AI company research: five prompts with an evidence audit

“Research this company and tell me whether it is a good prospect” asks AI to make several decisions at once. It must identify the company, interpret your offer, search for facts, infer fit and decide what matters. A polished paragraph can hide a mistake at any of those steps.

A better workflow asks a sequence of small, checkable questions. It keeps company identity, observed facts, uncertainty and your commercial decision separate. The result may be less fluent than a general summary, but it is easier to use responsibly.

Below are five copyable prompts, a complete example of a misleading result and a review plan. They are proposed research instructions, not a benchmark showing that a particular model achieves a given accuracy. Use them in a research assistant with source access; adapt each to the actual output contract of your tool.

What AI can contribute to company research

AI can help search for an official page, extract a specific observation and compare it with a written criterion. The useful output is a fact you can inspect and a source that supports it.

It should not silently turn missing information into a negative, estimate private facts without permission, or treat a source about a similarly named company as evidence. A source link is also insufficient if the linked page does not support the answer.

NIST's Generative AI Profile addresses risks of generative AI, including confidently presented false content. The operating rules below are our application of that concern to company research: verify identity, inspect decisive evidence and preserve uncertainty.

Before automating research, write the actual company definition. The B2B ICP template helps turn “good prospect” into an observable condition.

Prepare the input before writing the prompt

Supply the official domain if you have one. Otherwise provide enough context to distinguish the organization: name, country, city, known product or registry identifier. A company name alone is often ambiguous.

Keep three categories in separate places:

Do not put instructions embedded in retrieved pages on the same footing as your question. A page telling an assistant to ignore its task is source content, not authorization to change the research or send information elsewhere.

Ask only for information you need. A company-level research question usually does not require collecting personal contact details. Add people research later when the account is accepted and the intended use is appropriate.

Prompt 1: resolve the company identity

Use this first when names, brands, groups or subsidiaries could be confused.

Task: identify the organization described by the supplied input.
Start with the supplied domain and official pages. Confirm thecompany name, country and relationship between brand and legalor operating entity where the source makes that relationship clear.
Do not substitute a similarly named company. Do not transfer aparent group's characteristics to a subsidiary.
Return the official website and a concise explanation with sources.If identity remains ambiguous, report that uncertainty and thespecific fact needed to resolve it. Stop before qualification.

Review: Does the source concern the same entity as the input? Does an imprint or official company page explain a name difference? A search result heading alone may not resolve it.

Stop condition: two plausible entities remain. Additional commercial research would attach increasingly detailed facts to an uncertain identity.

Prompt 2: establish one business-model fact

Replace the bracketed criterion with an observable condition from your ICP. Avoid bundling size, region, technology and fit into one answer.

Task: determine whether this identified company meets this criterion:[the company manufactures its own physical products].
Use official product, production or company pages where available.Distinguish own production from distribution, consulting andcontract manufacturing. Retain the distinction described by the source.
Explain the observed fact and cite the page that supports it.Keep the result unknown if available evidence does not decidethe criterion. Do not infer production from the industry label.

Review: “We supply industrial components” does not necessarily mean “we manufacture them.” A product catalog may mix own brands and distributed brands. Inspect the exact claim.

Stop condition: the question needs non-public information. Keep the field unresolved and decide whether a human conversation would justify the effort.

Prompt 3: investigate a current hiring or expansion event

Task: check for a currently supported public event relevant to[our defined operating problem] at this identified company.
Prioritize the official careers page or original company announcement.For a vacancy, retain role, location, employing entity and currentavailability. For expansion, retain the announced location and stage.
Distinguish publication date, event date and date checked. Do nottreat a repost or an undated page as proof of a new event.
Describe the event separately from possible commercial relevance.Do not infer urgency, budget, dissatisfaction or purchase intent.

Review: Is the vacancy still available on the official page? Does the expansion concern the relevant subsidiary? A search snippet can remain visible after the underlying event has expired.

Stop condition: only old or unattributed copies remain. Record the candidate event as unverified rather than presenting it as current.

Use the buying-signal matrix to decide which kinds of event are relevant to your offer before submitting this question.

Prompt 4: test a technology claim in context

Task: investigate this specific claim about the identified company:[the company is implementing system X].
Find evidence that states the company's own relationship to thesystem. Separate current use, planned implementation, past use,customer work, partnership and desired employee experience.
Retain the date and wording necessary to explain the distinction.If the evidence supports only experience requirements or a partnerrelationship, do not report an internal implementation as confirmed.
Return the narrow fact supported by the source and what remains unknown.

Review: A consultancy hiring developers with SAP experience may implement SAP for clients. That does not prove the consultancy is replacing its own ERP. A three-year-old customer story establishes a different time frame from a current project notice.

Stop condition: indirect clues are the only evidence. They may support a follow-up question, but not a confirmed technology field.

Prompt 5: prepare a reviewable handoff

Use this only after the underlying facts have been checked. It is a summarization instruction, not permission to send messages or modify a CRM.

Task: summarize the reviewed evidence for the next human decision.
Use only the supplied, checked observations. Keep four sections:1. Confirmed company facts and their sources.2. Criteria contradicted by evidence.3. Unknowns that could change the decision.4. One recommended next research question, with its purpose.
Do not invent a contact, budget, urgency or relationship. Distinguishour interpretation from facts stated by the company. Do not startoutreach, update external records or purchase additional data.

Review: Can a colleague understand why the account is being investigated without rereading the entire research history? Does every decisive claim lead back to an actual observation?

Stop condition: the summary introduces a claim that does not exist in the checked inputs. Remove it and inspect whether the same error occurs elsewhere.

Worked example: why a plausible answer can be wrong

Consider a fictional industrial distributor called Westhaven Components. The only supplied product-page excerpt says: “We distribute precision components from international manufacturers.” Its careers page asks for “experience with SAP migration projects.” These excerpts are invented for this exercise, not live website evidence.

An unreliable summary might say: “Westhaven manufactures precision parts and is currently migrating to SAP, making it a strong prospect for production integration.” It makes three unsupported jumps:

The better result is shorter: the available excerpt describes distribution; the vacancy requests migration experience; neither confirms own production or an internal migration project. If own production is mandatory, the current evidence does not justify accepting the account.

This is why an audit should check whether the source entails the claim, not simply whether a citation exists. A real URL next to an unsupported sentence is still an unsupported sentence.

Run a small quality review before scaling

Choose a pilot that exposes likely mistakes. Include clear matches, clear exclusions, ambiguous names, subsidiaries, sparse sites, old announcements and pages where a technology appears only in a hiring requirement. Ten to twenty cases can reveal instruction problems; they cannot establish a general accuracy rate.

Before the run, write the expected fact and supporting source for cases you can adjudicate. Keep genuinely unknown cases unknown. Review results without changing the answer key to match what the model returned.

Use four outcomes:

• Supported: the answer concerns the right entity and the source establishes the claim.
• Contradicted: evidence establishes that the returned claim is wrong.
• Unsupported: a value was returned, but the evidence does not decide it.
• No answer: no usable value was returned.

Track execution failures separately. A technical failure is neither a company fact nor a completed no-answer research outcome.

Report raw counts, unsupported share, review minutes and cost. If six of ten answers are supported, two unsupported, one contradicted and one empty, publishing “90% coverage” alone hides the most important issue. Nine filled answers do not mean nine trustworthy findings.

Adapt the prompts to Leadbase's typed research

Leadbase's typed enrichment entry point accepts known input, a question and a defined output type. Ask for a URL when the task is finding an official page; use text or a permitted list of values for a classification. The prompt blocks above describe research behavior; they are not an arbitrary JSON response schema for the API.

For multi-field results, use the documented typed-list contract where appropriate. Inspect the explanation and sources separately from the returned value. An empty result must remain different from “false,” and a valid output shape does not guarantee a correct company match.

Leadbase publishes this guide. These instructions can also be used in another source-capable research tool. We do not claim that a prompt eliminates hallucination or that every question has an observable public answer.

Start with a narrow task in the research catalog, inspect a few results in Activity, and check settled costs before expanding. Do not automatically repeat an ambiguous paid submission.

Questions about AI company research

Can one prompt research everything about a company?

It can produce a broad briefing, but that makes errors harder to isolate. For a decision with a clear acceptance rule, separate identity, decisive facts and interpretation. You can summarize reviewed findings afterward.

Does requiring citations make the research reliable?

It makes inspection possible. You still need to verify identity, date and whether the page actually supports the claim. A dead link, an unrelated subsidiary or a source with a weaker statement can all undermine a cited answer.

What if public sources disagree?

Retain both observations with their dates and scope. Check whether they concern different entities or periods. Escalate the consequential disagreement rather than silently choosing the statement that makes the account look more attractive.

What should we optimize next?

Once the answers are usable, optimize the cost of obtaining them. Continue with data enrichment costs and waterfall economics, including a worked calculation that counts review work and pending billing.

Sources and product documentation checked on 17 October 2026. Prompts and fictional examples are editorial teaching material, not measured model-performance evidence.