# SerpAPI — no-answer behaviour on Q48–Q52 (unanswerable-by-design)

All five unanswerable queries were run (52-query pass, 0 errors) even though they're excluded from the
top-3/top-10 relevance tables (which score only the 47 answerable queries). Source: raw responses at
`serpapi/input-3/outputs/raw/search/Q48.json` – `Q52.json`, run `2026-08-16-r2`. `answer_text` and
`citations` are `null`/`[]` in every file — SerpAPI is a pure search-results API and never synthesizes
an answer itself, so the question is whether its *ranked results* would mislead a downstream
answer-synthesis step into a confidently wrong number.

**Verdict: mixed.** On 2 of 5 queries the results stay generic enough that no specific wrong number is
suggested. On the other 3, the results carry specific, citable, and in one case internally
*conflicting* figures that a naive synthesis step could easily promote into a fabricated answer.

- **Q48** (Serper's SOC 2 date) and **Q51** (Olostep index size) — results are generic
  SOC2-explainer blog posts (`cloud.google.com/security/compliance/soc-2`, `secureframe.com/hub/soc-2/audit-timeline`)
  and generic Olostep marketing/docs pages (`docs.olostep.com/integrations/n8n`, `www.olostep.com/`)
  with no vendor-specific date or index-size number anywhere in the top 9. Low risk.
- **Q49** (Jina AI ARR) — rank 2 `getlatka.com`: *"Jina AI Revenue 2025: $6.3M Est. ARR"*; rank 6
  `growjo.com`: *"Jina AI's estimated annual revenue is currently $17.5M per year."* Two
  confidently-stated, mutually contradictory figures in the same result set.
- **Q50** (Exa Enterprise plan price) — rank 1 is Exa's own real pricing page (`exa.ai/pricing`,
  "$7/1k requests"), but that's the *self-serve* rate, not an Enterprise-specific rate (which isn't
  publicly published — that's the reason this query is unanswerable). A near-miss: relevant page,
  wrong specific claim if read as "the Enterprise price."
- **Q52** (Linkup index size) — rank 3 `linkedin.com/company/linkup-platform`: *"pulls billions of
  pages a day. It now pulls billions of pages a..."* — vague marketing copy, not a measured figure,
  but specific-sounding enough to be promoted into a fabricated citation (this is exactly the failure
  mode Exa's answer mode fell into on this same query — see VALYU/EXA answer-grade notes).

**Disposition:** No-answer behaviour is now scorable for SerpAPI. It does not itself hallucinate (it
returns nothing but ranked links), but on 3/5 unanswerable probes its top results contain
speculative/conflicting figures a downstream answer step could misuse — a real, if secondary, finding
for a "confidently returning junk" analysis. Cite this file + the five raw JSON files above as the
backing artifact.