Claude Web Search 2026: Inside the $10 Citation Engine

By Alex Bain 2026-07-30 15 min read

Anthropic charges $10 per 1,000 searches for Claude web search, and in the 16 months since it first shipped, the feature has moved from a United States only preview into the default retrieval layer of an assistant that now helps decide which brands get named in AI answers. Claude web search reached the developer API on May 7, 2025, went live on every consumer plan including the free tier by May 27, 2025, and gained dynamic filtering plus agentic response controls across February and March 2026. For B2B marketing and SEO teams, that arc is the clearest signal yet that citations, not blue links, are becoming the unit of visibility.

What Claude web search actually is in July 2026

Claude web search is Anthropic's built-in, real-time retrieval tool. When a prompt depends on information that is current or sits outside the model's training data, Claude decides on its own to run one or more searches, reads the results, and returns an answer with inline citations to the sources it used. According to Anthropic's platform documentation, citations are always enabled for web search, and each cited passage carries the source URL, the page title, and up to 150 characters of quoted text. The help center states the user benefit plainly: Claude provides direct citations so you can fact check the sources.

The tool is steerable. Developers cap the number of searches with a max_uses parameter, and Anthropic notes that simple factual queries typically use one to three searches while comparative or multi-entity research can use ten or more. Claude does not search for everything, and the documentation is specific about the split.

Claude runs a live web search when the request involves:

• Recent events, news, or product announcements
• Current prices, rates, scores, or statistics
• Details about specific organizations, people, or products that may have changed
• An explicit instruction to search or look something up
• Facts that fall past the model's training cutoff
• Anything a system prompt has told it to verify before answering

It answers directly, with no search, for established facts, math and science fundamentals, coding concepts, creative writing, and analysis of content already provided in the conversation. That split matters for marketers. The queries where Claude reaches for the live web are exactly the commercial, comparison, and best-tool-for-the-job prompts where a citation carries revenue.

From a US-only preview to a global default

Historical context matters because the rollout was fast and deliberate. Anthropic launched web search first inside the consumer Claude app on March 20, 2025 as a preview for paying users in the United States, running on Claude 3.7 Sonnet. Seven weeks later, on May 7, 2025, it opened the capability to developers through the API. TechCrunch reported the API price at $10 per 1,000 searches and support for Claude 3.7 Sonnet, Claude 3.5 Sonnet, and Claude 3.5 Haiku at launch. By May 27, 2025, web search was live on every Claude plan worldwide, including the free tier, a point InfoQ covered as the capability moved from the consumer app into developer tooling.

The 2026 upgrades turned a convenience feature into infrastructure. On February 17, 2026 Anthropic shipped web_search_20260209, adding dynamic filtering. On March 18, 2026 it added web_search_20260318 with response inclusion control for agentic workflows. Today three tool versions coexist, and the timeline below tracks how the feature matured from preview to platform.

The cadence is telling. Anthropic compressed preview, API, and global availability into roughly ten weeks in 2025, then spent 2026 optimizing cost and agent behavior rather than reach. That is the pattern of a company treating retrieval as a permanent part of the stack, not an experiment. For growth teams, the practical takeaway is that Claude has had live web access for well over a year. Long enough that its citation habits are now stable enough to study, measure, and optimize against with confidence.

Dynamic filtering rewired the token math

The most consequential 2026 change is dynamic filtering, and it is a cost story as much as an accuracy story. With basic web search, every result gets loaded into the model's context window, including navigation, boilerplate, and passages that have nothing to do with the query. With web_search_20260209 and later, Claude instead writes and runs Python inside a code execution sandbox to parse and filter the results first, so only relevant content reaches the context window. Anthropic's dynamic filtering announcement reported an average accuracy gain of 11% while using 24% fewer input tokens.

"The model behaves like an actual researcher, writing Python to parse, filter, and cross-reference results rather than reasoning over raw HTML in context."
Gareth Jones, Product and Research Lead at Quora, in Anthropic's February 17, 2026 announcement.

The benchmark movement was not marginal. On two retrieval benchmarks, dynamic filtering lifted both the mid-tier and flagship models by double digits.

Dynamic filtering is available on Claude 4.6 and later models plus the Claude Mythos Preview. Because the filtering runs inside code execution, Anthropic charges no extra fee for those code execution calls beyond standard token costs. For agentic search, where a single task can trigger ten or more queries, cutting input tokens by roughly a quarter changes the unit economics of building on Claude. It also improves answer quality, which is the part marketers should care about, because a model that cross-references sources before answering is a model that rewards well-structured, quotable source pages over keyword-stuffed filler.

What $10 per 1,000 searches really costs

Pricing looks simple and hides nuance. Web search is billed at $10 per 1,000 searches on top of standard token costs, per Anthropic's documentation. Each search counts as one use regardless of how many results come back, and searches that error out are not billed. The content Claude retrieves still flows through the context window as input tokens, both within a single turn and on later turns of the same conversation, so a search-heavy agent pays twice: once for the search request and again for the tokens that search pulls into context.

That second cost is what dynamic filtering and the March 2026 response inclusion control attack. Setting response_inclusion to excluded drops raw search blocks from the API response once a code execution call has consumed them, which lowers output token cost for agents that do not need to echo source text back to the client. Developers shape behavior with a handful of controls:

• max_uses: a hard cap on the number of searches per request
• allowed_domains: restrict retrieval to a trusted set of sites
• blocked_domains: exclude specific sources, though you cannot use both lists at once
• user_location: localize results by city, region, country, or timezone
• response_inclusion: drop consumed search blocks to cut output tokens
• allowed_callers: choose direct search or code execution for dynamic filtering

At $10 per 1,000 searches, a product that runs three searches per user query pays three cents per query in search fees before a single token is counted. That is inexpensive next to building and maintaining a crawler, which is precisely the pitch Anthropic made at launch. It is not free at scale, and it explains why the 2026 roadmap reads like a sustained effort to keep the token bill down as agents search more aggressively.

Why Claude web search matters for AEO, AIO and GEO

Here is the marketing translation. When Claude searches the live web and cites sources, the pages it names become the answer the buyer reads. Being that cited source is the entire game of answer engine optimization, AI optimization, and generative engine optimization. The generative engine optimization services market is projected to grow from about $1.48 billion in 2026 to $17.02 billion by 2034, a compound annual growth rate near 45%, which tells you how much budget is moving toward this exact problem.

Muck Rack's May 2026 edition of its ongoing What Is AI Reading study analyzed more than 25 million links from ChatGPT, Claude, and Gemini across 17 industries and found that earned media accounts for 84% of all AI citations, with journalism alone at 27%.

"Three editions in, the data keeps telling the same story: earned media is what AI trusts. If your brand is not showing up in the media coverage AI is reading, you are not showing up in the answers AI is giving."
Greg Galant, co-founder and CEO of Muck Rack, in the study's May 7, 2026 release.

Claude's citation model is unusually legible for optimizers. Because every web search response carries the source URL and up to 150 characters of the exact cited text, you can see precisely which sentence earned the citation. That is a tighter feedback loop than a ranking position on a results page. If you want to understand where you stand today, an AI visibility audit that measures how often Claude, ChatGPT, and Gemini name you is now a sensible first step rather than a nice-to-have.

How Claude cites sources vs ChatGPT, Perplexity and Gemini

Context on the competitive field first. In May 2026, ChatGPT's share of AI chatbot web traffic fell to 53.9% from 79.0% a year earlier, while Gemini reached 27.9% and Claude 9.2%, according to reporting that flagged ChatGPT slipping below 50% for the first time. By monthly active users the gap is wider, with ChatGPT near 1.11 billion, Gemini 662 million, and Claude 245 million. Reach is not citation behavior, though, and the engines diverge sharply on which sources they name.

Aggregated research spanning more than 680 million citations, packaged in the 2026 AI Platform Citation Source Index, shows Reddit as the single most cited domain across every major engine at roughly 40% frequency. Beyond that the engines split. One analysis found ChatGPT drew 47.9% of its sources from Wikipedia while Perplexity pulled 46.7% from Reddit, and only 11% of cited domains overlapped between the two. A separate study of 34,234 AI responses found a 46 times gap in brand citation rates, with ChatGPT naming brands 0.59% of the time versus Perplexity at 13.05%.

Claude sits in the middle of this spread and leans on earned media and reference sources, consistent with the Muck Rack data. The strategic implication is that a single AI SEO playbook does not exist. The domains that win a Perplexity citation, often fresh Reddit threads, are not the domains that win a ChatGPT citation, which skews to Wikipedia and established news. Claude web search, with its always-on citations and exposed cited text, is the easiest of the three to reverse engineer, which makes it a useful proving ground for a broader search visibility program before you scale tactics to the noisier engines.

The enterprise cloud land grab

Claude web search is not only a consumer and API feature. It is being wired into the major enterprise clouds, which is where the durable revenue sits. Anthropic's documentation lists web search as available on the Claude API, Claude Platform on AWS, and Microsoft Foundry, where it requires a Hosted on Anthropic deployment. On Google Cloud only the basic web search tool is available, without dynamic filtering, and web search is not offered on Amazon Bedrock at all. Those distinctions are not trivia. They decide which retrieval and citation behavior your customers experience depending on the cloud their vendor chose.

The Google Cloud integration shows how quickly this became enterprise plumbing. Google's Gemini Enterprise Agent Platform documentation lists Claude web search support across newer model families including Claude Opus 5, Claude Sonnet 5, Claude Fable 5, Claude Opus 4.8, and Claude Sonnet 4.6, and it instructs administrators to configure the constraints/vertexai.allowedPartnerModelFeatures policy to retain access to partner model features after June 2026. When a hyperscaler writes governance policy around a competitor's retrieval tool, the feature has stopped being optional.

The enterprise momentum tracks Anthropic's standing with developers. By one widely cited measure, Anthropic leads enterprise LLM API spend at roughly 40%, ahead of OpenAI near 27% and Google near 21%. Web search is part of the reason. It removes the need for teams to build and maintain their own crawling and ranking stack, and it ships with the compliance affordances enterprises demand, including zero data retention eligibility and admin controls to disable search or restrict its domains from the Claude Console. For a marketing leader, the lesson is that Claude's answers will increasingly be generated inside tools your customers already pay for, from Azure to Vertex, so the citation you earn compounds across surfaces you never sold into directly.

What the citation studies say about getting named

The uncomfortable finding across 2026 research is that most brands cannot even measure this yet. Semrush's expanded 2026 AI Visibility Index, built on 126 million AI search prompts, reported that 45% of marketing leaders cannot accurately measure their brand's visibility inside AI answers and only 9% have tools to track every relevant metric across platforms. Rand Fishkin of SparkToro, working with Gumshoe.ai, had 600 volunteers run nearly 3,000 identical prompts through ChatGPT, Claude, and Google AI, and documented how inconsistent the recommendations were from one run to the next.

The patterns that do repeat are actionable. Content built for citation tends to share the same traits:

• Original statistics and first-party data, which several studies link to 30% to 40% stronger performance in AI answers
• Structured, list-formatted content, which accounts for a majority of cited URLs in multiple analyses
• Earned media and third-party coverage, the source type behind 84% of citations in Muck Rack's data
• Clear, quotable sentences that map cleanly to Claude's 150 character cited text window
• Freshness, which Perplexity in particular rewards, with recent content cited far more often
• Presence on the reference sites engines trust, from Wikipedia to established trade press
• Consistent entity information so engines can resolve exactly who you are

The gap between what works and what teams can measure is the opportunity. If only 9% of organizations have full cross-platform tracking, the other 91% are guessing, and the teams that instrument Claude, ChatGPT, and Gemini citations now will optimize while competitors wait. In practice that means logging the exact prompts your buyers ask, capturing which sources each engine cites for each one, and treating the returned cited text as keyword-grade data worth reporting on every month.

Five predictions for Claude web search through 2027

Based on the 2025 to 2026 trajectory, here is where this is heading.

1. Dynamic filtering becomes the default caller on every new Claude model by mid 2027, and basic direct search is positioned as a legacy path, since Anthropic already defaults allowed_callers to code execution on version 20260209 and later.
2. Web search pricing holds at or near $10 per 1,000 searches through 2027 while token costs keep falling, so the flat search fee becomes the larger line item for high-volume agents and pushes teams toward aggressive filtering and caching.
3. Claude's share of AI chatbot traffic climbs past 12% by mid 2027 on the strength of enterprise cloud distribution, even as ChatGPT remains the largest single assistant.
4. Earned media stays the dominant citation source, holding near or above 80% across the major engines, which keeps digital PR and original research at the center of AI visibility budgets.
5. At least one major analytics vendor ships native Claude citation tracking, turning cited text into a reportable metric alongside rankings, because the measurement gap Semrush documented is too large to leave unfilled.

These are directional calls, not guarantees, and any of them can bend if Anthropic changes pricing or a regulator reshapes how AI systems attribute sources. The consistent through line is that retrieval and citation are converging into a single competitive surface. Claude web search is the cleanest instrument to watch that happen, because its citations are always on and its cited text is exposed in the response, which is not true of every rival.

What to do Monday morning

Concrete steps for a B2B marketing or SEO team this week, in order.

1. Run your top 20 buyer prompts through Claude with web search enabled and record every source it cites, including the exact cited text.
2. Repeat the same prompts in ChatGPT and Gemini and compare the source lists, because the winners rarely match across engines.
3. Audit whether your original data and statistics live on pages Claude can reach and quote in 150 characters or less.
4. Prioritize earned media and original research, the source types behind 84% of AI citations, over another round of thin blog posts.
5. Check that your entity information is consistent across Wikipedia, LinkedIn, Crunchbase, and your own site so engines can resolve who you are.
6. Set a monthly cadence to re-measure, because recommendations shift from run to run and month to month.

The strategic frame is simple. For fifteen years the goal was to rank. The goal now is to be the sentence an AI quotes back to a buyer. Claude web search made that measurable by exposing the URL and the cited text behind every answer, and it did so at a price, $10 per 1,000 searches, that put real-time retrieval inside products used by hundreds of millions of people. Teams that build a repeatable generative engine optimization practice around that feedback loop now will own the citations their competitors are still trying to measure. Start with the prompts, follow the citations, and improve the exact source pages the engines already name.

Alex Bain

Head of Growth

Alex leads programmatic SEO and AEO at Skitrate. Twelve years in growth, previously head of organic at a B2B SaaS unicorn and an ex-agency SEO director. Specializes in scaling content from zero to seven-figure traffic on technical and high-intent commercial verticals.

• 12 years leading SEO and growth at agencies and in-house SaaS
• Built and scaled programmatic SEO catalogs across SaaS, fintech, edtech
• Active practitioner of AEO/AIO/GEO measurement and citation tracking
• Public talks at SearchLove, BrightonSEO, MozCon

More posts by Alex Bain →