Gumshoe.Open a case

The standard · a public benchmark

The 5 tests every
research API should pass.

Every research API demo looks identical: point it at a company, get a tidy summary. The differences only show up on the cases that break things. Here are five reproducible tests, our measured results, and the failure modes we saw. Run them on whatever you use today. We name no competitor: you will find out soon enough.

01

The homepage test

READ A COMPANY SITE AND CLASSIFY IT AGAINST A DEFINITION.

This is the floor, not a differentiator. Every serious tool passes it. If a vendor makes this their headline, they are selling you table stakes.

PROTOCOL

Point the tool at 5 sites: a pure-play agent company (sierra.ai), three products that added agent features on top (attio.com, notion.so, clay.com), and one with none (basecamp.com). Ask: "does this company sell AI agents as its primary product? Apply the removal test: if you deleted every agent capability, would the product still exist?"

OUR RESULT

5 out of 5. Sierra core, the three bolt-ons correctly demoted to watchlist or exclude, Basecamp excluded.

1 credit · ~11 seconds · reads the page and quotes it back

WHAT TO WATCH FOR

Most tools land 4 out of 5. The usual miss is calling a generic SaaS a watchlist instead of excluding it, because the prompt is doing the work, not the engine.

02

The careers test

FOLLOW FROM THE HOMEPAGE TO AN INTERNAL PAGE AND EXTRACT STRUCTURED FACTS.

A hiring signal lives on a careers page, not the homepage. This checks whether the tool can navigate a site rather than read one URL.

PROTOCOL

Ask: "are they hiring engineers right now? List the exact open role titles." Give only the root domain, not the careers URL.

OUR RESULT

Passed. Exact titles returned, including Engineering Manager Agent and Software Engineer Agent, with locations.

1 credit · ~7 seconds

WHAT TO WATCH FOR

Most tools pass this too. Multi-page reading is now standard. Do not pay a premium for it.

03

The interaction test

GET DATA OUT OF A PAGE THAT REQUIRES CLICKING, SEARCHING, OR FILTERING.

This is where the field splits in two. Scrapers and cached readers return the page title and nothing else, because the content only exists after interaction.

PROTOCOL

Point the tool at a JavaScript directory such as ycombinator.com/companies and ask it to search for a category and return 5 results with names, descriptions, and websites.

OUR RESULT

Passed. A real cloud browser typed in the search box, applied a batch filter, noticed the filter returned nothing, corrected itself, then opened 5 company profiles. 13 steps, screenshots at each one.

2 credits · ~55 seconds · screenshots kept as exhibits

WHAT TO WATCH FOR

This is the most common failure. A typical answer reads: "the provided content only includes the main page title. It does not contain any actual directory listings." Our own plain reader returns 80 characters on that page. Without a real browser, nobody passes this.

04

The off-site test

ANSWER A QUESTION WHOSE ANSWER IS NOT ON THE COMPANY SITE.

Funding, press, competitor mentions and reviews live elsewhere. A tool anchored to the target domain cannot reach them, no matter what its docs claim.

PROTOCOL

Ask: "did this company raise funding recently? Amount, round, date, with sources." Pick a company with a well covered raise that is absent from its own homepage.

OUR RESULT

Passed. Returned 950 million dollars, Series E, dated May 2026, with the investor list and four sources.

part of a deep research dossier · 15 credits · ~19 seconds

WHAT TO WATCH FOR

Site-anchored engines answer "no information about any recent funding" and cite the company site as proof of absence. Check whether the tool actually searched, or just read the domain again.

05

The honesty test

WHAT HAPPENS WHEN THE SITE IS DOWN, GATED, OR EMPTY.

The dangerous failure is not an error. It is a confident answer built on nothing, which then poisons your CRM at scale.

PROTOCOL

Point the tool at a domain that is parked, under maintenance, or unreachable. Read what comes back.

OUR RESULT

Passed. The report names the failure, states what blocked it, and refuses the verdict. On one unreachable domain we returned exclude with the reason "site unreachable, connection refused", and separately flagged that similarly named companies in search results were different entities.

1 credit · the failed read still costs 1, never more

WHAT TO WATCH FOR

Watch for two bad behaviours: a confident summary of a maintenance page, and silently pulling a cached copy from weeks ago without saying so.

Run them on us

1,000 free credits a month, no card. Every number on this page was measured, not estimated. Reproduce them.

Open your first case

Keep reading