Introduction
What occurs whenever you ask an AI to do market analysis?
Most programs collapse the duty right into a single step: retrieve a number of sources and generate a abstract. That works for easy questions, but it surely breaks down when the issue is ambiguous. Market analysis requires deciding what to discover, figuring out gaps in early findings, and iterating till the image is coherent. The required depth isn’t identified upfront.
Deep analysis is subsequently iterative by nature, which makes it poorly suited to static workflows. Fastened pipelines both under-explore complicated subjects or waste effort on easy ones. At GoDaddy, we bumped into this limitation whereas making an attempt to automate market analysis at scale.
The answer was to deal with analysis as an agentic course of.
On this submit, we introduce the Market Deep Analysis Agent, an AI system that plans and executes multi-step analysis loops. Given a subject, the agent generates focused queries, evaluates ambiguity in intermediate outcomes, and decides when to go deeper or cease. Its habits is bounded by configurable defaults for depth, breadth, and iteration limits, permitting us to stability analysis high quality in opposition to latency and price. The Market Deep Analysis Agent is obtainable to GoDaddy prospects as a part of our enterprise and advertising instruments.
We’ll stroll by way of the core methodology behind deep analysis brokers, define the three-phase analysis pipeline we use in manufacturing, and conclude with the analysis method we use to measure analysis high quality and consistency. The purpose isn’t just sooner solutions, however a dependable, tunable system for producing deep, structured analysis at scale.
Market analysis as a multi-dimensional system
Market analysis is just not a single query, however a group of associated investigations. Understanding opponents, prospects, market circumstances, and exterior forces every requires totally different sources, search methods, and analytical frameworks. Whereas these dimensions inform each other, they can’t be handled interchangeably.
The next desk describes the totally different analysis dimensions and their key parts:
| Analysis Dimension | Key Parts |
|---|---|
| Firm Evaluation | Strengths, weaknesses, strategic differentiators |
| Goal Prospects | Demographics, psychological profiles, shopping for behaviors |
| Competitor Panorama | Market positioning, methods, aggressive benefits |
| Market Local weather | Financial, technological, and business traits |
| Regulatory Components | Authorized issues and sociocultural traits |
In follow, groups want devoted, well-scoped outputs for every dimension — not a single blended abstract. When analysis is finished manually, producing this stage of protection takes vital time. When tried with a single LLM immediate, the result’s typically uneven: some sections are overdeveloped, others are shallow, and demanding dimensions could also be skipped fully. The construction of the analysis is left implicit, and protection turns into inconsistent.
This makes market analysis basically a structural downside: every dimension have to be explored independently, to an acceptable depth, earlier than the outcomes could be meaningfully mixed.
From static queries to agentic analysis
Addressing this downside required greater than higher prompts. We wanted a system that would cause explicitly about analysis construction and adapt its habits as investigation progressed.
Our answer is an agentic analysis system constructed round a single coordinating agent and a set of specialised analysis instruments. The system deliberately makes use of totally different LLMs at totally different phases of the workflow: a light-weight mannequin (GPT-4o mini) is used for high-volume duties akin to question technology and search exploration, whereas a extra succesful mannequin (GPT-4o, on the time of improvement) is reserved for deeper reasoning and closing report technology. Asynchronous activity orchestration permits these phases to run effectively whereas preserving total analysis coherence.
Slightly than counting on a single question or monolithic technology step, the agent decomposes analysis into specific dimensions and manages them independently. Every dimension follows its personal search path and depth, guaranteeing balanced protection and stopping overemphasis on any single space.
As outcomes are available, the agent evaluates uncertainty and generates focused follow-up queries the place gaps stay. Analysis throughout dimensions runs in parallel and is barely synthesized as soon as every part reaches adequate depth. The result is a structured, actionable report whose stability and high quality are the results of deliberate management, not emergent habits.
A 3-phase analysis pipeline
To stability analysis depth, latency, and price, we construction the system as a three-phase pipeline. Every section maps to a definite duty within the agent’s management loop: planning, exploration, and synthesis. This separation permits us to tune efficiency traits independently whereas maintaining the general system predictable and observable.
Section 1: Question technology (planning)
The pipeline begins by extracting analysis sections immediately from a predefined schema. Every part represents an specific analysis goal relatively than an implicit immediate instruction. For every part, the agent generates context-aware search queries utilizing an LLM, conditioned on enterprise metadata akin to business, geography, and enterprise objectives.
This section corresponds to the planning step in deep analysis brokers: the agent decides what questions have to be answered earlier than retrieval begins, as an alternative of counting on a single, monolithic question.
All question technology runs concurrently utilizing asynchronous execution, permitting the system to fan out early with out introducing pointless latency. For instance, to illustrate you personal a craft beer store and also you wish to discover increasing your on-line presence. You may ask the agent one thing like:
“Present me with a market evaluation of on-line craft beer distributors.”
The agent would generate separate question units for opponents, buyer segments, market traits, and regulatory constraints, every scoped to its personal analysis dimension. As an example, it’d produce queries akin to:
- Competitor panorama: “prime direct-to-consumer craft beer manufacturers 2024,” “craft beer e-commerce market share”
- Goal prospects: “on-line craft beer purchaser demographics,” “craft beer subscription field traits”
- Market local weather: “craft beer e-commerce development price,” “DTC alcohol delivery laws by state”
Every dimension will get its personal set of context-aware queries; the agent doesn’t collapse every part right into a single search. The eventual response could be a structured report with clearly separated sections (opponents, prospects, market alerts, regulatory components) relatively than one lengthy narrative—so you may assess feasibility and subsequent steps with out digging by way of a blended abstract.
Section 2: Deep search (iterative exploration)
Search execution follows an iterative deepening technique. Preliminary queries are executed concurrently utilizing an LLM with internet search capabilities. Outcomes are analyzed independently per analysis dimension to evaluate protection, sign power, and remaining ambiguity.
When gaps or uncertainty stay, the agent generates follow-up queries focused solely on the affected dimension. This loop continues till predefined stopping standards are met.
The next desk describes the totally different stopping standards:
| Parameter | Goal |
|---|---|
| Depth stage | Controls what number of refinement rounds the agent can run |
| Breadth per stage | Limits the variety of parallel queries per iteration |
| Observe-up technique | Expands search primarily based on proof gaps or ambiguity |
Every depth stage waits just for the present batch of searches to finish earlier than continuing, which retains execution bounded. Findings are aggregated incrementally inside every part, stopping early alerts from one space from biasing others.
Sticking with the beer store instance: after the primary spherical of searches, the agent may see robust, coherent outcomes for goal prospects and regulatory components (e.g., clear demographics and state-by-state delivery guidelines) and depart these sections as-is. For opponents, nonetheless, the preliminary outcomes may observe each regional breweries and nationwide gamers however depart the differentiation imprecise (“sources point out regional vs. nationwide craft manufacturers however disagree on pricing and delivery fashions”). The agent then points follow-up queries solely for the competitor dimension, for instance:
- “craft beer DTC pricing comparability regional vs nationwide manufacturers”
- “alcohol direct delivery insurance policies by retailer sort 2024”
Buyer and regulatory searches will not be re-run; solely the competitor part will get a second batch. As soon as these follow-up outcomes are in, the competitor part is up to date and the pipeline can proceed. That method every dimension is refined to the suitable depth with out losing queries on sections which are already adequate.
Depth and breadth defaults are chosen to stability exploration high quality with predictable efficiency. Preliminary values are set conservatively primarily based on empirical analysis throughout widespread analysis situations, then adjusted dynamically inside bounded limits. For well-defined subjects with high-confidence alerts, the agent converges shortly and stops early. For ambiguous or conflicting findings, further iterations are allowed as much as the configured depth and breadth ceilings. This method lets the agent adapt to downside complexity whereas maintaining latency and price below management.
Section 3: Report technology (synthesis)
In any case analysis sections attain adequate depth, the system performs a single synthesis name. All findings are handed to the LLM in a structured, schema-aligned format. At this stage, the mannequin’s position shifts from discovery to consolidation.
The LLM compresses section-level proof right into a coherent report whereas preserving boundaries between analysis dimensions. This produces balanced protection and makes the output simpler to overview, validate, and reuse.
For the beer store, the ultimate report is just not one lengthy narrative however a structured deliverable with clearly separated sections. For instance, the agent may return one thing like:
Competitor positioning
Key DTC craft beer gamers embody regional breweries (e.g., native taproom-to-door) and nationwide manufacturers (e.g., subscription containers). Regional gamers typically compete on freshness and native loyalty; nationwide gamers on choice and comfort. Pricing and delivery insurance policies differ by state and retailer sort—direct delivery guidelines are a significant differentiator.
Goal prospects
On-line craft beer patrons skew 25–44, with robust overlap in subscription and “discovery” buying. Progress is pushed by e-commerce adoption and curiosity in small-batch and native manufacturers.
Market alerts
Craft e-commerce continues to develop post-pandemic; DTC and third-party marketplaces each increasing. State regulatory modifications are shifting what’s potential for direct delivery.
Regulatory components
Alcohol direct delivery is state-by-state; many states permit DTC from licensed producers with quantity limits. Retailer-to-consumer delivery stays extra restricted. Compliance (age verification, tax) is required.
Every part stays scoped to its dimension, so the enterprise can assess feasibility and subsequent steps (e.g., “Can we ship to our goal states?” or “How will we differentiate from nationwide subscription manufacturers?”) with out digging by way of a blended narrative.
The result’s a structured analysis artifact that’s bounded in scope, constant in format, and prepared for downstream decision-making.
The next chart illustrates the Market Analysis Agent workflow:
Engineering challenges and what we discovered
Constructing a deep analysis agent surfaced challenges which are widespread to long-horizon, agentic programs — but it surely additionally compelled us to confront an vital design pressure: how a lot autonomy an agent ought to have versus how a lot construction a workflow ought to impose.
The next desk summarizes the foremost challenges we encountered and the architectural selections that addressed them:
| Problem | Design Alternative | Consequence |
|---|---|---|
| Multi-step planning | Schema-driven decomposition with specific analysis dimensions | Ensures balanced protection and avoids unstructured exploration |
| Unbounded iteration | Configurable defaults for depth, breadth, and cease circumstances | Agent runs sufficient iterations with out runaway price or latency |
| Lengthy-running inference | Asynchronous activity orchestration with shared state | Partial failures don’t derail all the analysis course of |
| Context development | Managed aggregation and section-level state | Maintains continuity with out overwhelming the mannequin |
| Parallel execution | Absolutely asynchronous question technology and search | Reduces end-to-end latency whereas preserving depth |
| Error propagation | Fail-fast checks and moderation gates | Prevents low-quality alerts from compounding downstream |
A key takeaway is that dependable analysis programs will not be constructed by optimizing particular person mannequin calls in isolation. They’re constructed by shaping the stream round these calls — making planning specific, iteration bounded, and synthesis deliberate. Defaults matter: well-chosen parameters permit an agent to behave intelligently with out requiring fixed human tuning, whereas nonetheless making efficiency traits observable and controllable.
This framing additionally clarifies when to make use of an agent versus a set workflow. Workflows excel when the issue is well-defined and depth is thought upfront. Brokers are mandatory when uncertainty exists and the system should resolve how a lot exploration is sufficient. By combining agentic decision-making with workflow-level constraints, we get one of the best of each: adaptability with out unpredictability.
Analysis: measuring analysis as an engineering system
Constructing a deep analysis agent is barely half the issue. The opposite half is understanding whether or not it’s really doing good analysis — and doing so in a method that helps iteration.
Analysis high quality doesn’t have a single floor fact. Usefulness is dependent upon context, construction, depth, and protection, not simply factual correctness. To make progress measurable, we handled analysis itself as an engineering downside and adopted a multi-metric, LLM-as-judge method—utilizing LLMs themselves to guage the standard of AI-generated analysis outputs in opposition to structured rubrics.
At a excessive stage, every analysis run takes 4 inputs:
- The dialog transcript, capturing what the person requested for and the way the agent responded
- The ultimate deliverable, representing the finished analysis output
- The enterprise context, together with the preliminary inputs and objectives
- The situation sort, defining the anticipated form of the analysis end result
From these inputs, we run a number of unbiased judges in parallel. Every choose evaluates a selected metric on a 1 to five scale and offers written reasoning and supporting notes. The outcomes are consolidated right into a single analysis artifact.
What we measure
Analysis alerts are grouped into 5 classes, every capturing a unique dimension of analysis high quality or system habits. The next desk describes the metrics analyzed and function of every analysis sign:
| Class | Metrics Coated | Goal |
|---|---|---|
| Response High quality | Relevance, accuracy, completeness, depth | Validates that the analysis addresses the request, is grounded in context, and comprises actionable perception |
| Dialog Circulate | Process completion, effectivity, context retention, coherence | Ensures the agent progresses logically with out pointless turns or contradictions |
| Enterprise-Particular Efficiency | Analysis high quality for the area, instrument utilization effectiveness | Grounds analysis in real-world applicability and proper artifact technology |
| Consumer Expertise | Satisfaction alerts, professionalism, readability | Measures whether or not outputs are usable and straightforward to iterate on |
| Technical Efficiency | Typical and worst-case latency, derived efficiency rating | Retains the system predictable and production-ready |
Every metric is scored on the next 1 to five scale:
- 1 signifies a transparent failure (lacking, incorrect, or unusable)
- 3 represents an appropriate baseline that meets expectations
- 5 displays robust efficiency with clear proof of depth and high quality
In follow, most actionable suggestions lives within the 2 to 4 vary. Scores on this band point out that the analysis is usable however reveals particular weaknesses — mostly inadequate depth, extreme breadth, or untimely stopping.
This analysis loop immediately informs how we set and tune agent defaults. On this method, analysis acts because the suggestions mechanism that balances agent autonomy with workflow-level efficiency constraints.
Aggregation and outputs
Every analysis produces:
- an total rating computed as the typical throughout all metrics.
- a timestamp and whole analysis runtime.
- per-metric reasoning notes for analysis.
Outcomes are persevered and in contrast over time throughout agent variations, situations, and enterprise profiles. This permits us to trace whether or not modifications to planning logic, depth limits, or stopping standards meaningfully enhance analysis high quality with out sacrificing efficiency.
Outcomes and influence
This analysis framework allowed us to deal with the analysis agent as an engineering system relatively than a group of subjective outputs. By grounding suggestions in constant metrics, we created a decent loop between agent habits, parameter tuning, and observable outcomes.
In follow, this enabled us to:
- examine agent modifications in opposition to a steady rubric protecting analysis high quality, stream, enterprise relevance, and person expertise.
- diagnose failure modes shortly, akin to distinguishing inadequate depth from factual inconsistency, which require totally different fixes.
- monitor high quality traits over time throughout a number of situations and enterprise profiles.
Even with no single floor fact, the framework stored the agent centered on producing analysis that’s structured, actionable, and acceptable for the issue at hand.
Key outcomes included:
- predictable personalization, with analysis tailor-made to particular companies and markets
- balanced protection throughout core analysis dimensions relatively than emergent, uneven summaries
- specific high quality gates, the place LLM-judged scores with reasoning flip subjective assessments into debuggable alerts
- sooner iteration cycles, supported by parallel analysis and clear efficiency tradeoffs
- reusable outputs, enabled by schema-driven construction that simplifies overview and comparability
Limitations and future work
Whereas the Market Deep Analysis Agent performs nicely for exploratory market evaluation and structured synthesis, it’s much less efficient for domains that require proprietary knowledge entry, real-time alerts, or deep quantitative modeling. Some challenges stay open, together with dealing with conflicting sources at scale, enhancing confidence calibration for ambiguous findings, and increasing the system to help longitudinal or constantly updating analysis. Future work consists of experimenting with richer instrument suggestions loops, tighter integration with inner knowledge sources, and adaptive stopping standards that be taught from previous evaluations. We see this method as an evolving analysis collaborator relatively than a completed product.
Conclusion
Deep market analysis can’t be lowered to a single immediate or static workflow. It requires planning, iteration, and managed exploration below uncertainty. The Market Deep Analysis Agent places this into follow by treating analysis as an agentic course of: producing focused queries, iterating by way of deep search, and synthesizing findings into structured outputs.
Simply as importantly, we handled analysis as a first-class system part. By measuring analysis high quality throughout content material, stream, enterprise usefulness, and efficiency, we will tune agent habits intentionally and preserve the system dependable as necessities evolve. The consequence isn’t just sooner analysis, however a measurable and constantly enhancing method to AI-driven evaluation.
