← All articles
AI Visibility 7 April 2026 15 min read

How AI Search Actually Works: When ChatGPT, Claude and Perplexity Search the Web

Understand when and how AI models trigger real-time web search versus relying on training data — and what that means for your B2B content strategy.

Not Every AI Answer Comes From the Same Place

When a Swiss procurement manager asks ChatGPT "Who are the leading cloud ERP providers in the DACH region?", the answer might come from the model's training data, from a live web search, or from a blend of both. Most B2B marketers do not realise this distinction exists — and it fundamentally changes how you should approach your content strategy.

Each major AI platform handles web search differently. Some always search. Some never do unless triggered. Some have permanently fresh data. Understanding these mechanics is not academic — it determines whether your latest case study, your new product launch, or your press release will ever surface in an AI answer.

This article provides a technical but accessible breakdown of how each major AI platform decides when to search the web, what data sources it draws from, and what this means for how you should structure your content strategy. Armed with this knowledge, you can make informed decisions about where to invest your marketing resources for maximum AI visibility.

How Each AI Platform Handles Search

ChatGPT: Selective Search via Tools

ChatGPT does not search the web by default. It uses a tool called web_search that it decides to invoke based on the nature of the query. When a user asks something that requires up-to-date information — recent events, current pricing, today's news — the model recognises this and triggers a Bing-powered web search before generating its response.

But here is the critical nuance: for many B2B queries, ChatGPT relies entirely on its training data. If someone asks "What are the best HR software solutions for Swiss SMBs?", ChatGPT may answer purely from what it learned during training — without searching the web at all. Your company either exists in that training data or it does not.

When does ChatGPT decide to search? The triggers include:

  • Time-sensitive queries. Anything implying "current", "latest", "2025", or "today" typically triggers search.
  • Specific factual lookups. Questions about specific companies, prices, or contact details often prompt a search.
  • Recent events. News, product launches, or industry developments from recent months.
  • Explicit search requests. When a user says "search for" or "find me" in their prompt.

For general category queries — the kind B2B buyers ask most — ChatGPT frequently answers from memory alone.

ChatGPT Search: The Bing Connection in Detail

When ChatGPT does decide to search, it sends queries to Bing's search API and receives results back. This has several important implications for Swiss B2B companies:

  • Bing indexation is mandatory. If your pages are not in Bing's index, ChatGPT literally cannot find them during web search. For Swiss companies that have never registered with Bing Webmaster Tools (the vast majority), this represents a complete blind spot.
  • Bing ranking factors apply. When ChatGPT receives Bing search results, the pages that rank highest in Bing are the ones most likely to be cited. Bing's ranking algorithm differs from Google's — it places relatively more weight on exact-match domains, social signals, and content freshness.
  • Multiple searches per response. ChatGPT often performs multiple searches to answer a single query. It might search for "best ERP software Switzerland," then "ERP pricing Swiss SMB," then a specific company name. Each search retrieves different results, and the final answer synthesises all of them.
  • Search results are processed, not displayed. Unlike Perplexity, ChatGPT does not show users the raw search results. It reads the retrieved pages, extracts relevant information, and weaves it into a natural-language response. This means the quality and clarity of your content matters — the AI needs to be able to extract useful information from your pages quickly.

Claude: Web Search on Demand

Anthropic's Claude uses a tool called web_search that functions similarly to ChatGPT's approach. Claude decides whether a query needs fresh information and triggers a web search when it determines the question requires current data or specific facts it cannot confidently answer from training alone.

Claude tends to be conservative about when it searches. For broad industry questions, it often relies on training data. This means your presence in the sources Claude was trained on — industry publications, authoritative websites, directories — is especially important.

Claude's Conservative Approach: What It Means for Your Strategy

Claude's reluctance to search creates a specific challenge for Swiss B2B companies. Because Claude relies more heavily on training data than ChatGPT, getting into Claude's training data is disproportionately important. Here is what this means in practice:

  • Long-established web presence matters more. Pages that have existed for a year or more are more likely to be in Claude's training data than recently published content.
  • Third-party mentions carry extra weight. If your company is mentioned on Wikipedia, in academic papers, or in established industry publications, Claude is more likely to "know" about you from training data.
  • Your llms.txt file helps. When ClaudeBot crawls your site, your llms.txt file provides structured information that can be incorporated into training data, giving Claude a clean understanding of your company.
  • When Claude does search, make it count. For the queries where Claude does trigger web search, ensure your content is the most relevant, specific, and authoritative result available. Claude's search tends to be more targeted than ChatGPT's, so highly specific content performs well.

Perplexity: Always Searching

Perplexity is fundamentally different. It always performs a web search for every query. There is no "training data only" mode. Every answer is grounded in real-time search results, and every answer includes source citations.

This makes Perplexity the most SEO-like of the AI platforms. Your current web presence, your Bing indexation, your page speed, and your content freshness all directly influence whether Perplexity finds and cites you. If your website ranks well in traditional search, you have a head start with Perplexity.

Perplexity's Citation Model: A Closer Look

Perplexity displays sources prominently — numbered footnotes throughout the response and a full source list at the bottom. This citation model has unique implications:

  • Your page title and meta description matter. These appear in Perplexity's source cards. A clear, descriptive page title like "Cloud ERP Solutions for Swiss Manufacturing — Pricing, Features, and Comparisons" is more likely to be clicked than "Muster AG — Solutions."
  • Content depth is rewarded. Perplexity tends to cite pages with comprehensive, detailed content over thin or superficial pages. A 2,000-word guide outperforms a 200-word overview.
  • Recent content has an edge. Perplexity's search prioritises fresh content. A page updated last week will often outrank an identical but older page.
  • Multiple citations per response. Perplexity typically cites three to eight sources per response. This means even if you are not the primary source, you can still appear as a supporting citation — building visibility gradually.

Google AI Mode and AI Overviews: Always Fresh

Google's AI features — Gemini-powered AI Overviews and the newer AI Mode — always have access to Google's full, real-time search index. They do not rely on a separate training data cutoff in the same way. If your page is indexed by Google and ranks for a relevant query, it can appear in an AI Overview immediately.

For Swiss B2B companies already investing in SEO, this is good news. Your Google SEO work directly feeds into Google's AI features.

The Search Behaviour Matrix

This summary table makes it easy to understand how each platform handles search and what it means for your content strategy:

Platform Searches Search Source Shows Citations Your Priority
ChatGPT Selectively Bing When searching Bing indexation + training data
Claude Conservatively Web search partners When searching Training data (highest priority)
Perplexity Always Multiple indices Always SEO fundamentals + fresh content
Google AI Always Google index Source cards Google SEO

The Two-Track Problem for B2B

This creates a fundamental challenge for B2B content strategy. Your content needs to work on two tracks simultaneously:

  1. Training data track. Content that has been published long enough and on authoritative enough sources to be absorbed into model training. This affects ChatGPT and Claude responses when they do not search.
  2. Real-time search track. Content that is currently indexed, fresh, and well-optimised for search. This affects Perplexity always, Google AI always, and ChatGPT/Claude when they decide to search.

Most Swiss B2B companies focus exclusively on one track — usually the second, because it resembles traditional SEO. But ignoring the training data track means you are invisible in a significant portion of AI interactions.

Quantifying the Two-Track Split

How much of AI interaction relies on training data versus real-time search? The split varies by platform and query type, but here are approximate ranges based on our testing across Swiss B2B categories:

  • Perplexity: 100% real-time search. Training data plays a minimal role.
  • Google AI: 90-100% real-time search. Google's AI features have full access to the live index.
  • ChatGPT: Approximately 40-60% of B2B category queries trigger web search. The remainder are answered from training data. Queries that include temporal markers, specific requirements, or comparison requests are more likely to trigger search.
  • Claude: Approximately 25-40% of B2B category queries trigger web search. Claude is the most training-data-dependent for general queries.

This means that if you only optimise for real-time search, you are invisible in roughly 30-40% of ChatGPT interactions and 60-75% of Claude interactions for your category. Conversely, if you only focus on training data, you miss Perplexity entirely and the search-triggered portions of ChatGPT and Claude.

How Training Data Inclusion Actually Works

AI models are trained on massive datasets scraped from the web at specific points in time. Common Crawl, a publicly available web archive, forms the basis for many models. But each AI provider also runs proprietary crawlers — GPTBot for OpenAI, ClaudeBot for Anthropic — that build additional training datasets.

Getting into training data requires:

  • Consistent web presence over time. Pages that have existed for months or years are more likely to be captured in training snapshots.
  • Authority signals. Pages with backlinks, citations, and cross-references from other authoritative sources are prioritised in training data curation.
  • Crawler access. If your robots.txt blocks GPTBot or ClaudeBot, your content will not be in their training data. Full stop.
  • Content quality. Training data pipelines include quality filters. Thin content, duplicate pages, and marketing fluff are often filtered out.

The Training Data Pipeline in Detail

Understanding the training pipeline helps you appreciate what kind of content is most likely to be included:

  1. Web crawling. AI providers crawl the web continuously, building massive archives of web content. GPTBot, ClaudeBot, and Common Crawl all contribute to different training datasets.
  2. Data cleaning. Raw web content is cleaned to remove duplicate pages, navigation elements, advertisements, and boilerplate text. Only the core content of each page is retained.
  3. Quality filtering. Automated systems assess content quality based on factors like text length, vocabulary diversity, factual density, and source authority. Low-quality content — thin pages, keyword-stuffed articles, auto-generated text — is filtered out.
  4. Deduplication. Content that appears identically across multiple domains (a common issue with syndicated content and directory listings) is deduplicated, with the most authoritative version retained.
  5. Training. The filtered, cleaned dataset is used to train or fine-tune the model. The model learns patterns, associations, and factual relationships from this data.

The practical implication: content that passes through quality filtering and deduplication is content that is original, specific, substantive, and hosted on authoritative domains. This is why a single well-researched article on a respected industry publication can have more training data impact than dozens of thin directory listings.

Practical Implications for Your Content Strategy

Knowing how AI search works changes what you should publish and where:

For the Training Data Track

  • Publish evergreen expert content now. Detailed guides, industry analyses, and technical comparisons published today will be included in the next training data snapshot — which might not happen for months. The sooner you publish, the sooner you enter the training pipeline.
  • Get mentioned on authoritative third-party sites. Industry publications, Swiss business directories, and established news sources are weighted heavily in training data. A mention on Handelszeitung carries more training-data weight than ten blog posts on your own site.
  • Ensure your llms.txt and schema markup are in place. These structured data sources are specifically designed to be captured by AI training pipelines.

For the Real-Time Search Track

  • Keep content fresh. Update key pages regularly. Perplexity and Google AI prioritise recent content.
  • Optimise for Bing, not just Google. ChatGPT's web search is Bing-powered. If you are not in the Bing index, ChatGPT cannot find you even when it does search.
  • Publish news and press releases. Time-sensitive content is exactly what triggers AI web searches. Regular press releases ensure you appear when AI tools go looking for current information.
  • Monitor what triggers searches in your category. Use a tool like per4mx to test different prompt formulations and see which ones cause AI models to search versus answer from memory. This tells you where the gaps are.

The Dual-Track Content Calendar

Here is a practical content calendar that addresses both tracks simultaneously:

  • Monthly: Publish one evergreen expert article (training data + real-time). A detailed, specific guide that answers a buyer question in your category. Examples: "How to Evaluate Swiss Cloud Hosting Providers for Regulated Industries" or "Complete Guide to Procurement Automation for Swiss Manufacturers." These articles build training data presence over time while being immediately available for real-time search retrieval.
  • Monthly: Update one existing key page (real-time). Refresh your most important product, service, or About page with current information, new statistics, or recent achievements. Adding a "Last updated: [date]" line signals freshness to both search engines and AI crawlers.
  • Quarterly: Issue a press release (real-time + training data). Press releases create immediate multi-source coverage for real-time search while also entering the historical record that informs training data. A strong company boilerplate ensures your key facts are distributed consistently.
  • Quarterly: Contribute an expert article to an industry publication (training data). Guest articles on sites like Inside IT, Handelszeitung, or sector-specific publications carry high authority weight in training data curation. These also generate backlinks and brand mentions that compound over time.
  • Annually: Update your llms.txt file and schema markup (training data + real-time). Ensure these machine-readable assets reflect your current company information, product offerings, and key metrics.

The Prompt Matters More Than You Think

A subtle but critical point: the exact wording of a user's prompt determines whether an AI searches or not. "Best ERP systems" might get a training-data answer. "Best ERP systems in 2025" will almost certainly trigger a search. "Compare current ERP pricing for Swiss manufacturers" — definitely a search.

This means your AI visibility can vary dramatically depending on how your prospects phrase their questions. Monitoring a range of prompt variations — not just one — is essential to understanding your true visibility. per4mx runs multiple prompt variations across all major AI platforms to give you a comprehensive picture.

Prompt Variations and Their Search Implications

Here are examples of how slight prompt variations can change whether an AI searches or not, using a Swiss IT consulting category as an example:

  • "Best IT consulting firms in Switzerland" — Likely answered from training data by ChatGPT and Claude. Search triggered by Perplexity and Google AI.
  • "Best IT consulting firms in Switzerland 2026" — Search triggered on all platforms due to the temporal marker.
  • "Compare IT consulting firms in Zurich for a cloud migration project" — Mixed. The specificity may or may not trigger search depending on the platform.
  • "Who should I hire for SAP S/4HANA migration in Switzerland? Budget around CHF 200,000, 18-month timeline" — Almost certainly triggers search on all platforms due to the specificity and recency implications.
  • "Search for Swiss IT consultancies that specialise in banking sector compliance" — Explicit search request triggers search on ChatGPT and Claude.

The takeaway: your content needs to serve both search-triggered and training-data-based queries. Content that is specific, fact-rich, and addresses detailed scenarios works for both modes.

What This Means for Swiss B2B Companies

The companies that win in AI visibility will be those that understand the mechanics, not just the surface. Knowing that ChatGPT sometimes searches and sometimes does not, knowing that Perplexity always searches, knowing that Claude is conservative about triggering search — these insights should directly shape your content calendar, your PR strategy, and your technical SEO priorities.

The actionable takeaway: build for both tracks. Invest in lasting, authoritative content that will enter training data. Simultaneously, maintain fresh, well-indexed content that AI tools can find in real time. Cover both, and you are visible regardless of how the AI decides to answer. For a complete action plan, see our 30-day GEO roadmap, and learn why being present in multiple AI indices is the foundation of both tracks.

Frequently Asked Questions

Can I force ChatGPT to search for my company?

You cannot control when ChatGPT decides to search. However, you can influence it indirectly. If your content is framed around current, time-sensitive topics — "2026 pricing," "latest features," "current compliance requirements" — prompts that reference these topics are more likely to trigger search. Additionally, if a user explicitly asks ChatGPT to "search for" or "look up" information, it will search. Your strategy should focus on being present in Bing's index (so ChatGPT can find you when it does search) and in training data (so ChatGPT knows about you when it does not search).

How often does training data get updated?

Training data update schedules vary by provider and are not publicly disclosed on a fixed cadence. Major model releases (like GPT-4o, Claude 3.5, etc.) typically include updated training data, but the exact cutoff dates vary. As a general guideline, expect training data to lag reality by three to twelve months. This is why the real-time search track is important for timely content while the training data track is important for establishing long-term presence. Content published today may not appear in training-data-based answers for months — but it will be immediately available for search-based answers on Perplexity and Google AI.

Does page speed affect AI search results?

Yes, both directly and indirectly. AI crawlers have timeout thresholds — if your page takes too long to load, the crawler may abandon the attempt, resulting in incomplete or missing content in the index. For real-time search retrieval, slow pages may be deprioritised in favour of faster alternatives. Additionally, page speed is a ranking factor for both Google and Bing, meaning slow pages rank lower in the search results that AI models retrieve. Aim for page load times under two seconds for your key content pages.

If my company is new, how long before AI models know about me?

New companies face a cold-start problem with AI visibility. Training-data-based visibility takes the longest — potentially six to twelve months or more, depending on when the next training data update occurs and whether your content has accumulated enough authority to pass quality filters. Real-time search visibility can be established much faster: register with Bing Webmaster Tools, ensure your site is technically accessible to AI crawlers, and publish high-quality content. Perplexity can surface your content within days of publication if it ranks in web search. ChatGPT can find you within weeks once your pages are indexed by Bing. For new companies, the priority should be establishing real-time search visibility first while building the kind of authoritative web presence that will eventually enter training data.

Ready to take action?

Check your AI visibility for free

See how ChatGPT, Claude, Perplexity, and Gemini describe your company today. Get a free visibility report in minutes.