An agent finds a company name on a page about a new distribution center. It adds the company to a prospect list, cites the page, and writes that the company is expanding its warehouse operations.
Open the source and the problem becomes clear: the company supplied equipment to the facility. It does not operate it.
The citation is real. The qualification is wrong. This is the gap between finding text that mentions a subject and establishing a claim about that subject. A sales research system has to cross that gap explicitly.
Bind the Claim to the Right Subject
Consider a fictional announcement involving a retailer, a property developer, an equipment supplier and a logistics operator. All four names occur near “new distribution center.” Only some of them fit an offer for warehouse operators.
A source URL cannot express those relationships. The research needs the passage that connects the selected company to the claimed role. It also needs to distinguish the company's brand, legal entity and parent where that distinction changes the buying decision.
Fama's development research code includes logic for locating a quoted passage within the retrieved source and retaining nearby context, including relevant headings. It also narrows the text used to connect a person's identity to a profile reference. These are inspected implementation mechanisms; they are not proof that the system resolves every ambiguous page correctly.
They exist to prevent a specific kind of error. A name in one table row should not inherit a role, link or fact from the neighboring row. A speaker from last year's event should not become a confirmed participant in this year's event because both appear on the same site.
Ask What the Passage Actually Establishes
After resolving the subject, examine the claim's strength.
“Plans to open a facility” differs from “has opened a facility.” “Supports clients in Europe” differs from “employs a European operations team.” A person speaking about procurement does not automatically own a purchasing budget.
For the distribution-center example, the research record could retain three separate fields: the named entity, its stated relationship to the facility, and the event's status. Qualification then uses those fields against the seller's brief.
If the offer requires an operating facility, a planned opening may warrant monitoring rather than immediate contact. If the offer serves equipment suppliers, the originally rejected company may become relevant under a different brief. Evidence does not have a universal sales meaning; it supports a particular decision under stated criteria.
Count Origins, Not Links
Suppose the announcement appears on the company's site, a wire service and three trade publications. Five links may represent one originating statement.
That distinction matters when a workflow tries to increase confidence by asking for two sources. If it counts domains alone, syndicated material can satisfy the rule without adding independent reporting.
The development implementation in Fama records source-lineage signals and includes synthetic tests for duplicated wording, wire-service attribution and credited origins. The inspected tests express an intended behavior: repeated versions of one originating document should not automatically become independent corroboration. They were read for this article, not executed.
Lineage detection has limits. Different wording does not prove independent reporting; similar wording does not prove dependence in every case. A source count should therefore remain an explanation of the available evidence, not a mathematical guarantee of truth.
A buyer should ask whether the second source contributes new support or merely repeats the first.
Carry Uncertainty Through to the Answer
Research systems can be cautious internally and overconfident in their final prose. A row says “reported,” but the summary says the company definitely meets every criterion. A question asks the user to accept an assumption, and the final paragraph presents that assumption as externally verified.
Fama's development answer format separates facts, assumptions, derived values, recommendations and gaps. It checks references against stored rows and claims. Those structural checks are intended to make it harder for unsupported prose to disappear into a polished summary, but they do not provide complete semantic verification of every sentence.
That limit is important. A sentence can contain the right number while making the wrong claim about it. Matching a value is useful validation; understanding the whole sentence still requires evaluation.
Keep the evidence label visible at the point of use. A salesperson deciding whether to ask a clarifying question needs to know that a role is reported by one source rather than confirmed by the person's employer. Hiding that distinction behind a generic confidence score removes information the seller could use.
Recheck What Depends on a Changed Source
Imagine the operator later corrects its announcement: the warehouse is a partner facility, not a directly managed site. A new source fetch should do more than update one sentence in storage.
The affected qualification, proposed message and reusable account fact all deserve reconsideration. Preserve the earlier source version so the team can understand why the old decision was made, while preventing it from silently supporting a new one.
This is the reason to retain the relationship between a claim and its evidence. Provenance is useful both when the claim is accepted and when it must be withdrawn.
Review One Record End to End
When evaluating prospect research, pick a borderline account and follow the entire chain: selected entity, retrieved passage, claim, qualification rule, uncertainty, recommendation. Introduce a correction and inspect what changes.
Do not settle for seeing a citation icon beside the company name. The record should explain why that source supports that statement about that entity under the current brief.
The standard is the same whoever did the research. A list built by hand should meet it, and so should anything an agent returns.