TL;DR
- Treat a citation as a route to evidence, not evidence that the attached sentence is true. Verify the company, the supporting passage and the date before using a claim in an account decision.
- Use Exa when your workflow needs search results plus retrievable text and highlights. Tavily is useful for a separate search-and-extract process. Linkup fits structured research with sources. None removes the need to check whether the source actually supports the inference.
- Separate facts from interpretations. A new office is observable; “the company needs our product” is a judgement requiring more context.
- Spend human attention on consequential contradictions, stale facts and ambiguous entities. A failed verification should remove or soften the claim, rather than generate another confident paragraph.
A source URL is the start of the check
An account brief can contain five linked claims and still mislead a seller. A source may refer to a different subsidiary, report an old announcement as a current event or support a weaker statement than the model writes. The problem is not solved by requiring a URL at the end of every paragraph.
Verification asks a narrower question: does this evidence support this claim about this account at the time we intend to act? It is a different job from finding relevant pages or tracing the agent's tool calls. For retrieval procurement, start with our search API guide. Here the purchase is the evidence workflow after retrieval.
Use three statuses: supported, qualified and unsupported. “Qualified” is useful when the evidence supports a limited conclusion. A company announcement can support “the company announced a new office”; it may not establish that the office has opened, how many people work there or whether a budget exists for your offer.
Avoid collapsing those states into a numerical confidence score. A score cannot explain whether uncertainty concerns the date, entity or interpretation. Sellers need the reason for a qualification, especially when the claim will appear in a message to a named prospect.
Choose the retrieval path around the evidence
Exa documents search responses containing URLs, publication dates where available, text, highlights and structured grounding fields. That makes it useful when the verification service should work directly with passages rather than search snippets. Its search reference describes the response contract; the presence of a grounding field is not an independently measured truth guarantee.
Exa's current pricing distinguishes search modes and content retrieval. Auto Search is listed at $7 per 1,000 requests for up to ten results; Contents is $1 per 1,000 pages per content type. An operator checking a known URL can use content retrieval without purchasing another discovery search. Additional results and summaries can alter the bill. Choose it for an evidence pipeline that benefits from highlights and explicit retrieval control; don't buy a deeper search mode simply to make an already-known source look more authoritative.
Tavily separates search from extraction. Its Extract reference returns extracted content and failed URLs, making failure handling a first-class part of the job. This fits teams that discover a source with one tool and want to fetch a shortlist of known pages afterwards. Check the actual extracted text: the correct URL with an empty or incomplete body is not a supported claim.
Tavily's credit documentation describes a monthly free allowance and different consumption by operation. Price the depth and operation you will use, rather than counting every URL as an identical search. It is a sensible first evaluation for an existing Tavily user; it is less compelling if its extraction repeatedly misses the document format that contains your decisive evidence.
Linkup offers web search for grounded answers and structured information through its documented API. It is worth evaluating when the desired result is a structured research answer rather than a bag of pages. Retain the sources separately from the generated answer so that a reviewer can inspect the claim without rerunning the whole research task. Budget its chosen search depth from the current offer; this article does not assign an unverified flat price to every Linkup task.
For difficult HTML, our account-research scraping guide covers the collection layer. A stronger extractor can recover evidence, but it cannot fix an inference that the evidence never supported.
The five-part claim record
Store the smallest useful unit: one proposition with its own evidence. A paragraph containing a funding round, headcount and expansion interpretation should become three records.
| Field | What to retain | Failure it exposes |
|---|---|---|
| Account identity | Domain, legal/trading name and relevant business unit | Evidence about a similarly named company |
| Proposition | The exact factual sentence or clearly labelled interpretation | A citation that supports only part of a paragraph |
| Evidence | Source URL, short supporting passage and document location | A link that cannot be reconciled with the claim |
| Time | Publication/event date when available, plus retrieval time | Old evidence presented as a new trigger |
| Decision | Supported, qualified or unsupported, with reason | A confidence score with no usable explanation |
Add source type and independence where those change interpretation. Two syndicated copies of one announcement are one originating claim, not two independent confirmations. A company's own release is strong evidence that it made an announcement; it is weaker evidence of achieved results in that announcement.
Preserve a lawful snapshot or hash when your use rights and retention policy permit it. If full content cannot be retained, store an allowed excerpt, location and retrieval details. Don't turn a verification archive into an uncontrolled copy of licensed articles or personal data.
Use the ABM Account Scoring Template to connect evidence to a qualification rule. Keep “source supports fact” separate from “fact earns points”: the first is verification, the second is your commercial policy.
Work through an expansion claim
Suppose the research output says: “Northstar is opening a German office and hiring twenty salespeople, so it is entering a buying window for CRM implementation.” This is an illustrative account, not a finding from a live run.
The linked announcement says Northstar plans a Berlin office next year. A careers page has four German sales vacancies. The CRM conclusion has no direct source. Split the sentence before trying to validate it.
The office claim becomes: “Northstar announced plans for a Berlin office next year.” Confirm that the release concerns the target group or subsidiary. The hiring claim becomes: “Four German sales vacancies were visible when checked.” Four vacancies do not prove twenty planned hires, and a listing may remain online after recruitment ends. The buying-window sentence becomes an interpretation: “The expansion may create a need to review regional sales operations.” It should not automatically qualify the account for an expensive campaign.
The revised brief is less dramatic and more useful. A seller can ask how the team is setting up regional operations without claiming to know a procurement project. If the office announcement predates a cancellation notice, the brief should include the contradiction and stop using expansion as a trigger.
The same process applies to funding, technology adoption and headcount. A provider's company match or a model's confident answer is an input to the check, not permission to merge all the evidence into one story.
Cost a hundred account briefs
Consider 100 briefs containing five consequential claims each: 500 claim checks. Assume 500 known-page text fetches and 100 extra fetches for failed or stale evidence. At Exa's listed Contents rate of $1 per 1,000 pages for one content type, 600 fetches cost $0.60. If 100 claims also need a fresh Auto Search, the listed $7 per 1,000 request rate adds $0.70, for $1.30 of those retrieval operations.
This is an illustrative workload using observed published rates, not a complete provider quote. It excludes additional content types, extra results, summaries, model calls, storage, taxes and any premium data sources. Nor does it assert that those calls will resolve every claim.
Now assume model checking costs $8, storage and orchestration $5, and 50 exceptions require three minutes each at a loaded operator rate of $60 an hour. Exception review is 150 minutes, or $150. Total assumed operating cost is $164.30. If 90 briefs meet the declared acceptance rules, cost per accepted brief is $1.83, rounded.
Those 90 accepted briefs are a scenario assumption, not a measured yield. At 100 exceptions, review becomes $300 and total $314.30. The main economic lever is the ambiguity rate and review time, not the fraction of a cent between two page-fetch tariffs. A retrieval vendor earns its place when it supplies inspectable evidence that reduces avoidable rework in your account sample.
Decide which failures stop the workflow
Fail closed on account identity and consequential contradictions. A brief about the wrong Northstar should not be sent to a seller with a caution symbol. An outdated expansion trigger should not continue to generate personalised outreach while an exception waits in a queue.
Less consequential gaps can produce a shorter brief. Missing an employee count does not necessarily prevent a useful account summary if the buying policy does not depend on size. Mark the field unavailable, rather than filling it from a vague snippet.
Keep unsupported claims out of downstream actions. The action layer should receive accepted propositions and their qualifications, not the original generated paragraph. Preserve enough provenance to answer “why did we contact this account?” later. Our pre-call account research guide explains how to deliver that evidence to the seller without burying the useful call context.
Start with a verification record and exception policy before purchasing a second research platform. Add a different retrieval provider when the first has a demonstrable evidence gap in the documents you need. Paying twice for similar search results is a poor substitute for a clear acceptance rule.
FAQ
Does a cited AI answer count as verified research?
It counts as research with a route to evidence. Verification still requires checking the relevant passage, target entity and time. A citation can be accurate while the attached conclusion is too strong. Retain the supported version and label the commercial interpretation separately.
Must every claim receive a human check?
Not every low-impact field needs the same treatment. Automate clear support checks where the source and entity are stable, then send contradictions and consequential ambiguity to a person. For a new workflow, inspect a representative sample of automated passes as well as failures; otherwise you only learn about the errors the system already recognises.
Can we use search snippets without fetching the source?
A snippet is useful for deciding what to retrieve. It is a weak final evidence record because it can omit qualifications, dates and context. For a fact that will drive an account score or appear in outreach, retrieve the original source or remove the unsupported precision.
Should a failed source check discard the whole account?
Only when the missing fact is necessary to the account's acceptance rule. An ambiguous legal identity should stop the match. An unavailable secondary detail can leave a narrower usable brief. Define the rule before the workflow runs so that the model cannot decide ad hoc which uncertainties matter.
