Multi-provider parallel deep-research aggregator with cross-reference synthesis

Send one research question to four deep-research providers in parallel — Gemini Deep Research, OpenAI Deep Research, Perplexity, and Parallel.ai — then run a cross-reference synthesis step (a Claude skill) that tags each finding as consensus (2+ providers agree), unique (one provider only), or contradiction (providers disagree). Contradictions and uniques get a targeted secondary research pass to decide what to keep. Consensus findings ship into the final report; verified contradictions get flagged with confidence levels. Parallel.ai's structured JSON output makes it the right pick for programmatic agent use; Gemini's browser-mode DR is best for human-driven multidisciplinary reviews. NotebookLM is the working synthesis surface when the question stays open across multiple sessions.

Reported numbers from operators running this:

• One ICP study for AI consulting work, 4 variant prompts × 4 providers = 16 reports → synthesis produced 32 consensus + 30 unique + 4 contradictions = 66 total findings vs an average of ~30 findings per single provider. That's +120% findings vs the best single provider.
• Parallel's Ultra 8x mode reads ~50k pages in 30–50 minutes on a single hard run; an operator running 6–8 parallel Ultra researches brackets a topic in an evening.
• Cursor productized the underlying primitive (ask two models the same question, compare) into a public two-model UX in May 2026, validating that the pattern works at consumer scale.

1. Define the research question and 4 variant prompts. Variants reduce single-prompt brittleness. Keep them concrete: same target topic, different angles ("competitive landscape", "customer pains", "regulatory exposure", "channel economics").
2. Send each variant to each provider. 16 reports total in the reference workflow. Use the providers' native deep-research surfaces — they each have their own pacing, retry logic, and citation style.
3. Drop the raw reports into a synthesis surface. Two paths: (a) a Claude Code skill that reads all 16 reports and emits a structured synthesis; (b) NotebookLM as the conversational surface where you keep asking follow-up questions across the corpus. Source uses both depending on whether the deliverable is a one-shot report or a live research thread.
4. Run the synthesis prompt. It tags each finding with consensus / unique / contradiction and writes a structured output. The synthesis is where the value lives — not in the individual reports. Each provider has blind spots that only show up in comparison.
5. Secondary research on contradictions and uniques. Don't ship a contradiction; resolve it with a targeted follow-up. Source warns: contradictions are where each provider is most confident-wrong; do not default to majority vote.
6. For deep technical topics, use Parallel Ultra 8x. Source operator's example: scanning the full Confidential Computing literature to validate his own architecture, ~50k pages digested in <1 hour per Ultra run; output dumped into NotebookLM for continued conversation.
7. For programmatic / agent-driven research, lean on Parallel. Its JSON output integrates cleanly with downstream agents; GPT and Gemini DR output is markdown-shaped and needs additional parsing for agent pipelines.

Synthesis prompt for the Claude skill that reads N reports and emits a categorized output:

You are reading N deep-research reports on the same question, from
different providers. Your job is to produce ONE synthesis with three
sections:

1. CONSENSUS findings — claims supported by 2+ reports. Cite which.
2. UNIQUE findings — claims appearing in only one report. List with
 the source provider and a one-line note on whether the claim is
 verifiable from the report's own citations.
3. CONTRADICTIONS — claims where reports disagree directly. List each
 with the conflicting positions and the providers holding each.

Do not invent. If a provider made a claim and others didn't address
it, that's UNIQUE, not contradiction. If reports approach the same
question from incompatible framings, surface that as a meta-note at
the end — not as a contradiction.

Field-tested deep-research provider mix:

Parallel.ai — best for programmatic / JSON output; Ultra 8x for hard topics
Gemini DR — best for human-driven multidisciplinary research (browser)
OpenAI DR — strong on synthesis; markdown output
Perplexity — fast; good for grounding short-cycle questions
NotebookLM — synthesis surface for cross-session research threads

• Don't trust majority vote on contradictions. Source explicitly warns: contradictions are where each provider is most confident-wrong. Resolve them with targeted follow-up, not by tallying.
• The synthesis is the deliverable, not the raw reports. Operator: "Synthesis is where the value lives. Each provider has blind spots on its own output."
• Parallel's JSON output is the differentiator for agents. GPT / Gemini DR markdown needs parsing before agent integration; if you're piping into a downstream agent, Parallel is the right surface.
• Cost compounds fast on Ultra 8x. One operator reports a $17/run economic anchor on Gemini Search Grounding via Vertex service-account; Parallel Ultra 8x is in a similar class. Budget per question, not per session.
• NotebookLM is a conversational surface, not a synthesis API. For one-shot synthesis, run the Claude skill against the reports; for an ongoing research thread, drop them into NotebookLM and converse with it.