Choosing the engine: hosted research vs. BMAD, data included

Chapter 32 described a clean pairing pattern: BMAD frames the question and checks the answer, a hosted engine does the wide crawling in between (Draft the prompt, run it outside, Process the report back). This chapter answers the two questions that pattern leaves open, and they are the two an engineering team actually argues about. Which engine? And, the one nobody asks until legal does: what happens to the question you typed?

A research question is not neutral data. "Does adding image embeddings beat our text baseline?" tells a reader what your ranker is made of and what your next quarter probably contains. Sent to the wrong place under the wrong settings, your research plan is itself a leak. So this chapter compares the engines the way a team should: artifacts, verification, repeatability, automation, and data handling first; benchmark scores last, and with a surprise in them.

All product facts below were verified against official documentation in July 2026; policies change, so treat the specifics as dated claims (the staleness discipline applies to this chapter too).

The contenders, practically

Two structural notes before the comparison dimensions. Gemini is the only one that shows you its research plan for approval before browsing, which is Deep Recon's plan gate as a product feature; if your team has learned to love that checkpoint, it transfers. And Claude's Research is a claude.ai product, not an API: as of mid-2026 the Anthropic API gives you the building blocks (server-side web search and web fetch tools) but you assemble the orchestration yourself, which is exactly what Deep Recon's Run mode does inside your IDE.

What the benchmarks actually say

The most useful independent result is FutureSearch's Deep Research Bench (89 multi-step web research tasks with human-keyed answers, run against a frozen snapshot of the web so tools compare fairly). Its headline finding is the one nobody markets: for several vendors, the plain frontier model with web search beat the branded deep-research product. o3 with search outscored OpenAI's own deep research mode; Perplexity's standard Pro search outscored Perplexity Deep Research. The label earned its keep in two cases: Gemini's Deep Research clearly beat Gemini with search, and Claude's Research slightly beat Claude with search.

Vendor-reported scores tell a similar partial story: OpenAI reported 26.6% on Humanity's Last Exam for deep research at launch, Perplexity 21.1% for its version. And a separate evaluation (DeepResearch Bench) found the citation crowns split: Gemini's reports carried the most supported citations per report, Perplexity's had the highest citation accuracy.

The lesson for a team is not "buy Gemini" or "avoid the mode." It is that "deep research" is a product label, not a capability guarantee, and the spread between tools is smaller than the spread between well-framed and badly-framed questions. Which is an argument for exactly what the last four chapters built: keep the framing, verification, and artifact discipline on your side of the fence, and treat the engine as a swappable crawler. Better yet, measure: a golden set of five research questions your team already knows the answers to, run through each candidate engine once a quarter, is the Chapter 29 validation harness pointed at research tools.

The data question

Here is the table that should precede any "which engine" debate, because it is the one your security review will write on the whiteboard. Consumer tiers and business tiers of the same product have opposite defaults.

Read the consumer column again with your research plan in mind. Three of the four default to learning from what you type, and the one that made it an explicit choice retains opted-in conversations for five years. None of this is hidden; all of it is missed, because the person choosing the research tool is usually not the person who read the data processing addendum.

Now map BMAD's three research modes onto that table, because the modes are also data paths:

• Run mode executes in your IDE against your configured model API or subscription. API traffic sits in the no-training-by-default column at every major provider, and artifacts land as files in your repo. This is the confidentiality-preserving path, and the research firewall adds minimization on top: subagents receive their brief, not your codebase.
• Draft mode produces a prompt you paste somewhere, and the somewhere decides everything. Pasted into a business-tier engine, fine. Pasted into a consumer tier with default settings, your competitive research question just joined a training corpus at three of four vendors. The prompt itself is the leak surface: Deep Recon's drafts deliberately contain the question and source standards, not your architecture, which limits the damage but does not eliminate it.
• Web bundles run whole planning conversations (brief, PRD-shaping, research) inside consumer Gemini Gems or ChatGPT custom GPTs to save metered tokens. That is a strategy-grade conversation on a consumer surface. The flat rate is real; so is the setting you must check first.

The practical policy fits in four lines, and belongs in the same repo as the research templates:

1. Confidential or strategy-revealing research runs in Run mode or on a business tier, never on consumer defaults.
2. Anyone using Draft mode or web bundles verifies the training toggle state of the target account first, once, in writing.
3. Public-domain sweeps (literature, standards, competitor public docs) may use any engine; the findings come back through Process mode into the repo regardless of where they were gathered.
4. The report of record is research.md in git. A shared chat link is not an artifact; it is a bookmark into someone else's retention policy.

Automation: research on a schedule

The staleness map from Chapter 32 implies a recurring job: re-check the claims that age. As of mid-2026 the automation paths are real but uneven. OpenAI exposes deep research as API models (o3-deep-research, and a mini variant at roughly a fifth the price) with per-token pricing plus web-search call costs; practitioner reports put typical runs between cents and tens of dollars, which makes a weekly refresh of five aging claims affordable. Google ships its Deep Research agent through the Gemini API with server-side background execution, built for exactly the disconnect-and-collect pattern a CI job wants. Perplexity's Sonar Deep Research is the fast, inexpensive option. On the Anthropic side you compose the loop yourself from the web-search and web-fetch tools, or run Deep Recon on a schedule in a headless coding agent session.

Both consumer scheduling features (ChatGPT scheduled tasks, Gemini scheduled actions) exist but are consumer-tier conveniences with caps and caveats; for a team, the API path plus a cron job that opens a pull request updating research.md is less magic and more reviewable. The Chapter 32 rule applies: prefer the deterministic interface for a deterministic job (Deep Recon hands its counting to recon_kit.py for exactly this reason).

The decision table

The pattern behind every row: gathering is a commodity; custody is not. Engines compete on crawling, and the benchmarks say even that race is close. What no engine sells you is the part BMAD holds: the framed question, the source standards, the verification pass, the artifact in your repo under your retention policy, consumed by the next document in your pipeline. Choose engines freely; never outsource custody.

One research surface remains, and it is the one this whole part has been circling: the questions nobody on the team even knows to ask, because they live outside the codebase entirely, in regulations, licenses, and adversaries. 👉