How to appear in ChatGPT: the B2B guide (2026)

Why being cited by ChatGPT changes the B2B game

ChatGPT claims several hundred million weekly active users, according to OpenAI’s public communications. Some of these questions are commercial or comparative: "best ERP for an industrial SME", "GEO agency Paris", "Stripe vs Adyen". When the answer names your brand, the prospect discovers you before even opening Google.

These mechanisms have been studied. The reference academic study, "GEO: Generative Engine Optimization" (Aggarwal et al., Princeton, Georgia Tech, Allen Institute for AI, IIT Delhi), presented at ACM SIGKDD 2024, tested several optimisations on 10,000 queries against generative engines. It measured up to +41% visibility in answers for content enriched with sourced statistics, and +28% with attributed quotations.

How ChatGPT actually picks its sources

ChatGPT draws from two pools. The first is its training corpus, a mass of text frozen at a date that varies by model. The second is live web search (ChatGPT Search), historically backed by Bing’s index, which it uses whenever a question calls for recent information or recommendations.

To get cited, a brand therefore needs to appear in the public texts that feed the models, be indexed by Bing as well as Google, and write its pages so that passages can be quoted without ambiguity. That’s the opposite of pre-2020 SEO, where you tried to occupy the results page. Here, you try to become a source the AI cites.

Step 1. Citability audit

Before any action, measure where you stand. The NEXUS GEO audit scores each brand against the 47 criteria of the public GEO-47 framework, grouped into 8 pillars, with a score out of 100. The framework is public; applying it (gathering evidence, weighting, sector trade-offs) is the agency’s work. Because it covers what makes a site citable, the audit grid is valid for all AIs (ChatGPT, Gemini, Perplexity, Claude, Mistral, Copilot and Google AI Overviews).

The deliverable is a 30-page PDF report with an overall score out of 100, a score per pillar, a visibility test in 5 AI engines (ChatGPT, Gemini, Perplexity, Mistral and Google AI Overviews), a benchmark against 3 named competitors and a 6-month action plan, presented in a 60-minute debrief. Until you have that baseline, any action is blind. Our methodology page presents the approach.

Step 2. Schema.org: the non-negotiable technical layer

Schema.org is the vocabulary machines read to understand your site without guessing. Without valid Schema.org Organization markup, ChatGPT has to guess who you are, and it can get it wrong, or worse, attribute your signals to a namesake. Google states that structured data helps it understand a page, without making it a condition for appearing in AI Overviews.

Priority types for B2B are Organization (identity), ProfessionalService or SoftwareApplication (offer), FAQPage (extractibility), Article (editorial content), Person (authors). Our Schema.org for AI guide details the JSON-LD implementations line by line.

Step 3. llms.txt: what it is and what it isn’t

llms.txt is a standard proposed by Answer.AI in September 2024 (llmstxt.org). It’s a Markdown file placed at the site root that summarises your offer and main pages for language models. Many tech sites have adopted it, but no major AI provider has announced official support for it, and its effect on citations has not been publicly demonstrated. You can see AI crawlers visiting the file in your server logs.

It is cheap to set up, so it can sit alongside the other signals; it does not replace them. Details in llms.txt: the complete guide.

Step 4. Citable content: 5 writing rules

1. One claim, one source. Every figure, every claim should be attributed to a dated source: a dated, attributed figure can be quoted, a slogan cannot.
2. Definitions in the first sentence. AI engines readily reuse sources that define a concept clearly and factually. Put the definition under an H2 and answer in fewer than 40 words right below.
3. High factual density. One number, date or proper noun every few sentences. The Princeton GEO study found that adding relevant, sourced statistics is among the optimisations that most improve visibility in generative answers.
4. Scannable structure. An H2 every 200-300 words, ordered lists, comparison tables: well-delimited blocks make extraction easier.
5. Explicit FAQ. A FAQ section with Schema.org FAQPage markup at the end of every commercial page, answering the exact questions your prospects ask assistants. ChatGPT often builds its answers from question-and-answer pairs.

Step 5. Domain authority: the slowest layer

Technical citability is a prerequisite; authority weighs on how often you appear. ChatGPT gives more weight to brands that others cite: sector media, independent comparisons, studies, specialist press. Public knowledge bases such as Wikipedia, Wikidata, Crunchbase or the SIRENE company registry for France are among the sources that help models recognise a company.

Three levers matter here. First, entity recognition: being present and unambiguous in the public knowledge bases AI engines lean on, so they can tell you apart from any namesake. These bases have their own admission rules, and Wikipedia penalises self-promotion: an entry created too early, or without secondary sources, can be deleted and hurt the brand. Second, regular third-party coverage: dated mentions in credible sector media build authority over time. Third, proprietary studies that the press picks up, one of the most durable ways to earn those mentions. Deciding which lever to pull first, in what order and with which supporting signals is the sequencing work of a GEO engagement.

4 illustrative scenarios

The four profiles below are cases built for illustration, not client records. They show how priorities change from one brand to another: they describe a method, not results.

Scenario 1. B2B SaaS publisher

Priorities: structured product pages and comparisons, a consolidated brand entity, and tracking of a panel of industry questions.

Scenario 2. Direct-to-consumer clothing brand

Priorities: product pages with complete Product and Offer data, and presence in the buying guides and comparisons AI engines consult for questions such as "ethical clothing brands made in Europe".

Scenario 3. Regional law firm

Priorities: factual pages for each practice area, consistent listings in professional directories, and strict compliance with the profession’s communication rules.

Scenario 4. Climate deeptech B2B startup

Target: citations in thematic summaries ("French startups decarbonising industry"). Priorities: proprietary studies aimed at the sector press, a consolidated entity and a complete llms.txt. This takes several months, the time for the press to pick up the studies and for engines to index them.

Step 6. Measure to steer

You can’t steer what you don’t measure. After the audit, you track a panel of representative industry questions over time, with three metrics: citation rate (share of questions where the brand is named), average position in the answer, and AI share of voice (the brand’s share of all brand citations). NEXUS GEO runs this tracking on ChatGPT, Gemini, Perplexity, Mistral and Google AI Overviews, and adjusts priorities at each cycle.

To go further, see our dedicated article on measuring AI citations, which compares free tools (Google Search Console, Bing Webmaster Tools), paid tools (Profound, AthenaHQ, Goodie AI) and DIY approaches.

FAQ

Sources

• Aggarwal P. et al. ("GEO: Generative Engine Optimization") Princeton, Georgia Tech, Allen Institute for AI, IIT Delhi, ACM SIGKDD 2024 (arxiv.org/abs/2311.09735).
• OpenAI: official communications on ChatGPT usage and web search.
• Google Search Central: documentation on structured data and AI features in Search, 2025.
• Semrush: study on the value of AI-referred traffic (an AI visitor converts 4.4× better), 2025 (semrush.com).
• Answer.AI: llms.txt standard proposal, September 2024 (llmstxt.org).
• GEO-47 framework: NEXUS GEO, 8 pillars and 47 criteria (/referentiel-geo-47).