Deep research is a chain, not a button — Multi-step research with AI

Why strong research does not come from one giant request, but from a sequence of questions, interim findings, checks and follow-up investigations.

Deep research sounds like a feature: write a question, launch the process and wait for a long report. That may be enough for a small information task. A project needs something else. It needs knowledge that is sufficiently reliable and sufficiently specific to support a decision.

Research should therefore be treated as a process rather than a single retrieval, starting with database construction as the foundation of project work. A proven practice is to explicitly hand the result of one deep-research run into a new chat or browser and continue from there. Later, many separate research documents are consolidated into a shared knowledge base.

The practical conclusion is simple: deep research is a chain. Every stage answers something and creates new questions. Good research ends not when the first long answer ends, but when the remaining knowledge gaps are small enough for the next project decision.

The first research run is orientation, not the final product

After Article 2, the project has a workable question rather than a vague topic. Research can now begin with a purpose. The first round should map the problem space rather than close it prematurely.

Its job is to identify important terms, existing approaches, relevant actors, methods and constraints, useful source types, contradictions and unknowns. It should also show which assumptions from the problem definition become stronger and which become weaker.

The result is not yet a concept. It is a map of the knowledge space.

If you ask for a final recommendation in round one, the model is pushed to compress before the evidence base has been examined. The answer may look efficient while hiding the very uncertainty that could change the project later.

Research starts with knowledge gaps

A useful research chain is built from what the project does not yet know. Typical gaps include the real user group, the size and frequency of the problem, existing solutions, technical requirements, data needs, operating constraints, costs, risks and success criteria.

Instead of "research this topic," create a research backlog:

# RESEARCH BACKLOG

R1 — foundations and terminology
R2 — market / existing solutions
R3 — technical feasibility
R4 — processes and stakeholders
R5 — risks and boundaries
R6 — cost and resources
R7 — counter-arguments
R8 — project-specific deep dive

The backlog does not have to be followed mechanically. Its purpose is to prevent one long report from being mistaken for completeness.

Round 1: broad enough to see the field

The first run should be broad enough to reveal the main perspectives. A stronger prompt than "Which AI solution should I use?" is: "Examine the problem from technical, organisational, economic and user perspectives. Identify solution families, key concepts, common risks, unresolved debates and the questions that must be answered before an architecture decision. Separate established information, plausible assumptions and open points."

This produces an orientation layer rather than a premature tool choice.

Round 2: the previous result becomes input

This is where a chain differs from independent searches. The second run does not restart from zero. It inherits the relevant results of the first.

A practical sequence is:

1. preserve the first result;

2. mark key findings and open points;

3. identify contradictions or weak evidence;

4. formulate concrete follow-up questions;

5. pass only the relevant material into the next research context.

The next run therefore receives not merely more context, but better context.

Follow-up questions are a quality signal

Research quality is visible in the questions it enables. If the first run identifies three technical options, the next questions can compare data requirements, operating costs, skills, vendor dependencies, security constraints and prototype feasibility.

Each round should become narrower and closer to a decision.

Not every finding deserves another deep dive

Multi-step research does not mean following every interesting lead. Before opening another branch, test its relevance, uncertainty, risk, time sensitivity and verifiability.

A question with high decision impact and high uncertainty belongs early in the chain. An interesting fact that cannot change scope, concept or execution can be deferred or discarded.

Collecting sources is not evaluating sources

AI can gather sources very quickly. That increases the need for source discipline. For project work it is useful to distinguish at least four levels: primary sources; authoritative secondary sources; orientation sources; and unconfirmed claims.

The distinction need not dominate the published article. It should remain visible inside the research process so that repetition is not confused with evidence.

Every round needs an interim research state

Do not simply store another long document. Summarise each round in a compact structure:

# RESEARCH STATE

## What we know now
## What is still probable
## Contradictions
## Open questions
## Decision being prepared
## Next research step

This separates knowledge from document volume and makes the reason for the next run explicit.

Research results are allowed to disagree

Two research runs can legitimately produce different conclusions because they use different definitions, dates, source sets or assumptions. In fast-moving AI projects, this is normal.

The contradiction itself becomes a research object: Which definition differs? Which source is closer to the original? Which information is newer? Does each claim refer to the same usage context? Which hidden assumption creates the difference?

Contradictions improve the question when they are made visible.

General research and project-specific research should remain separate

For larger topics, a two-stage method works well: research the general field first, then adapt the findings to the concrete project.

General research asks which strategies, patterns, solution families and recurring risks exist. Project-specific research asks which of those patterns fit our resources, data, audience, scope and constraints.

Mixing both in one huge prompt often produces a smooth answer in which general knowledge and project assumptions can no longer be distinguished.

Research chains can branch in parallel

Once the problem space is understood, independent research paths can run in parallel: architecture, market, user needs, data, risk and business model. The requirement is that these branches are later consolidated.

This creates the transition to Article 4. Separate research files only become reusable project knowledge when they are brought into a structured research pool and knowledge base. This is easy to see in practice: dozens of research documents were collected and semantically connected before they were used for further planning.

The value did not come from one perfect research run. It came from the consolidation of perspectives.

When is the research deep enough?

Research can always continue, so project management needs a stopping rule. A round can close when the main assumptions for the next decision are sufficiently tested, important contradictions are visible, new sources add little decision-relevant information, requirements can be derived, residual uncertainty is documented and the next useful activity is concept work, testing or a decision.

The goal is not certainty. It is decision-ready uncertainty: knowing enough to choose a next step while knowing what remains unknown.

A practical six-cycle model

Cycle 1 — Clarify the question and decision

What will the research enable you to decide or design?

Cycle 2 — Orientation

Map concepts, actors, solution families, risks and source landscape.

Cycle 3 — Deepen

Investigate the three to five most decision-relevant knowledge gaps.

Cycle 4 — Challenge

Search for contradictions, counter-positions and weak assumptions.

Cycle 5 — Translate to the project

Apply general findings to scope, resources, data, audience and constraints.

Cycle 6 — Consolidate

Move reliable findings, open questions, sources and decision implications into a structured knowledge base.

The cycles may repeat. A new finding can send you back to Cycle 3. A contradiction may require a new orientation run. These loops are part of reliable research, not evidence of bad planning.

Deep research is a project process

Treating deep research as a button optimises the length of the answer. Treating it as a chain optimises the quality of the decision.

AI then becomes a research partner inside a controlled process: structuring questions, locating material, condensing interim states, surfacing contradictions, generating follow-ups and comparing perspectives. Human responsibility remains with scope, source standards, priorities and stopping rules.

The most important question after a research run is therefore not "Is the report long enough?" but:

"Which decision can I now make better — and what question should I ask next?"

That is the point at which research begins to become project knowledge.

Worksheet: Build your first research chain

Use the project question from Article 2 or another real project.

1. Define the decision

Which concrete decision should the research prepare?

2. List knowledge gaps

Write at least six things you do not yet know reliably.

3. Draft an orientation research request

Map the field without forcing a solution.

4. Generate follow-up questions

Derive three targeted questions from the first interim state.

5. Classify sources

Separate primary, authoritative secondary, orientation and unconfirmed sources.

6. Define a stopping rule

Write one sentence describing when you know enough for the next project decision.

Reflection

The most important knowledge gap in my project: __________________________________________

The follow-up question most likely to change my decision: _________________________________

All materials to download — the topic overview and the worksheet: