{"id":28102,"date":"2026-09-06T21:30:06","date_gmt":"2026-09-06T20:30:06","guid":{"rendered":"https:\/\/inboundcph.dk\/?p=28102"},"modified":"2026-09-10T13:27:03","modified_gmt":"2026-09-10T12:27:03","slug":"claude-vs-chatgpt","status":"publish","type":"post","link":"https:\/\/inboundcph.dk\/en\/claude-vs-chatgpt\/","title":{"rendered":"Claude vs. ChatGPT: Web Searches, Source Selection, and What It Takes to Be Cited"},"content":{"rendered":"
\n

What is the difference between Claude and ChatGPT in 2026?<\/h2>\n

Claude is Anthropic\u2019s AI assistant, and ChatGPT is OpenAI\u2019s. Both search the web, both write fluent Danish, and both cost $20 per month on the standard paid plan. The difference lies in the model families, the user profile, and the way they select and display sources.<\/strong> At InboundCPH, we often get asked the question \u201cClaude vs. ChatGPT\u201d\u2014both from people who need to choose a tool and from those who want to know if they\u2019re mentioned in either one.<\/p>\n

The models have different names, and those names change frequently. In September 2026, Anthropic will be selling the Mythos, Fable, Opus, Sonnet, and Haiku model families. The free Claude plan provides access to Sonnet and Haiku; the Pro plan unlocks Opus; and the Fable models are accessed using consumption credits. OpenAI runs the GPT-5.6 family with the Sol, Terra, and Luna variants, and GPT-6 Astra is being rolled out to Pro, Business, and Enterprise customers, according to OpenAI\u2019s documentation from September 2026. For a Danish company, the names matter less than what they represent. Both offer a heavy-duty model for complex tasks, a mid-range model for everyday use, and a lightweight model for high-volume processing.<\/p>\n

Prices are similar at the lower tiers. Claude Pro costs $20 per month, or $17 with annual billing, and Claude Max starts at $100 with five or twenty times the usage allowance. ChatGPT Plus costs $20, and ChatGPT Pro is available in two tiers at $100 and $200 per month. Both offer free plans with web search. These figures were taken from claude.com\/pricing and OpenAI\u2019s ChatGPT documentation on September 6, 2026, and may have changed by the time you read this.<\/p>\n\n\n\n\n\n\n\n\n\n\n\n
<\/th>\nClaude (Anthropic)<\/th>\nChatGPT (OpenAI)<\/th>\n<\/tr>\n<\/thead>\n
Model Families, Sept. 2026<\/td>\nMyth, Fable, Opus, Sonnet, Haiku<\/td>\nGPT-5.6 Sol, Terra, and Luna; GPT-6 Astra on the way<\/td>\n<\/tr>\n
Free Plan<\/td>\nYes, with Sonnet, Haiku, and web search<\/td>\nYes, using a web search<\/td>\n<\/tr>\n
Standard Plan<\/td>\nPro: $20\/month ($17 with annual payment)<\/td>\nPlus: 20 USD\/month.<\/td>\n<\/tr>\n
Top Plan for Individuals<\/td>\nMax: starting at 100 USD\/month (5x or 20x usage)<\/td>\nPro: 100 USD\/month (5x) or 200 USD\/month (20x)<\/td>\n<\/tr>\n
Team Schedule<\/td>\nTeam: $20 per seat per month with annual payment<\/td>\nBusiness: per user\/month.<\/td>\n<\/tr>\n
Answers in Danish<\/td>\nYes, you can<\/td>\nYes, you can<\/td>\n<\/tr>\n
Help Center in Danish<\/td>\nNo (12 languages; Danish is not one of them)<\/td>\nYes (help.openai.com\/da-dk)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

The Danish language is not a problem for either of them. Both Claude and ChatGPT write in Danish at a level suitable for customer emails, analyses, and proposal texts. The difference lies outside the model itself: ChatGPT has a Danish help center and user interface, while Claude does not. For a department that needs to onboard 15 colleagues, this is a practical difference, not a difference in quality.<\/p>\n

When do Claude and ChatGPT search the web, and how many sources do they cite?<\/h2>\n

Claude searches the web when the question depends on current or changing information, and draws on its training to provide answers based on stable knowledge. In practice, ChatGPT searches more frequently and typically cites more sources per answer. Both cite the pages they use with clickable links.<\/strong> That is the most important difference if you want to be evaluated based on whether your company is mentioned.<\/p>\n

Anthropic describes the rule in its developer documentation. Claude performs searches for news, prices, statistics, information about specific companies and products, and when the user requests it. Claude responds without performing searches when dealing with established facts, code, creative writing, and analysis of material that the user has entered themselves. A simple factual question typically triggers 1 to 3 searches, while a comparison of multiple providers can trigger 10 or more. Web search has been enabled globally on all Claude plans since May 27, 2025. Anthropic has not disclosed which search index Claude uses, but documentation from February 2026 shows that Anthropic now also indexes content itself using the Claude-SearchBot.<\/p>\n

ChatGPT has its own search tool and an index maintained by OpenAI itself. The OAI-SearchBot builds the index, and pages that have opted out of it appear only as navigation links and not as sources in the response. When we at InboundCPH tested the question \u201cHow does Claude choose its sources?\u201d in ChatGPT on September 6, 2026, with a Danish location setting, the model ran 2 searches, evaluated 25 web addresses, and cited 3 domains in the final response: a Danish agency website, an international developer blog, and Anthropic\u2019s own documentation. Google\u2019s AI mode cited three sources in response to the same question, none of which were Danish. We see this pattern consistently across our measurements: AI responses typically mention three to five companies, regardless of how many pages the model has read.<\/p>\n

The competition to be cited is therefore a competition for a limited number of spots. We\u2019ll walk you through how ChatGPT selects them in the guide to Visibility in AI Search<\/a>. Here, we focus on what makes Claude special: fewer searches, fewer sources, and a preference for pages that convey something specific.<\/p>\n

How do you compare Claude and ChatGPT on 10 Danish business questions?<\/h2>\n

They compare Claude and ChatGPT by asking the same 10 questions with web search enabled in both and recording four pieces of information per response: whether the model performed a search, how many sources it cited, how many were Danish, and which companies were recommended. InboundCPH calls this procedure the InboundCPH Claude Test Protocol.<\/strong> Our platform currently measures Google AI summaries, AI mode, and ChatGPT automatically, while Claude is measured manually using the protocol. We do not publish figures that we have not measured, so there are no percentages for Claude listed here yet.<\/p>\n

The questions cover the customer journey from information to decision\u2014the same breakdown as in our platform: Information, Compare, Selection, and Decision. According to 6sense (2025), 94 % of B2B buyers use AI assistants in the purchasing process, so the questions should resemble those a buyer actually asks.<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Questions<\/th>\nTravel Phase<\/th>\nWhy it's included<\/th>\n<\/tr>\n<\/thead>\n
What are the rules for e-invoicing in Denmark in 2026?<\/td>\nInformation<\/td>\nForcing both to seek help: the rules are changing<\/td>\n<\/tr>\n
Is AI SEO better than regular SEO?<\/td>\nInformation<\/td>\nFrom InboundCPH's own portfolio, known ChatGPT behavior<\/td>\n<\/tr>\n
How much does an interim CFO cost per month in Denmark?<\/td>\nInformation<\/td>\nPrice inquiries trigger a search on both sites<\/td>\n<\/tr>\n
Which ERP system is suitable for a manufacturing company with 80 employees?<\/td>\nCompare<\/td>\nComparison with Multiple Partial Searches<\/td>\n<\/tr>\n
Which CRM system is best suited for a B2B company with 20 salespeople?<\/td>\nCompare<\/td>\nTests that mention Danish retailers<\/td>\n<\/tr>\n
How much does an AI-powered SEO campaign cost?<\/td>\nCompare<\/td>\nFrom InboundCPH's portfolio: a Danish pricing issue<\/td>\n<\/tr>\n
Which accounting firms are recommended for medium-sized businesses?<\/td>\nSelection<\/td>\nShortlist question: Who is mentioned by name?<\/td>\n<\/tr>\n
Which leasing partner is best for company cars?<\/td>\nSelection<\/td>\nTesting the Effect of the \u201cBest\u201d Phrase on Source Selection<\/td>\n<\/tr>\n
Which agency should we choose for Google Ads in Denmark?<\/td>\nSelection<\/td>\nLocal shortlist; requires Danish sources<\/td>\n<\/tr>\n
How much does it cost for a company to switch pension providers?<\/td>\nDecision<\/td>\nClose to the decision-making process, combining price and rules<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n
1<\/div>\n
\n

Ask the question exactly as it is in both assistants<\/h3>\n

Use the same wording, on the same day, with web search enabled in both Claude and ChatGPT. Do not write \u201csearch the web\u201d in the prompt, because then you won\u2019t be measuring the model\u2019s own choice.<\/p>\n<\/p><\/div>\n<\/p><\/div>\n

\n
2<\/div>\n
\n

Save the entire answer and the list of sources<\/h3>\n

Copy the answer and all the links cited into a spreadsheet. A simple \u201cyes\u201d or \u201cno\u201d to the question \u201cWere we mentioned?\u201d isn\u2019t enough when, in three months, you\u2019ll have to explain what changed.<\/p>\n<\/p><\/div>\n<\/p><\/div>\n

\n
3<\/div>\n
\n

Write down the four goals for each answer<\/h3>\n

Whether the model was searched (yes\/no), the number of cited sources, the number of Danish sources, and the names of companies mentioned in the response. Four columns per assistant result in eight figures per question.<\/p>\n<\/p><\/div>\n<\/p><\/div>\n

\n
4<\/div>\n
\n

Repeat after 30 days<\/h3>\n

Trends change, and your content evolves. Only the second measurement will show whether a new page has made a difference.<\/p>\n<\/p><\/div>\n<\/p><\/div>\n

The example from the previous section illustrates what the protocol detects. In \u201cHow Does Claude Choose His Sources?\u201d, ChatGPT read 25 pages and cited 3. Among the 25 were six guides to source criticism from Danish university libraries, which the model found via the word \u201csources\u201d and subsequently rejected. You\u2019ll only see that noise if you save the entire list of sources.<\/p>\n

What does it take to be cited in Claude?<\/h2>\n

To be cited in Claude, three conditions must be met: the page must be accessible to Claude-SearchBot and Claude-User; it must answer the question precisely in the first sentence; and the sender must be unambiguous. InboundCPH checks these three layers in that order when a customer\u2019s page does not appear in Claude\u2019s response.<\/strong> The order is important because a problem in the first layer renders the next two irrelevant.<\/p>\n

In short, the way it selects sources works the same as ChatGPT, just with fewer options. Claude assesses whether the question requires up-to-date knowledge, sends one or more search queries, retrieves the pages that look most promising, and extracts the passages that contain a specific answer. The passage is cited, not the page as a whole. That\u2019s why a single sentence with numbers and units trumps three paragraphs about why AI is changing everything. In an analysis of 75,000 brands, Ahrefs found that brand mentions correlate about three times more strongly with AI visibility than backlinks (0.66 vs. 0.22). This aligns with what we see in the raw responses: the model cites pages that discuss named entities with numbers.<\/p>\n

\n
\n
Level 1 \u00b7 Access<\/div>\n

Claude-SearchBot and Claude-User can retrieve the page<\/h3>\n

Check your robots.txt file, CDN rules, and bot protection. A well-intentioned \u201cblock all AI bots\u201d rule from 2024 often affects both types of bots. Without access, the content will never be considered.<\/p>\n<\/p><\/div>\n

\n
Level 2 \u00b7 Precision<\/div>\n

The first sentence under each heading is the answer<\/h3>\n

Claude quotes passages. Include numbers with units, years, and sources in the visible text. State any reservations openly: \u201cdepending on the starting point\u201d is easier to quote accurately than an unconditional promise.<\/p>\n<\/p><\/div>\n

\n
Layer 3 \u00b7 Sender<\/div>\n

Named author, clear organization, consistent facts<\/h3>\n

The same company name, the same address, and the same information about yourselves throughout the site and in trade publications. The model selects sources it can vouch for, and conflicting information about the sender undermines trust.<\/p>\n<\/p><\/div>\n<\/p><\/div>\n

\n

Claude is also reading your offers<\/h2>\n

Claude is primarily used to analyze documents that users upload themselves: reports, contracts, and proposals. Your content thus encounters Claude twice\u2014as a web source and as an attachment to a decision. Write in a way that ensures your text can withstand critical review by a model: clear statements, figures, and disclaimers. This is the same standard we at InboundCPH follow when we write. AI-SEO Content<\/a> for customers.<\/p>\n<\/p><\/div>\n

How do you add ClaudeBot, Claude-User, and Claude-SearchBot to robots.txt?<\/h2>\n

Anthropic uses three bots, each with its own user-agent name: ClaudeBot retrieves data for model training, Claude-User fetches pages on a user\u2019s behalf, and Claude-SearchBot builds the search index. If you want to be cited in Claude, Claude-User and Claude-SearchBot must have access. If you block only ClaudeBot, you are opting out of training while maintaining your visibility.<\/strong> This breakdown is available in Anthropic's Help Center, last updated on April 7, 2026. The older name, Claude-Web, no longer appears in the documentation, and any rules related to it can be deleted.<\/p>\n

OpenAI uses the same three-part structure. GPTBot fetches data for training, OAI-SearchBot builds the search index, and ChatGPT-User fetches pages when a user makes a query. OpenAI itself states that robots.txt rules do not necessarily apply to ChatGPT-User because the fetching is initiated by a human. A change to robots.txt takes effect in search results after about 24 hours. Anthropic supports the \"Crawl-delay\" directive to slow down the crawl and advises against blocking by IP address, as this prevents the bots from reading the robots.txt file at all.<\/p>\n\n\n\n\n\n\n\n\n
Purpose<\/th>\nAnthropic (Claude)<\/th>\nOpenAI (ChatGPT)<\/th>\nEffect of Blocking<\/th>\n<\/tr>\n<\/thead>\n
Training Models<\/td>\nClaudeBot<\/td>\nGPTBot<\/td>\nFuture content is excluded from training. No effect on quotes<\/td>\n<\/tr>\n
Search Index<\/td>\nClaude-SearchBot<\/td>\nOAI-SearchBot<\/td>\nThe page disappears as a source in AI responses<\/td>\n<\/tr>\n
Pickup for a user<\/td>\nClaude-User<\/td>\nChatGPT User<\/td>\nThe page cannot be accessed live. At OpenAI, robots.txt does not always apply<\/td>\n<\/tr>\n
Advertise<\/td>\nNone<\/td>\nOAI-AdsBot<\/td>\nApplies only to pages used as landing pages in ChatGPT Ads<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

A robots.txt file that allows indexing but disallows crawling for both looks like this. Copy the block and customize it according to your own rules for other robots.<\/p>\n

User-agent: ClaudeBot\nDisallow: \/\n\nUser-agent: GPTBot\nDisallow: \/\n\nUser-agent: Claude-SearchBot\nAllow: \/\n\nUser-agent: Claude-User\nAllow: \/\n\nUser-agent: OAI-SearchBot\nAllow: \/\n\nUser-agent: ChatGPT-User\nAllow: \/<\/code><\/pre>\n
Training is a business decision, and there may be good reasons to say no. Search and user acquisition are visibility decisions, and for most B2B companies, the answer is yes. The basic rules for the file, including order and wildcard syntax, are outlined in our Guide to robots.txt<\/a>. If ChatGPT is new to you, it offers Our ChatGPT Guide<\/a> the basics before you go for your checkup.<\/p>\n
Frequently Asked Questions About Claude and ChatGPT<\/h2>\n
\n
\nIs Claude better than ChatGPT at Danish?<\/summary>\n
\n
Both Claude and ChatGPT write fluent Danish, and InboundCPH has not observed any difference in the quality of responses to Danish business-related questions. The difference lies outside the model itself: ChatGPT has a Danish help center and user interface, while as of September 2026, Claude\u2019s help center is available in 12 languages but not Danish. Choose based on how you work, not on language.<\/p>\n<\/div>\n<\/details>\n
\nHow much will Claude and ChatGPT cost in 2026?<\/summary>\n
\n
Claude Pro costs $20 per month ($17 with annual billing), and Claude Max starts at $100. ChatGPT Plus costs $20, and ChatGPT Pro costs $100 or $200 per month, depending on usage. Both offer free plans with web search. Prices were obtained from Anthropic and OpenAI on September 6, 2026.<\/p>\n<\/div>\n<\/details>\n
\nHow does Claude search the web?<\/summary>\n
\n
Claude determines on its own whether a question requires up-to-date information and only performs a search if that is the case. According to Anthropic\u2019s documentation, a simple factual question triggers 1 to 3 searches, while a comparison can trigger 10 or more. Claude then fetches the pages, selects passages, and displays them as quotes with links. Web search has been available on all Claude plans since May 27, 2025.<\/p>\n<\/div>\n<\/details>\n
\nShould I block ClaudeBot in robots.txt?<\/summary>\n
\n
ClaudeBot only retrieves data for training Anthropic\u2019s models. If you block it, it will not affect whether your pages are cited in Claude\u2019s responses. Quotes depend on Claude-SearchBot and Claude-User, which Anthropic documented as separate bots in February 2026. If you want to opt out of training while maintaining visibility, block ClaudeBot and allow the other two.<\/p>\n<\/div>\n<\/details>\n
\nCan InboundCPH measure our visibility in Claude?<\/summary>\n
\n
In September 2026, InboundCPH\u2019s AI visibility platform automatically measures Google AI summaries, AI mode, and ChatGPT, and saves the raw response for each question. We evaluate Claude manually using InboundCPH\u2019s Claude test protocol, which consists of 10 questions and four metrics per response. The results are included in the AI visibility analysis alongside the automated measurements.<\/p>\n<\/div>\n<\/details><\/div>\n
\n
See what Claude and ChatGPT have to say about you<\/h2>\n
An AI visibility analysis shows you exactly what ChatGPT and Google AI respond with when your customers ask questions, and who is recommended instead of you. We evaluate Claude using the test protocol from this article. You\u2019ll get the raw response behind each number.<\/p>\n
    Order an AI Visibility Analysis<\/a>\n  <\/div>\n
Sources<\/h2>\n