The hub for Linkup's public, reproducible benchmarks — how a web-search API performs on the retrieval jobs behind real GTM and legal work. Every number below links to a repo that ships its dataset, queries, scoring code, and results, so it can be re-run and re-checked.
Repos: Search Benchmarks on GTM Use-cases · Finance — FinSearchComp · RedlineBench × Linkup (legal)
Linkup vs Exa, Perplexity, and Parallel across the three signals that cover the people-data lifecycle: get the record right, make it useful, keep it current. Linkup leads all three — decisively on the two that matter most for triggers (correct-person and freshness).
| Signal | What it measures | Linkup | Exa | Perplexity | Parallel |
|---|---|---|---|---|---|
| Enrichment | Correct person from a LinkedIn URL (n=500) | 94% | 56% | 64% | 63% |
| Richness | Pre-meeting brief quality, /100 (n=100) | 64.8 | 55.8 | 53.1 | 59.8 |
| Freshness | Caught a just-happened job change (n=~50) | 74% | 14% | 9% | 11% |
Enrichment is checked against an independent profile DB, so the gap measures whether the record points at the right human before a rep sees it. Freshness is the highest-value trigger and the widest gap — the failure mode for others is reporting the right person at their stale employer.
Firmographics and research that drive account scoring, prioritization, and outbound. Linkup leads both live company signals.
| Signal | What it measures | Linkup | Exa | Perplexity | Parallel |
|---|---|---|---|---|---|
| Answer quality | On-target sources, 100 companies × 5 sections (n=500) | 80.7% | 76.8% | 73.4% | 71.0% |
| Specific information (example here: Funding) | Total funding within ±25% of Crunchbase (n=93) | 83% | 72% | 56% | 74% |
Answer quality tracks five distinct GTM signals per company, each fired as its own query set: what the company offers, its pain → solution angle, case studies (proof + metrics), named customers & partners, and buying-motion CTAs (demo · pricing · sign-up) — the exact fields a rep needs to research and prioritize an account. Every engine gets one search endpoint, no chaining or extract, so the score reflects retrieval across all five, not the writer.
Funding is the hard, specific case: given only a company's name, HQ, and founding year, return its total equity raised. That number is scattered across press releases, filings, and trackers — often partial, stale, or conflicting — so landing within ±25% of Crunchbase means reconciling multiple rounds into one correct total, not just surfacing a headline figure.
Financial retrieval on FinSearchComp — 635 questions written by 70+ professional financial analysts, each with a single gold answer verified against primary sources (SEC filings, exchange data) and graded strictly right/wrong. We run the two historical Global tiers identically across every system, taking each provider's best research/deep config. Linkup is 1st on both tiers.
| Tier | What it measures | Linkup | Perplexity | Parallel | Tavily | Exa |
|---|---|---|---|---|---|---|
| T2 | Simple historical lookup (n=119) | 82.4% | 72.3% | 73.1% | 40.3% | 42.0% |
| T3 | Complex multi-step investigation (n=84) | 58.3% | 56.0% | 46.4% | 36.9% | 26.2% |
T2 pulls a single fixed figure from a company's past filings — a line item like revenue, R&D expense, or inventory for a given fiscal year — so it rewards finding and reading the right primary source. T3 is the hard tier: reaching the answer takes multi-step work — pulling figures across several sources or periods and computing a ranking, ratio, maximum, or derived difference — so it tests aggregation and calculation on top of retrieval, where the field drops off fastest.
Does live web search make a frontier model a better contract negotiator? We layer a Linkup-backed search skill on top of Crosby's RedlineBench and run GPT-5.5 against its published closed-book score — same model, prompts, and 3-judge grading; the only change is the agent can look up market norms while it redlines.
| System | Score (turn-weighted) |
|---|---|
| GPT-5.5 + Linkup search (140 tasks) | 58.6% |
| GPT-5.5, published baseline (closed-book) | 50.5% |
Adding Linkup search lifts GPT-5.5 by +8.1 points, ahead of the published baseline and every other model on the board.
Shared principles across every bench:
- Each API at its own native surface — no LLM wrapper between the API and the score, so we measure the retrieval engine itself. Where a model step is needed, it's held identical across engines.
- Independent ground truth — profile DB, Crunchbase, verified company identity, or Crosby's own 3-judge panel.
- Deterministic where possible, judged where necessary — field-fill and matching are computed; an LLM judge is used only where strings fail.
- Reproducible — datasets, queries, and scoring scripts are committed. Search is non-deterministic, so committed numbers are one representative run.
Full method, caveats, and reproduction steps live in each source repo.
linkup-benchmarks/
├── README.md # this file — the consolidated view
└── assets/
├── make_hero_charts.py # regenerates all four composites
├── people_hero.png # 3 people signals in one image
├── company_hero.png # 2 company signals in one image
├── finance_hero.png # FinSearchComp T2 + T3 tiers
└── legal_hero.png # RedlineBench leaderboard
Regenerate the charts: pip install -r assets/requirements.txt && python3 assets/make_hero_charts.py