AWS turned web research into one tool call. Audit the answer, not just the citations

THE SIGNAL IN ONE SENTENCE

A citation is an address. It is not a witness, a guarantee or a tiny academic force field around the sentence next to it. Amazon Web Services made that distinction more important on November 5 when it announced general availability for Web Answer on Amazon Bedrock AgentCore. The product takes a natural-language question and returns a synthesized answer with source links through a single tool call. AWS says Web Answer plans several searches, runs them in parallel, reads key pages, searches again when the evidence looks thin and attaches sources to the claims it produces. The result returns through an AgentCore Gateway as a Model Context Protocol tool response. The answer appears as markdown with inline citation markers, while structured output carries the answer and a list of source URLs. That removes a great deal of plumbing. A team no longer has to assemble separate search, retrieval, synthesis and citation components just to let an agent answer a current web question. Web Answer can sit behind an AgentCore Gateway, use IAM or JSON Web Token authentication and be discovered by an MCP-compatible client. AWS says it works with Strands Agents, LangChain, LangGraph, CrewAI and other MCP clients. It is available in US East, Northern Virginia, with more regions planned. Pricing is 10 dollars per 1,000 Web Answer calls, plus billable Gateway use. That is the product story. The editorial story starts one click later. An answer can contain a citation and still be wrong. The linked page may be stale. It may describe a different country, product version or population. It may repeat somebody else's claim without showing the underlying evidence. The page may support half of a sentence while the answer quietly stretches it across the other half. The link may point to a respected institution whose actual page says nothing useful about the claim. The URL can be real while the relationship is imaginary. AWS describes Web Answer as grounded, current and source-cited. It says each grounded claim maps to a source URL where one is available. The words “where one is available” deserve a yellow highlighter. Some claims may not carry a source. Some sources may not be accessible later. Some pages may change. A current search can retrieve a page published years ago, and a page with today's date can summarize evidence gathered decades earlier. This does not make the tool useless. It makes citation auditing part of using it responsibly. The Plain Signal's practical answer is a claim audit. Start by preserving the question. Record the exact wording, time, country setting, published-date range, target-level domain rules and request-level filters. A small change in the question can change the search plan, sources and answer. If the result matters, “we asked about capital requirements” is not a reproducible record. Preserve the runtime context too. Record the AgentCore Gateway, target configuration, connector version, region, caller identity, tool name and latency. Keep the raw MCP result, including the markdown answer, structured citations and error state. Do not reduce the evidence to a screenshot of the pleasant final paragraph. Then split the answer into claims. One sentence can contain a date, number, causal explanation and recommendation. Those are four claims, even if the sentence carries one citation marker. Audit the pieces separately. For each claim, open the linked page. Find the exact passage that supports the statement. Record the page title, publisher, author when available, publication date, last-updated date and the relevant excerpt or section locator. If the source is a PDF, store the page. If it is a living webpage, preserve an approved snapshot or archive reference when policy allows. Now ask what kind of source it is. A regulation is not the same as a consultancy's summary of the regulation. A research paper is not the same as a press release about the paper. A company's launch post can verify what the company announced, but it cannot independently prove that the product performs as advertised. Primary sources should carry the factual backbone when they exist. Authority is also local. A central bank can be authoritative for its own rules and still be irrelevant to a neighboring jurisdiction. A national health agency can publish excellent guidance for one population that does not automatically transfer to another. An old product manual can come from the correct vendor and still describe the wrong release. Check scope, date and version beside the publisher's name. Next, compare the claim with the source language. Watch for certainty inflation. “May improve” is not “improves.” “Associated with” is not “caused.” “The company reports” is not “independent tests found.” “Available in one region” is not “available globally.” A synthesized answer can make several cautious source statements sound like one clean conclusion. Mark every unsupported extension. If the page supports the number but not the explanation, split the result. If two sources disagree, show the disagreement. If the evidence is missing, write “not established” instead of asking the model to fill the hole with professional-sounding grout. Domain filters can improve the search boundary. AWS lets a Gateway administrator set target-level include and exclude lists that apply to every call. A calling agent can add narrower domain limits or exclude more sites, but it cannot add a domain blocked by the administrator's target configuration. Request-level filters can also specify a publication-date range and country. That hierarchy is useful. A financial-research agent can be restricted to regulators, standards bodies and approved market sources. A medical workflow can exclude forums and content farms. A legal workflow can limit the search to courts, legislatures and official guidance for the correct jurisdiction. But a clean allowlist can still contain stale, promotional or incomplete pages. Domain control is a boundary, not a truth machine. The audit should also compare the finished answer with raw search evidence. Run a small independent search for the central claim. Look for a missing primary document, a newer version, a correction, a contrary result or an official page that ranks lower than a polished explainer. The goal is not to rebuild the entire internet. It is to detect whether the synthesis has hidden the most important disagreement. Sampling should follow risk. For a casual travel idea, opening one or two representative links may be enough. For a pricing page, verify every number and effective date. For legal, medical, financial, security or safety decisions, require claim-level review by a qualified person and do not let a cited answer become the final authority. Track coverage as well as accuracy. Count how many material claims have direct support, indirect support, conflicting evidence or no evidence. A response with eight links can have poor coverage if all eight support the same background paragraph while the recommendation at the end floats free. Record source diversity carefully. Ten sites can repeat one wire story, one press release or one flawed study. That is one evidence lineage wearing ten hats. Preserve the upstream source when you can identify it. The answer also needs a freshness label. AWS says Web Answer runs new searches for each question and gives its model the current date. That can help the search process look for recent information. It does not freeze the returned pages or guarantee that the most recent source is correct. Record when the question was asked, when each source was published and when the claim was last verified. If a decision remains active, set an expiration or review date. Privacy and access controls belong in the receipt too. AWS says Gateway calls can be authenticated with IAM or JWT and describes the service as keeping data inside its environment. Customers still need to decide what questions may include confidential information, which identities can invoke the tool, what logs and content are retained, and whether the selected region and source set meet internal requirements. Do not put a customer secret into a search question because the Gateway has a nice lock icon. Close the audit with a decision. Accept the claim, narrow it, qualify it, replace its source, mark it unresolved or remove it. Record who reviewed the evidence and which answer version they reviewed. If the source changes later, the record should show which page supported the original decision. Web Answer can make research dramatically easier to add to an agent. That is useful. It can also make a polished synthesis appear before anyone has inspected the joins between claims and sources. That is where judgment belongs. Do not count the citations. Follow them. Open the page. Find the passage. Check the date, authority, scope and disagreement. Then decide whether the answer earned the confidence its formatting already borrowed.