Category: Artificial Intelligence

An AI citation refers to the fact that a generative artificial intelligence model mentions and references a specific source in its response to a user. It is the equivalent of the blue link in Google, but in a conversational universe where a brand only appears if the model considers it trustworthy according to its own criteria. And those criteria vary radically from one model to another.

According to a study published by Yext on March 18, 2026, based on the analysis of 17.2 million AI citations generated by ChatGPT, Gemini, Claude, and Perplexity, each platform follows a distinct informational architecture. The direct consequence for SMBs: being well positioned in one model’s responses guarantees no visibility in the others.

Why AI search is not a single channel

For years, businesses optimized their digital presence for one central algorithm: Google’s. The logic was unified, the rules relatively stable, and efforts could be concentrated on a single lever. That time is over.

Today, a growing share of information queries flows through conversational agents. These agents do not operate according to PageRank logic or an advertising auction system. Each applies its own sourcing policies, often opaque, to decide which sources deserve to be cited in a response.

The Yext study (https://www.yext.com/research/ai-citation-behavior-across-models) confirms this with a striking figure: approximately 89% of domains cited by these models are specific to a single platform. Only 11% of domains appear across multiple AI engines simultaneously. In other words, the vast majority of sources cited by Gemini do not appear in Claude, and vice versa.

This fragmentation imposes a new reality: AI visibility is not a single channel to optimize globally. It is a set of distinct channels, each with its own rules.

What each model prioritizes

Gemini: the primacy of structured and controlled sources

Gemini is the model that displays the strongest preference for proprietary and verifiable content. According to the Yext study, approximately 93% of its citations come from sources directly controlled by brands: official websites and business or directory listings. News media and forums account for only 3% of its cited sources.

This logic aligns closely with the E-E-A-T signals valued by Google Search (Experience, Expertise, Authority, Trustworthiness). For an SMB, this means that Gemini rewards above all the consistency of structured data, semantic markup, and a well-maintained presence on reference directories. An incomplete Google Business profile or a poorly structured website mechanically reduces the chances of appearing in its responses.

Claude: online reputation carries more weight

Claude stands out for a significantly greater integration of reviews and user-generated content. The Yext study indicates that 15% of its citations come from consumer reviews, two to four times more than the other models (Gemini and ChatGPT each cite approximately 3% reviews, Perplexity approximately 7%).

The majority of its citations remain anchored in brand-owned sites and listings (81%), but customer sentiment exerts a perceptible influence on how Claude presents a business. For an SMB, this means that active online reputation management, the volume and quality of online reviews, and discussions on independent platforms become direct levers of AI visibility.

Perplexity: the most stable behavior

Perplexity displays the most consistent sourcing behavior across sectors. Brand sites account for between 37% and 50% of its citations depending on the industry. This relative stability makes it a more predictable model to optimize, even if the share of proprietary sources is less dominant than with Gemini.

Perplexity tends to diversify its sources more than its competitors, which means that a broader content strategy, including industry publications and specialized media, can help gain visibility there.

ChatGPT: the highest sectoral variability

OpenAI presents the highest sectoral variability of all the models analyzed. According to the Yext blog (https://www.yext.com/blog/how-chatgpt-perplexity-gemini-claude-decide-what-to-cite), the share of brand sites in its citations ranges from 28% in the food sector to 44% for nonprofit organizations. In hospitality, ChatGPT cites brand sites at a rate of 38%, compared to only 17% for Perplexity in that same sector.

This variability implies that an SMB cannot apply a uniform strategy for ChatGPT: the sourcing logic depends heavily on its industry. A prior analysis by vertical is essential before investing in optimization.

What this concretely changes for an SMB

Faced with this fragmentation, four priority actions emerge for building an effective and lasting AI visibility strategy.

1. Structure and unify first-party data

The underlying trend is clear: between the two Yext studies (October 2025 and March 2026), the share of brand-controlled sources in AI citations rose from 86% to 90%. Proprietary digital assets are taking an increasing place in AI responses. This means that a well-structured website, with careful semantic markup (schema.org, structured data), self-contained pages, and direct answers to frequently asked questions, forms the indispensable foundation of any GEO (Generative Engine Optimization) strategy.

2. Maintain business listings and directory presence

Gemini, which represents a considerable volume of traffic through its integration into Google, draws heavily from business listings and third-party directories. An SMB must ensure that its information (name, address, phone number, hours, description) is accurate, consistent, and up to date on Google Business Profile, industry directories, and reference platforms. The slightest inconsistency between these sources can undermine the trust the model places in them.

3. Actively manage online reputation and reviews

Claude gives significantly more weight to consumer reviews than the other models. At the same time, the overall share of citations from reviews declined from 8% to 5.5% between the two Yext studies, which does not mean reviews have become negligible, but that they need to be high quality rather than simply numerous. Responding to reviews, encouraging authentic feedback, and monitoring reputation on independent platforms are actions that directly influence visibility in Claude.

4. Segment strategy by target model

Since 89% of cited domains are specific to a single platform, applying a single strategy across all models is counterproductive. An SMB will benefit from identifying which models are most used by its target audience (based on sector, usage patterns, and tools deployed by its prospects), then adapting its efforts accordingly. A player in the tourism industry would do well to work differently for ChatGPT and for Perplexity, whose behaviors diverge significantly in that sector.

• Audit current presence in each AI model’s responses.
• Identify gaps between visibility on Gemini, Claude, Perplexity, and ChatGPT.
• Prioritize actions based on the models most used by target audiences.
• Set up regular tracking of AI citations, model by model.

An essential limitation to keep in mind

The data presented in this article represents a snapshot of model behavior in the fourth quarter of 2025. Yext itself warns that these patterns can evolve rapidly, in line with updates to the models and their underlying retrieval systems. Coverage of this study by the independent press (https://cmotech.uk/story/study-finds-ai-models-develop-distinct-citation-habits) also highlights this evolving nature.

What is true today for Gemini or Claude may change in six months. Models are updated frequently, their sourcing policies evolve, and new players can shift the balance. This does not make optimization pointless, quite the contrary: it requires treating it as an ongoing discipline, not a one-time project.

The right approach is to build solid foundations (structured data, consistent listings, controlled online reputation) that remain relevant regardless of how algorithms evolve, while maintaining active monitoring of changes in model behavior.

Conclusion

Visibility in AI responses is not a single channel: it is an archipelago of platforms with distinct sourcing logics, which only a differentiated strategy can address effectively. Gemini rewards structure and authority, Claude integrates online reputation, Perplexity offers relative stability, and ChatGPT varies by sector. For an SMB, ignoring these differences amounts to optimizing for a single search engine in a world where four exist.

If you would like to audit your current visibility in AI responses and build a GEO strategy tailored to your sector, our teams can support you in this process, model by model.

Background: Anthropic and Its Ambitions in Generative AI

Anthropic is one of the most influential companies in the field of generative artificial intelligence. Founded in 2021 by former members of OpenAI, including Dario and Daniela Amodei, the company has positioned itself as an actor committed to AI safety, particularly through its Claude assistant. However, like all major American technology companies operating in such a sensitive sector, Anthropic is not immune to the regulatory and political pressures that are intensifying in the United States.

It is important to clarify a crucial point from the outset: as of the date this article was written, there is no official and verifiable confirmation that the American government explicitly compelled Anthropic to remove products bearing the names ‘Fable 5’ and ‘Mythos’ from its catalog. These names do not correspond to any publicly documented products in Anthropic’s official communications. It is therefore possible that these designations refer to internal projects, undisclosed code names, or information circulating in specialized circles without public confirmation. This article therefore offers an analysis of the general regulatory context surrounding this type of decision, while remaining factual about what is actually known.

The American Regulatory Context Around AI

Since 2023, the United States has considerably strengthened its oversight of the artificial intelligence sector. The Biden administration signed a landmark executive order on AI safety in October 2023, requiring developers of powerful systems to share their safety test results with the federal government before any public deployment. This measure targeted in particular so-called dual-use models, meaning those that could potentially be used for military purposes, disinformation, or cyberattacks.

Under the Trump administration, which returned to power in January 2025, American AI policy took a new direction. While some regulatory constraints were relaxed in the name of economic competitiveness against China, other control mechanisms were maintained or strengthened, particularly those related to national security and the protection of critical infrastructure.

• The Bureau of Industry and Security (BIS) monitors exports of sensitive AI technologies.
• The NSA and the Department of Defense have oversight authority over certain models with advanced capabilities.
• The National Institute of Standards and Technology (NIST) establishes reference frameworks for evaluating the risks of AI systems.
• Regular congressional hearings put the leaders of major AI companies under pressure.

Why Certain AI Projects Can Be Withdrawn Under Government Pressure

Risks Related to Advanced Content Generation Capabilities

American authorities are particularly concerned about models capable of generating highly persuasive content, convincingly simulating characters, or producing complex narratives that could be used for large-scale manipulation. A system designed to create immersive narrative universes, as a name like ‘Fable’ or ‘Mythos’ might suggest, would naturally fall into this sensitive category.

American intelligence agencies have repeatedly expressed concerns about AI tools capable of generating textual deepfakes, automated disinformation campaigns, or content designed to influence elections. These concerns have led several companies to modify or suspend projects even before their public announcement.

Confidential Agreements Between Companies and Federal Agencies

A well-established practice in the American technology sector involves what researchers call ‘soft requests,’ meaning informal requests from government agencies that do not take the form of laws or official orders, but that nonetheless exert real pressure on companies. These requests may concern the withdrawal of a product, the modification of its capabilities, or the restriction of its access to certain populations or regions.

Anthropic, as a company that has received significant investment from actors close to the American government, would be particularly exposed to this type of influence. The company has, for example, benefited from investments by Amazon, whose ties to American government contracts are well documented.

What Is Known About Recent Developments at Anthropic

In 2025, Anthropic continued to deploy and refine its range of Claude models, with the release of Claude 3.5 and the first announcements around Claude 4. The company also launched autonomous agent features capable of executing complex tasks on a computer, which the company calls ‘computer use.’ These developments have sparked renewed interest from regulators.

• Anthropic has strengthened its team dedicated to public policy and government relations in Washington D.C.
• The company has published several reports on the safety of its models, as part of a proactive transparency approach.
• Discussions are underway in the U.S. Congress to regulate autonomous AI agents, which represent a new regulatory frontier.
• Anthropic has signed voluntary commitments with the White House regarding safety testing of its most powerful models.

The Stakes of Creative and Narrative Projects in AI

If projects bearing evocative names like ‘Fable’ or ‘Mythos’ did indeed exist internally at Anthropic, they were likely part of the broader experimentation around narrative and creative AI. This field is under particular scrutiny, as it touches on intellectual property, the generation of potentially harmful content, and the ability of systems to create immersive experiences that are difficult to distinguish from reality.

Several companies in the sector have already had to abandon similar projects. Meta, for example, suspended certain developments related to the creation of persistent virtual characters following criticism about the psychological risks for users. Google DeepMind also revised the scope of certain narrative projects after consultations with safety experts.

The Tension Between Innovation and Safety

The AI sector today is defined by a fundamental tension between the desire to innovate rapidly in order to maintain a competitive advantage over rivals like China, and the need to ensure that deployed systems do not pose a risk to society. This tension is particularly acute for a company like Anthropic, whose stated mission is precisely to develop AI that is safe and beneficial for humanity.

When an internal project comes into conflict with safety guidelines, whether imposed externally by the government or defined internally by research teams, withdrawing or suspending the project may become the only reasonable option. This type of decision is rarely announced publicly, which explains why projects can disappear without any official explanation.

The Implications for the AI Industry as a Whole

The matter, whether or not its precise details are confirmed, illustrates a broader reality: governments, and in particular that of the United States, are exerting growing influence over the development of artificial intelligence. This influence does not always take the form of formal laws, but is also expressed through informal pressure, conditions of access to public markets, or threats of future regulation.

• AI companies are increasingly required to integrate regulatory compliance teams from the earliest stages of development.
• The concept of ‘safety by design’ is becoming an essential criterion for obtaining institutional funding.
• Projects with high creative or narrative potential are receiving particular attention from regulators.
• Transparency about the actual capabilities of models has become an implicit requirement for maintaining the trust of government partners.

Conclusion: Between Opacity and Responsibility

The story of Anthropic and the projects supposedly withdrawn under American pressure highlights the fundamental challenge that governing artificial intelligence represents in the contemporary era. In a sector where technological advances often outpace legislators’ ability to regulate them, decisions are frequently made in the shadows, far from public debate.

For observers and citizens alike, it is essential to remain vigilant, to follow official announcements from companies and regulators, and to treat with caution unconfirmed information circulating about undisclosed projects. What is certain, however, is that the question of regulating generative AI is now at the heart of American political priorities, and that this dynamic will continue to profoundly shape the global technological landscape in the years to come.

A Government Directive That Arrived Without Warning

On June 12, 2026, Anthropic received an official directive from the American government ordering it to immediately suspend access to two of its most advanced artificial intelligence models: Fable 5 and Mythos 5. The notification arrived at 5:21 PM (Eastern Time), leaving the company an extremely short window to comply with the order.

The directive, issued under national security and export control authorities, imposes a total suspension of access to these two models for all foreign nationals, whether located inside or outside American territory. This restriction also applies to Anthropic employees who are not American citizens. In practice, to ensure compliance, Anthropic was forced to disable Fable 5 and Mythos 5 for all of its users, regardless of nationality.

The other models offered by Anthropic, including Opus, Sonnet, and Haiku, are not affected by this measure and remain normally accessible.

The Official Reason: A Potential Bypass of Safety Guardrails

According to information communicated by Anthropic, the American government justifies its decision by the discovery of a method for bypassing the protections built into Fable 5, a technique commonly known as a ‘jailbreak’ in the AI field. The letter received by Anthropic does not, however, specify the technical details of the identified threat.

Based on Anthropic’s understanding of the matter, the government reportedly became aware of a specific technique for forcing the model to identify a small number of previously known software vulnerabilities considered minor. In concrete terms, this would involve a method of asking the model to analyze a particular source code and correct its flaws.

Anthropic’s Response: Deep Disagreement, but Full Compliance

Anthropic published a detailed statement in which the company expresses both its compliance with the legal directive and its profound disagreement with the basis of this decision. The company believes that the conclusions presented by the government do not justify such a radical withdrawal of a model deployed to hundreds of millions of users.

Months of Preventive Work Highlighted

Anthropic points out that before the launch of Fable 5, considerable work had been carried out to evaluate and strengthen the model’s protections. The company highlights the following points in particular:

• Thousands of hours of stress testing (‘red-teaming’) were conducted in collaboration with the American government, the UK AI Safety Institute (AISI), several private third-party organizations, and internal teams.
• These tests demonstrated that Fable 5’s protections are significantly more effective than those of any previously deployed model.
• No tester succeeded in finding a ‘universal jailbreak,’ meaning a method capable of globally bypassing all of the model’s protections.
• The guardrails put in place are so strict that many users have complained about their overly restrictive nature.

A Defense-in-Depth Strategy Claimed

Anthropic defends an approach it describes as ‘defense in depth.’ Rather than aiming for perfect resistance to jailbreaks, which it considers impossible to achieve for any model provider, the company sought to make such attempts either very limited in scope or extremely costly to carry out. This strategy is complemented by an intensive monitoring system designed to quickly detect and neutralize any successful attack.

It is within this logic that Anthropic imposed a 30-day customer data retention policy for Fable users, a decision that, according to the company, generates real costs and commercial friction, but allows it to analyze and correct bypass attempts.

Anthropic’s Technical Arguments Against the Decision

On the substance of the matter, Anthropic contests the scope and the actual danger of the jailbreak presented to the government. The company puts forward several technical arguments to support its position.

• The vulnerabilities identified through the bypass technique are described as ‘relatively simple’ and already known.
• Other publicly available models, without requiring a jailbreak, are capable of discovering these same vulnerabilities.
• Anthropic has examined the report that would form the basis of the government directive and concludes that the level of capability demonstrated is widely available through other models on the market, including OpenAI’s GPT-5.5.
• These capabilities are used daily by cybersecurity professionals responsible for protecting computer systems.
• At no point has the government presented evidence of a concretely harmful outcome linked to a Fable 5 jailbreak.
2026-06-18

In a world where artificial intelligence is transforming every aspect of our professional lives, Anthropic continues to innovate with an interface dedicated to creativity and design: Claude AI Design. Whether you are a designer, developer, writer, or project manager, this platform is built to support you at every stage of your creative work. In this article, we present everything you need to know about this tool, its practical uses, its recent additions, and our suggestions for structuring your projects effectively.

What is Claude AI Design?

Claude AI Design is a specialized interface developed by Anthropic, the company behind the Claude language model. Accessible via the URL claude.ai/design, this platform stands apart from classic chatbot interfaces by offering an environment designed for creative and visual tasks. It builds on Claude’s advanced capabilities in natural language understanding, structured content generation, and contextual analysis, all within a design-oriented framework.

The main goal of Claude AI Design is to provide a coherent workspace where users can generate ideas, refine concepts, write creative briefs, obtain layout suggestions, and collaborate with the AI smoothly. Unlike a simple text assistant, this interface integrates features tailored to the needs of creative professionals.

What is Claude AI Design actually used for?

The uses of Claude AI Design are numerous and cover a wide range of professional and personal needs. Here are the main use cases that make this tool an indispensable ally for creative professionals.

Generating and structuring creative ideas

One of the major strengths of Claude AI Design lies in its ability to generate relevant and original ideas from a simple description. Whether you are looking for a concept for an advertising campaign, a brand name, a color palette, or a website architecture, the tool analyzes your request and proposes concrete, well-argued, and immediately usable directions.

Writing briefs and specifications

Writing a creative brief is often a time-consuming task. Claude AI Design helps you structure your reference documents in just a few minutes. It can generate complete briefs for graphic projects, specifications for websites, or functional requirements for mobile applications. The result is clear, professional, and customizable to your needs.

UX and UI design assistance

For interface designers, the tool offers valuable support. It can suggest navigation structures, propose visual hierarchies, describe interface components suited to a given context, and even explain the design principles that apply to your project. It is a true thinking partner for user experience professionals.

Analyzing and improving existing content

You can submit a text, a project description, or an existing concept, and Claude AI Design will provide a detailed analysis with improvement suggestions. This feature is particularly useful for refining a pitch, improving the clarity of a message, or strengthening the visual consistency of a project.

• Generation of creative ideas and original concepts
• Automated writing of briefs and specifications
• Support for UX, UI, and information architecture design
• Critical analysis and improvement of existing content
• Creation of user personas and usage scenarios
• Suggestions for typography, color palettes, and visual styles
• Help writing pitches and creative presentations

Recent additions that make a difference

Anthropic regularly updates its tools to meet the growing expectations of users. Claude AI Design is no exception and benefits from continuous improvements that enrich the overall experience.

A more intuitive and streamlined interface

The latest updates have brought a partial redesign of the interface, with a cleaner look that makes navigation easier. Input areas have been optimized, responses are better formatted, and the readability of exchanges has been noticeably improved. The goal is to reduce friction between the user and the tool in order to encourage a more natural workflow.

Better long-context management

One of the most significant improvements concerns the handling of long conversations. Claude AI Design can now maintain an extended context across many exchanges, meaning it remembers the details of your project throughout the session. This capability is essential for complex projects that require consistency across multiple design stages.

Enhanced multimodal capabilities

With the evolution of Claude models, the platform now integrates more advanced image analysis capabilities. You can submit screenshots, mockups, or existing visuals, and the tool analyzes them to provide detailed feedback. This feature opens up considerable possibilities for design reviews and visual audits.

Customization of responses and tone

Users can now better configure the style of Claude AI Design’s responses. Whether adopting a formal tone for a professional document or a more creative register for a brainstorming session, the tool adapts to your preferences with greater precision. This flexibility is particularly appreciated by teams working on projects with distinct brand identities.

Project suggestions to get started

If you are not sure where to begin with Claude AI Design, here are some concrete project ideas that will allow you to discover the full potential of the tool and integrate it naturally into your workflow.

Creating a complete brand style guide

Describe your brand, your values, and your target audience to Claude AI Design, and ask it to generate a brand style guide proposal. The tool will suggest color palettes, typography choices, visual styles, and even application examples across different media. This is an excellent starting point for startups or entrepreneurs launching a new visual identity.

Designing a website architecture

Submit your website concept to the tool and ask it to structure the sitemap, define the main pages, subsections, and user journeys. Claude AI Design can also suggest the types of content suited to each page and the most relevant calls to action based on your industry.

Developing an advertising campaign concept

Describe your product, your audience, and your objectives, and let Claude AI Design propose several creative concepts for a campaign. The tool can generate slogans, visual ideas, key messages, and even a distribution strategy tailored to your communication channels.

Preparing a creative portfolio

For freelancers and agencies, Claude AI Design can help structure a compelling portfolio. It can write project descriptions, formulate convincing case studies, and suggest the best way to present your work to attract new clients.

• Creation of a brand style guide and visual identity
• Architecture and sitemap of a website or application
• Development of concepts for advertising campaigns
• Structuring a professional portfolio or presentation
• Generation of personas and user journeys for a UX project
• Writing creative content for social media
• Designing a consistent design system for a team

How to access Claude AI Design and use it effectively?

To access Claude AI Design, simply go to claude.ai/design from your browser. An Anthropic account is required to take advantage of all the features. Registration is simple and quick, and a free version is available to explore the tool before opting for a premium subscription that unlocks the most advanced capabilities.

To get the most out of the tool, a few best practices are worth following. First, be precise in your requests: the more context you provide, the more relevant and tailored the responses will be to your situation. Second, do not hesitate to iterate by refining your requests with each exchange to converge toward the desired result. Third, use the tool as a collaborator rather than an executor: ask questions, request alternatives, and challenge the proposals to obtain the best possible outcome.

Why adopt Claude AI Design now?

In a context where speed of execution and creative quality are major competitive advantages, Claude AI Design represents a sound investment for any creative professional. The tool makes it possible to significantly reduce the time spent on repetitive tasks or initial brainstorming phases, freeing up time for the most strategic and human aspects of creative work.

Moreover, the barrier to entry is very low: no particular technical skills are needed to use the tool. Its natural conversational interface makes it accessible to all profiles, from experienced designers to entrepreneurs discovering the world of creative work. This is one of its great strengths: democratizing access to high-quality creative assistance.

Finally, Anthropic is committed to continuous improvement, which means the tool will become even more powerful over time. Adopting Claude AI Design today means ensuring you stay at the forefront of innovations in artificial intelligence applied to creative work.

Conclusion

Claude AI Design is far more than a simple text assistant. It is a complete creative platform that supports professionals through every stage of their projects, from idea generation to the finalization of deliverables. Thanks to its enriched features, regular updates, and intuitive interface, it stands out as a reference tool for anyone who wants to combine creativity and efficiency. Visit claude.ai/design to discover for yourself everything this platform can bring to your creative work.