Blog By Esteve Castells
When marketers ask how their brand shows up in AI answers, the research they find is almost always about ChatGPT. The large citation studies of the past two years measured ChatGPT, Google AI Overviews and Perplexity. Claude is often left out, and when it is included, its numbers are rarely published on their own.
Claude.ai started searching the web in March 2025, and Anthropic made web search available on every Claude plan in May 2025, with citations shown in the answer. People now ask it which CRM to buy, which VPN to trust and which running shoe to start with. What they see is one answer. Nobody outside Anthropic sees the others: the answers the same question would have produced a minute earlier, in another chat, or in another language.
We wrote 80 buying questions with no brand names in them, put each one in five languages, and asked all 400 of them to the real Claude.ai app 18 times, with Claude Sonnet 5.5 and with Claude Haiku 5.5. Every answer came from a fresh chat with web search on. That gave us 14,400 answers, and with them a view of Claude that a single conversation cannot give: how much the brands change, which brand comes first, when Claude decides to search, and what it cites when it does.
Headline findings #
- One answer shows part of the list. A Sonnet 5.5 answer named a median of 6 brands. The 18 answers to the same question named a median of 15 different brands, or 11 counting only brands named at least twice.
- Identical answers are rare. Two Sonnet answers to the same question named exactly the same set of brands 11.9% of the time, and the same brands in the same order 6.4% of the time. Only 7 of the 400 questions got the same set in all 18 answers.
- The first brand moves. Sonnet mentioned the same brand first in all 18 answers for 32.5% of the questions. Haiku did so for 24.8%.
- Haiku searches, but rarely shows where it looked. Both models searched the web in about 6 of 10 answers. Of the Sonnet answers that searched, 97.8% showed a source link. Of the Haiku answers that searched, 19.0% did.
- The category decides whether Claude searches. Sonnet searched in every answer about smartphones, earbuds and robot vacuums, and in 1.4% of answers about skincare and language apps.
- No Reddit, no Wikipedia, no YouTube. Not one of the 15,937 sources the two models cited came from those three sites.
- Each market gets its own list. Of the brands Sonnet named in at least half of the answers to a question, only 22.0% did the same in all five languages.
How much the brand list changes #
Ask Claude "What is the best VPN?" once and you get a tidy shortlist. Ask it 18 times and a pattern appears: NordVPN, Surfshark, Proton VPN and Mullvad were named in every answer, and NordVPN came first in 14 of the 18. Few questions were that stable.
Most questions are looser. For "Which CRM integrates best with email and calendar for a startup?", Sonnet named 18 different brands across 18 answers. Attio and HubSpot each came first 7 times, and Copper 4 times. Two teams checking that question once could each come away convinced a different CRM leads.
Across all 400 questions, the numbers look like this.
| Sonnet 5.5 | Haiku 5.5 | |
|---|---|---|
| Brands in one answer (median) | 6 | 5 |
| Different brands across 18 answers (median per question) | 15 | 13 |
| Same, counting only brands named at least twice | 11 | 9 |
| Brands named in at least 16 of 18 answers (share of all brands named for a question) | 18.8% | 14.7% |
| Brands named in exactly 1 of 18 answers | 26.4% | 29.0% |
| Two answers name exactly the same brands | 11.9% | 11.8% |
| Two answers name the same brands in the same order | 6.4% | 6.7% |
| Two answers share the same top three, in order | 31.6% | 22.8% |
Behind every question there is a core of brands Claude names almost every time, a middle that comes and goes, and a long tail that shows up once. With Sonnet, it took a median of 11 answers to see 90% of the brands that all 18 answers named. A quarter of the questions produced 21 or more different brands, and one produced 68.
This fits earlier work. In January 2026, SparkToro had 600 volunteers run the same prompts through ChatGPT, Claude and Google's AI and found that getting the same list twice was very unlikely. Momentic compared same-day reruns across 11 platforms in 2026. Neither published Claude figures at this depth, and the setups differ, so the numbers are not directly comparable. What 18 runs per question add is the shape of the variation: which brands are fixtures, and how long the tail gets.
How often the first brand changes #
Being named first is what most brands want, so we looked at it separately. "First brand" here means the first brand mentioned in the answer.
Some questions have a clear winner. "What are the best running shoes for beginners?" opened with Brooks in all 18 answers, and "What is the best project management tool for a remote team?" opened with Asana in all 18. Others are split. "Which phone has the best camera?" opened with Apple 7 times, Samsung 7 times and Vivo 4 times.
Across the 400 questions, Sonnet kept the same first brand in all 18 answers for 32.5% of them and Haiku for 24.8%. The most frequent first brand still came first in most answers: 78.9% of the 18 for Sonnet and 75.9% for Haiku, on average across questions. The leader usually leads, yet a ranking built from one answer per question can change from one run to the next with nothing else changing.
Sonnet and Haiku behave differently #
The two models agreed with each other less than each agreed with itself. Two Sonnet answers to the same question had a brand-set overlap (Jaccard index) of 0.60. A Sonnet answer and a Haiku answer to the same question in the same round had 0.48. Their most frequent first brand was the same for 70.5% of the questions. Haiku also gave shorter answers, with a median of 1,429 characters against 1,886 for Sonnet, and named one brand fewer.
The biggest difference is in sources. Haiku ran a web search in 61.0% of answers and Sonnet in 56.6%. When they searched, both received about 17 to 18 pages per answer. But 97.8% of the Sonnet answers that searched showed at least one source link, against 19.0% of the Haiku answers that searched. Across all answers, that is 2.31 source links per answer for Sonnet and 0.23 for Haiku. With Haiku, a page can be among the search results behind an answer without the reader ever seeing a link to it.
When Claude searches, and when it does not #
Claude decides for itself whether to search. Sonnet searched in 56.6% of answers overall, but that average hides a sharp split by category.
| Category | Sonnet 5.5 answers with a web search |
|---|---|
| Wireless earbuds, robot vacuums, smartphones | 100% |
| Electric cars | 99.7% |
| Banks and money transfers | 98.3% |
| Laptops | 90.0% |
| VPNs | 81.9% |
| CRM software | 33.3% |
| Project management | 23.9% |
| Online courses | 3.6% |
| Travel booking | 3.1% |
| Skincare, language apps | 1.4% |
Consumer electronics and electric cars were searched almost every time, and so were banks and money transfers. In skincare, language apps, travel booking and online courses, Sonnet nearly always answered without searching. The same question does not always get the same treatment either: for 37.2% of questions, Sonnet searched in some of the 18 answers and not in others.
When it does search, Claude searches like someone shopping this year. Its queries had a median of six words, and 98.9% of Sonnet's queries and every one of Haiku's included "2026". Queries like that tend to surface review roundups and "best of 2026" lists.
What Claude cites, and what it does not #
We counted each domain once per answer. Across both models, the answers cited 15,937 sources and the web searches returned 97,457 domains.
None of the cited sources was Reddit, Wikipedia or YouTube. Of the 97,457 domains the searches returned, 32 were. In our study of AI shopping answers across ChatGPT, Google AI Mode, AI Overviews, Gemini and Perplexity, Reddit and YouTube together accounted for 10.4% of citations. That study used different questions and engines, so read it as context rather than a like-for-like comparison.
The answers alone cannot tell us why. Claude's web search has been reported to run on Brave Search, which Anthropic has not confirmed, and what Claude can cite depends on what its search returns. Whatever the reason, a plan for Claude visibility built on Reddit threads or a Wikipedia page relies on sources Claude did not cite in this study.
Cited sources also rotate. Two Sonnet answers to the same question had a cited-domain overlap (Jaccard index) of 0.46 on average, so roughly half the cited sites change between identical questions.
Each market gets its own list #
Every question was asked in English from the United States, Spanish from Spain, German from Germany, French from France and Brazilian Portuguese from Brazil, so language and country change together. We called a brand "core" for a question when a model named it in at least 9 of the 18 answers.
For Sonnet, only 22.0% of core brands were core in all five versions of the same question, and a brand that was core in one language was core in another 58.3% of the time. Haiku was similar, at 24.7% and 59.9%. Some of that is simply local markets. Asked for the best bank for freelancers, Sonnet's core list in the United States included Mercury and Relay, in Spain BBVA and Santander, in Germany Fyrst and Finom, in France Shine and Qonto, and in Brazil Nubank and Inter. But global categories move too: for "What is the best smartphone right now?", the English core list held three brands (Samsung, Apple and Google), while the Spanish one held eight.
What this means for brands #
- Measure Claude over many answers. One chat shows part of the list. A visibility number for Claude should come from repeated answers to the same question, read as a rate.
- Treat first position as a frequency. Even a clear leader comes first in most answers and misses some. Report the share of answers in which you come first.
- Know which Claude you are measuring. Sonnet and Haiku name different brands and show sources very differently.
- Check whether Claude searches in your category. Where it rarely searches, the answer does not draw on the web, and a new page has few ways in. Where it always searches, the review sites and comparison pages it finds matter most.
- Check where your Reddit and Wikipedia work shows up. Claude never cited those sites in these answers.
- Measure each market in its own language and country. Visibility in one only partly carries over to the next.
Methodology #
Questions. We wrote 80 buying questions for this study: 20 categories, from CRM software, VPNs and accounting tools to running shoes, mattresses, electric cars and skincare, with 4 questions each, ranging from a general "best X" question to narrower use cases, budget or free options and specific needs. None names a brand. Each question was written in English, Spanish, German, French and Brazilian Portuguese with natural local phrasing rather than a word-for-word translation, which gives 400 prompts. Where a question only made sense locally, such as a price in local currency or which foreign language to learn, the wording follows the local version.
Models and setting. Every answer came from the Claude.ai web app: Claude Sonnet 5.5 and Claude Haiku 5.5, a new incognito chat for every answer, no memory or previous conversation, web search switched on, and the country set to match the language. Claude decided on its own whether to search. Claude Opus could not be included.
Collection. Each prompt was asked 18 times per model, in 18 complete rounds. Every round covered all 400 prompts on both models in a shuffled order, so each question's 18 answers are spread across the collection window rather than asked back to back. Collection ran from 31 October 2026 at 17:14 to 1 November 2026 at 20:09 (Europe/Madrid). An unfinished 19th round was discarded. The result is 14,400 answers, 7,200 per model, and every answer reports the model we asked for.
Brands. We extracted the brands each answer offered as options with a language model, in order of first mention, and normalised the names for case, accents and punctuation. A brand counts once per answer. Sites cited only as references are not brands. We checked a sample of the extracted lists against the answers by hand. Extraction occasionally picks up a non-brand, such as a university named through an online course, which is why we also report the number of brands named at least twice.
Measures.
- Brands across 18 answers: the union of brands over the 18 answers to one question from one model.
- Identical set: two answers to the same question name exactly the same brands, order ignored, over all 153 pairs of answers per question.
- Jaccard index: the brands (or domains) two answers share, divided by all the brands (or domains) either one names.
- First brand: the first brand mentioned in the answer text.
- Core brand: named in at least 9 of the 18 answers.
Sources. "Searched" means the model ran at least one web search in that answer. "Returned" means a page its search returned to it; we cannot tell whether the model used a given page. "Cited" means a source link shown in the answer. Domains are counted once per answer.
Limits. This was a one-off research experiment, not an ongoing measurement: one two-day window and one version of each model, and Claude changes over time. The questions lean commercial by design and cover 20 categories, not everything people ask Claude. Results describe the Claude.ai app. Brand extraction by a language model is accurate but not perfect.
About LLM Pulse #
LLM Pulse tracks how brands appear in AI answers across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, AI Mode and other models: visibility, position, sentiment and the sources behind each answer, measured across repeated answers. See how it works for Claude tracking.
Frequently asked questions #
Does Claude give the same answer to the same question?
No. Across 18 identical buying questions, Claude Sonnet 5.5 named a median of 15 different brands where one answer named 6. Two answers to the same question named exactly the same brands 11.9% of the time.
How many times should I ask Claude a question to measure my brand's visibility?
In this study, Sonnet 5.5 needed a median of 11 answers to surface 90% of the brands that 18 answers named. One answer is not enough to judge visibility.
Does Claude always name the same brand first?
Not often. Sonnet 5.5 kept the same first brand in all 18 answers for 32.5% of questions, and Haiku 5.5 for 24.8%. The usual leader still came first in about three out of four answers.
Does Claude Haiku cite sources?
Rarely in this study. Haiku 5.5 searched the web in 61.0% of answers, but only 19.0% of its answers that searched showed a source link, against 97.8% for Sonnet 5.5.
When does Claude search the web?
It depends on the category. Sonnet 5.5 searched in every answer about smartphones, earbuds and robot vacuums, and in 1.4% of answers about skincare and language apps.
Does Claude cite Reddit or Wikipedia?
Not in this study. None of the 15,937 sources cited by Sonnet 5.5 and Haiku 5.5 across 14,400 buying answers came from Reddit, Wikipedia or YouTube.
Does the language of the question change Claude's recommendations?
Yes. Asked in five languages from five countries, only 22.0% of the brands Sonnet 5.5 named in at least half of the answers to a question did the same in all five.
Discover your brand's visibility in AI search effortlessly