Last updated: October 28, 2026
Claude reaches for the web far more often than you might expect. In a study posted in September 2026, researchers sent the same 1,000 real user prompts through four AI models' APIs. Claude called its web search tool on 825 of them (83%), the most of the four. The most cautious model searched on just 140 (14%), and the other two fell in between. The models tested were versions released between February and May 2026.
My verdict: for a B2B SaaS brand, Claude is the assistant where your own website matters most, so win it with your own pages. Claude searches often and quotes short passages. Its own documentation says it searches when a question involves specific products or prices that may have changed. That describes most software buying questions. What earns the citation is a page on your domain with self-contained, dated, specific sentences Claude can lift without rewriting.
A caveat on that opening number before you build a strategy on it. The researchers called the models through the developer API, not the Claude app your buyers use, and they don't say exactly when they ran the prompts. The prompts came from real users' everyday chats, and the paper doesn't single out software buying questions. Treat 83% as a sign that Claude is search-hungry. It doesn't tell you what share of your buyers' chats will reach your site.
This guide covers Claude only. If you also care about OpenAI's assistant, read the guide on how to get recommended by ChatGPT, because the retrieval works differently there. For how answer engines fit together, start with the AEO pillar guide.
When Claude searches and when it answers from memory
Anthropic's web search tool documentation is unusually direct about this. Claude searches when a request depends on information that is current, changing or outside its training data. The examples it lists include current prices and statistics, and information about specific organizations or products that might have changed.
It answers from memory for stable knowledge: established facts, math, coding concepts, creative writing and anything already in the conversation.
Map that onto the prompts your buyers type:
"What does [your product] cost for 50 seats?" Prices change, so Claude will usually search.
"Best SOC 2 automation tools for a Series A startup" names a category of products, so expect a search.
"Does [your product] integrate with NetSuite?" is a product-specific fact that might have changed, so it's another likely search.
"What is SOC 2?" is stable knowledge. Claude can answer from training, and your blog post on the topic may never be read.
The same documentation gives a sense of scale. Simple factual queries typically use one to three searches. Comparative research across several companies can use ten or more. A shortlist prompt is exactly that kind of multi-entity question, which means Claude will read several sources before it names anyone.
Two settings change this behaviour. First, on Team and Enterprise plans, an Owner must turn web search on for the whole workspace in Organization settings. If a buyer's company never did, Claude answers that buyer from training data alone. Second, Anthropic's help page says that in the new Claude experience there's no web search toggle at all. Claude simply searches when it helps. Whether your prospects see the old or the new behaviour varies by plan and rollout, so don't assume either.
The memory case still matters. When Claude doesn't search, it describes you from what it learned in training. You can't edit that directly. The only fix is consistent, accurate public information over time, which brings you back to the same pages.
Write for span-level citation
When Claude cites a search result, the citation object includes a field called cited_text, which holds "up to 150 characters of the cited content", along with the URL and page title. So Claude quotes a sentence or two at a time.
That has practical consequences for how you write product pages:
Put the product name in the sentence. "It supports SAML" is useless as a 150-character quote. "[Product] supports SAML SSO on the [plan name] plan" stands on its own.
Use numbers. A sentence with a real user limit or a real per-seat price is more quotable than "scales with your team".
Carry a date where facts change. "As of [month and year], [Product] integrates with [number] identity providers" tells Claude, and the reader, how fresh the fact is.
Keep one fact per sentence. A sentence that packs pricing, limits and integrations together is hard to quote cleanly.
Avoid pronouns that point to the previous paragraph. A quoted span loses its context.
There's a freshness angle too. Each search result Claude receives includes a page_age field, described as when the site was last updated. Anthropic doesn't publish how heavily Claude weighs it. Still, if the page says 2024 and the competitor's says this month, you're handing Claude a reason to prefer the competitor.
Recent tool versions add dynamic filtering. Claude writes code to filter search results so that only relevant content reaches its context. You can read that as another reason to keep facts tightly worded and near the top of the page. Long, padded intros are more likely to be filtered away before the useful sentence appears.
Structured data and an llms.txt file are optional extras. The guides on schema markup for AI search and llms.txt cover what they can and can't do.
The pages Claude is most likely to quote for software prompts
For B2B software prompts, the pages that hold the facts Claude searches for are almost always on your own domain. These are the four to get right first.
Product and feature pages. Each core feature should have a plain-language paragraph that names the product, says what it does and states the limits. If the only description of a feature lives in a demo video, there's nothing for Claude to quote.
Documentation. Docs pages tend to be specific, factual and current, which is exactly what a 150-character span needs. Keep docs public and crawlable where you can. A setup guide behind a login is invisible to search.
Comparison pages. When a buyer asks Claude to compare you with a rival, an honest comparison page on your site gives it a source that covers both. The guide to comparison pages shows how to write one that's fair enough to be trusted.
Pricing facts. If you publish prices, state them in text, not only inside a JavaScript pricing toggle. Give the plan name, the price, the billing unit and what changes between tiers. If you don't publish prices, say what pricing depends on, so Claude has something accurate to quote instead of guessing.
Whether Claude can reach these pages at all depends partly on your robots.txt. Anthropic's crawler documentation separates ClaudeBot, which collects training data, from Claude-SearchBot and Claude-User, which support search and user requests. It says blocking the latter two may reduce your site's visibility in answers. A separate guide to AI crawlers and robots.txt covers the full bot-by-bot setup.
Research mode and connected workspaces
Claude's Research feature, available on paid plans, changes the shape of the problem. Instead of one or two searches, Research runs many searches that build on each other. Anthropic's help center describes it as handling five or more tool calls over one to three minutes, and lists comparing business competitors as a use case.
Research can also search connected tools such as Gmail, Google Calendar and Google Docs alongside the web. Picture a buyer asking Claude to compare three vendors against the requirements doc in their Drive. Claude reads that doc, then searches the web for each vendor's facts.
Three things follow for you:
Depth gets rewarded. A multi-step investigation will open your security page, your integrations list and your pricing page in turn. Thin pages break the chain.
Your facts get checked against the buyer's own requirements. If their doc says "must support data residency in the EU" and your site never says where data is stored, you lose that line item by default.
Contradictions get noticed. Research reads multiple sources and cross-references them. If your pricing page and an old press release disagree, expect Claude to mention the conflict or pick one.
Research on mobile searches only the web, according to the same help page. Research also needs web search switched on to work.
A small Claude prompt panel to run every month
You don't need a tracking platform to start. A short panel of prompts, run by hand once a month, tells you most of what you need.
Pick around 20 prompts in four groups:
Category prompts: "best [category] tools for [buyer type]".
Comparison prompts: "[you] vs [main rival]" and "alternatives to [rival]".
Fact prompts: pricing, key integrations, security certifications, limits.
Problem prompts: the job your buyer is trying to get done, without naming a product.
For each prompt, record four things: whether Claude searched, whether you were mentioned, which of your URLs it cited and which sentence it quoted. The quoted sentence is the useful part. It shows you which spans work, and which facts Claude got wrong.
Run the panel in a fresh chat each time, on the plan your buyers most likely use. If you sell to enterprises, also test in a Team workspace where web search is enabled, because the defaults differ. The guide to measuring AI search visibility covers the full setup, including how to keep results comparable month to month. When the spreadsheet outgrows itself, the roundup of AI visibility tracking tools compares the options.
What to do first, by situation
If Claude doesn't mention you at all, start with category and fact pages. Make sure each core product page has three or four quotable sentences that name you, state a number and give a date. Then check that Claude-SearchBot and Claude-User aren't blocked.
If Claude mentions you but cites a competitor or a review site for your facts, your own pages aren't answering the question clearly enough. Find the sentence Claude quoted from the other source. Write a better, more specific version on your own page.
If Claude gets your pricing or features wrong, look for the outdated source. It's often an old blog post, a stale docs page or a press release that still ranks. Update or redirect it, and add a dated statement of the current fact on the canonical page.
If you sell mostly to large enterprises, assume many buyers are in workspaces where an admin controls search. Some will get answers from Claude's training data alone. That makes long-standing, consistent public information more important, and because it's a slow fix, it's worth starting this quarter.
If you're a new company with little coverage, Claude's training data probably knows little about you. You depend on search. Make your docs and pricing public, keep them current and make the prompts that matter most trigger a search by being about specifics.
What to watch out for
The research is thin. Apart from Anthropic's own documentation, there's very little independent, recent work on how Claude chooses citations. The study in the intro tested the API, not the consumer app. One of its authors works at Microsoft, which also builds a rival assistant. The authors also found that Claude's heavy searching had at best mixed effects on answer quality.
Model changes move the goalposts. Anthropic's help center now lists many more models than the ones in that study. A new model can search more or less often, and your panel results may jump without any change on your side. Note the model name with every panel run.
Don't write for the quote and forget the reader. Pages stuffed with short, keyword-heavy sentences read badly and can drift into the kind of scaled, low-value content that search engines penalise. Write the specific sentence a good sales engineer would say, then stop.
Beware of vendor statistics about Claude's citation habits. Several AI visibility tool companies publish confident numbers about which sites Claude favours. Most come from the vendor's own customer prompts. Treat them as marketing until someone independent reproduces them.
Don't block the search crawlers by accident. A blanket rule written to stop training crawlers can also stop the agents that fetch pages for answers. Check what your robots.txt actually allows.
Final thoughts
Claude rewards the boring parts of a SaaS website: accurate product pages, public docs, honest comparisons and clear pricing facts. Write each important fact as a short sentence that names the product, gives a number and carries a date. Then check monthly whether Claude is quoting it. The same page-level work carries over to LLM SEO for the other assistants.