Wikidata, Companies House, Crunchbase, X.
7. Set up share of voice tracking on a defined prompt universe of fifty to three hundred prompts, run weekly across at least ChatGPT, Perplexity and Google AI Overviews.

If you would rather have this engineered for you across every engine that matters, that is what our AI Visibility pillar covers. The diagnostic, the schema work, the citation engineering and the weekly share of voice monitoring run inside one named programme.

The sister posts in this cluster cover the surfaces you optimise for next: Ranking in Google AI Overviews for the Google side, and llms.txt: The New robots.txt for AI for the file format that may matter when models start consuming it. For the classical SEO foundation that underwrites all of it, our Google Maps SEO guide and Google Business Profile setup walkthrough are the starting points for UK SMBs.

Common questions

Next stepSee how Irvale engineers AI-search visibility→Diagnostic, schema, citation engineering and weekly share-of-voice monitoring across every engine that matters