# Linkup `sourcedAnswer` mode — supplementary 11-query test (2026-08-25)

**Scope note: this is an 11-query mini-panel run specifically to close the "untested coverage gap"
flagged for Linkup's answer mode — it is NOT a full 52-query run and must not be bar-charted directly
against Exa/OpenAI/Valyu's 47/52-query answer-mode figures without this caveat.** Panel: Q01, Q15,
Q23, Q32, Q37, Q43 (one per answerable block) + Q48–Q52 (all 5 unanswerable). Endpoint: `POST
https://api.linkup.so/v1/search`, `outputType: "sourcedAnswer"`, `depth: "standard"`. All 11 calls
returned `http_status: 200`, 0 errors. Raw responses: `_human-tasks/review-artifacts/_answer-mode-tests/linkup/Q*.json`.

## Answerable queries — 4/6 correct

| Query | Verdict | Notes |
|---|---|---|
| Q01 (Exa price) | ✅ correct | "$7 per 1,000 requests" — exact match |
| Q15 (ruff B007) | ✅ correct | "B007 (unused-loop-control-variable)" — exact match; cites `docs.astral.sh/ruff/rules/unused-loop-control-variable/` at rank 3, matching ground truth's own supporting URL |
| Q23 (Apollo pricing) | ✅ correct | Resolves to Apollo.io (sales engagement), not GraphOS; $49–119/mo tiers plus the 900-credit/year free tier matching ground truth |
| Q32 (Gujarat+Maharashtra pop.) | ❌ **wrong** | States "~200 million" using 2021 estimates (Gujarat 72.7M, Maharashtra 129.9M) instead of the 2011 Census figures (60.4M + 112.4M = 172.8M) — same base-year-mixing failure mode Valyu showed on this exact query in the main benchmark |
| Q37 (GDPR Art 17(3)) | ⚠️ partial | Names 4 of 5 exceptions (legal obligation, public interest task, archiving/research, public health, legal claims) but **omits the freedom-of-expression exception** |
| Q43 (Python 3.11.16) | ❌ **wrong** | States "does not contain the changelog for Python 3.11.16" — despite ranking `python.org/downloads/release/python-31116/` (the correct page) **as its own #1 source**. Failed to extract from a source it had already retrieved. |

## Unanswerable queries — 5/5 abstained, 4/5 cleanly

| Query | Verdict |
|---|---|
| Q48 (Serper SOC2 date) | ✅ clean abstention — names the 5 unrelated companies its sources actually discuss, explicitly notes none mention Serper |
| Q49 (Jina ARR 2026) | ⚠️ abstained, not fully clean — correctly states no 2026 figure exists, but surfaces the $6.3M 2025 estimate as context (same pattern as OpenAI's Q49 in the main benchmark) |
| Q50 (Exa Enterprise price) | ✅ clean abstention — explicitly distinguishes the $5/1k Pro rate from the custom-quoted Enterprise rate |
| Q51 (Olostep index size) | ✅ clean abstention |
| Q52 (Linkup's own index size) | ✅ clean abstention — correctly declines to state a page count even while discussing the unrelated "LinkUp" jobs company's job-posting count |

## Disposition

**Citation accuracy (answer APIs): scorable now — 4/6 answerable claims backed by on-topic, resolvable
sources; the one verified exact-match citation (Q15, matches ground truth's own URL) confirms the
citation mechanism is real, not decorative.**

**Answer quality (answer APIs): scorable now — 4/6 correct on the answerable mini-panel, 5/5 abstained
on unanswerable (4/5 cleanly).** Weaker than You.com's mini-panel result (6/6 / 5/5 clean) on this
same 11-query set, but with no outright fabrication of the kind Tavily's `include_answer` showed on
Q48/Q49/Q51 — see `TAVILY-include-answer-test-2026-08-25.md`.