Market research used to mean a week of browser tabs, spreadsheets, and increasingly stale notes. A researcher would spend Monday pulling industry reports, Tuesday scraping competitor pages, Wednesday digesting LinkedIn threads and Reddit forums, and Thursday trying to synthesise it all into something useful before Friday’s presentation. At Choco Media, we started applying ai market research workflows to this problem seriously about eighteen months ago, and the difference is not marginal — a process that used to take four to five days now takes four to five hours, with better coverage and a cleaner output structure. This post walks through the workflow in detail: what we scrape, how we summarise, how we synthesise, and where we still rely on human judgment.
This is for teams who already use AI writing tools but have not yet systematised the research phase. If you are still doing desk research the manual way — opening ten tabs, copying passages into a doc, trying to remember which source said what — this workflow will change your week.
Why Manual Desk Research Breaks at Scale
The core problem with manual research is not speed, it is consistency. A researcher who spends a day on competitive analysis produces a different output each time depending on where they start, how tired they are, and what they happen to notice. The research is good, but it is also fragile — it depends entirely on individual judgment and memory.
When you start doing research at volume — multiple clients, multiple markets, multiple briefs per week — that fragility compounds. You either hire more researchers, which is expensive, or you produce thinner work, which is a different kind of expensive.
AI does not solve the judgment problem. It solves the consistency and volume problem. The workflow we use is not about replacing thinking with AI — it is about using AI to handle the mechanical parts (gathering, formatting, summarising) so that thinking time is spent on interpretation rather than data entry.
What Actually Takes the Most Time in Manual Research
Before building the workflow, we mapped where time went in a typical desk research project:
- Finding and opening sources: 20–25% of total time
- Reading and taking notes from sources: 35–40%
- Organising notes by theme or question: 20–25%
- Writing the synthesis document: 15–20%
AI addresses the first three categories almost entirely. The fourth — writing the synthesis — benefits from AI assistance but still needs a human layer for accuracy and judgment.
Step One: Define the Research Brief Before Touching Any Tool
The biggest mistake teams make when switching to AI-assisted research is jumping into prompts before they know what they are actually trying to answer. The output quality is almost entirely determined by how clearly you define the questions upfront.
A good research brief has three components: the central question, the sub-questions, and the source types. For example, if a client wants to understand the competitive landscape for B2B SaaS onboarding tools in the Nordic market, the brief might look like this:
- Central question: Who are the main competitors, what do they charge, and what do customers say about them?
- Sub-questions: What features are table stakes? Where are the gaps? What does the negative review pattern look like?
- Source types: Company websites, G2/Capterra reviews, LinkedIn posts from practitioners, Reddit threads, relevant industry reports
With this brief, every subsequent AI prompt has a clear job. Without it, you get volume without direction — which is worse than manual research because it looks like it worked.
Step Two: Structured Scraping with Source Tagging
We do not use AI to scrape — we use it to process what we scrape. The scraping itself happens through a combination of tools depending on the source type.
Tools We Use for Gathering
- Competitor websites: Direct reading or tools like Firecrawl for structured extraction of pricing pages, feature lists, and case study pages
- Review platforms: Manual sampling from G2, Capterra, and Trustpilot — we pull 30–50 reviews per competitor, split between most helpful positive and most helpful negative
- Social listening: Perplexity for synthesised summaries of what practitioners are saying across LinkedIn and Reddit; we treat this as a starting point, not a primary source
- Industry reports: PDF extraction using standard tools, then chunking for AI processing
The critical habit is source tagging from the start. Every piece of content that enters the pipeline gets a tag: source type, URL or reference, date retrieved. This sounds tedious but it takes about thirty seconds per source, and it is the thing that makes the final synthesis document trustworthy rather than vague.
The difference between a research output people trust and one they quietly ignore is usually source attribution. If someone can trace a claim back to its origin in two clicks, they believe it. If they cannot, they do not.
Step Three: Summarisation Prompts That Preserve Specificity
Once material is gathered and tagged, we run it through a summarisation stage. This is where most teams go wrong — they use a generic “summarise this” prompt and end up with outputs that are accurate but have had all the interesting specificity stripped out.
The prompts that work for us follow a consistent structure:
The Core Summarisation Prompt Pattern
- Role: “You are a market research analyst summarising competitive intelligence for a marketing strategist.”
- Task: “Summarise the following source in 200–300 words.”
- Constraints: “Preserve specific claims, numbers, product names, and direct quotes where present. Do not paraphrase claims into generalisations. Flag anything you are uncertain about.”
- Format: “Use bullet points grouped by: key claims, evidence, notable gaps or contradictions.”
That last constraint — “do not paraphrase into generalisations” — is the most important line. Without it, AI will turn “the product has a 45-day free trial and no credit card required” into “the company offers a generous trial period,” which is useless for competitive research.
We run this prompt per source, not across batches. Processing sources individually is slower but produces summaries that are actually specific enough to use.
Step Four: Thematic Synthesis Across Sources
Once every source has a structured summary, we move to synthesis. This is the step that turns a pile of notes into insight — and it is where AI earns its keep most clearly.
The synthesis prompt asks AI to work across all the summaries at once. A typical version:
- “Here are structured summaries of [n] sources on [topic]. Your job is to identify: (1) the 3–5 themes that appear across multiple sources, (2) where sources agree, (3) where they contradict each other, (4) what is conspicuously absent from all sources.”
- Format: One section per theme, with specific source references for each claim.
The “conspicuously absent” question is one of the most valuable outputs. In our experience, what the available material does not mention is often as strategically interesting as what it does — it points to gaps in the competitive conversation that a client might be able to own.
We typically run synthesis in 2,000–3,000 word chunks if the source material is long, then do a second-pass synthesis across the chunk outputs. This avoids the context-window degradation that happens when you throw too much material at a model in one go.
Step Five: The Human Review Layer
We want to be direct about this: AI-assisted research is not unreviewed research. The workflow above produces a first draft of understanding, not a final deliverable.
The human review step has three jobs:
- Accuracy check: Verify specific claims, especially numbers, against the original sources. AI summarisation occasionally introduces errors — not fabrications exactly, but rounding, misattribution, or dropped context. Spot-checking 20–30% of claims is usually enough to catch patterns.
- Interpretation layer: The synthesis tells you what is there. The human layer adds what it means for the specific client and situation. This is not something AI can do reliably without deep context it does not have.
- Gap identification: Are there source types you did not cover? Markets you missed? Stakeholders whose perspective is absent? The human review step is where you decide whether the research is complete enough to act on.
In client work we have found that researchers who embrace the workflow spend most of their time on this interpretation layer — which is where their expertise actually adds value — rather than on the mechanical gathering and formatting work AI now handles.
The Prompt Templates We Use Most
Rather than abstracting these, here are the literal prompt structures we return to most often:
Competitor Pricing Page Summarisation
- “Summarise this pricing page. Extract: all plan names and prices mentioned, what each plan includes (feature list), any limits or caps, trial/freemium terms if present, and anything about enterprise or custom pricing. Preserve exact numbers. Flag anything that seems ambiguous or missing.”
Review Sentiment Extraction
- “Here are [n] customer reviews of . Identify: the top 3 things reviewers most commonly praise, the top 3 complaints or friction points, any patterns in which types of users are positive vs. negative, and any specific features mentioned by name (positive or negative). Use direct quotes where they are short and specific.”
Cross-Source Contradiction Finder
- “Here are summaries from [n] sources about [topic]. Identify any claims that directly contradict each other, and list the sources on each side of the contradiction. Then give your best assessment of which claim is better supported and why.”
These are not proprietary — they are the result of running similar research projects dozens of times and keeping the prompts that consistently produced usable output. If you are building your own AI content creation workflows, a similar approach to prompt refinement through repetition is worth the investment.
What We Still Do Manually — and Why
The workflow compresses research time dramatically, but it does not automate judgment. We still do manually:
- Primary source interviews: Talking to people — customers, practitioners, subject matter experts — produces insight that no scraping workflow reaches. AI can help you prepare interview questions and process transcripts, but the conversation itself needs a human.
- Trend interpretation: Identifying which emerging signals matter for a specific client in a specific market is a judgment call that depends on context AI does not have.
- Final framing: The way you present research to a client — what you lead with, what you downplay, what you recommend — is strategic communication, not data processing.
The goal is not to remove humans from research. It is to make sure the human time is spent on the parts that require human judgment. If you are building any kind of systematic content or marketing intelligence operation, the AI automation layer under the research workflow is where the leverage lives.
Getting Started Without Overhauling Everything
The fastest way to see the return from this approach is to run a single research project in parallel: do it your current way, and also run the AI-assisted version alongside. This surfaces the tradeoffs concretely — where the AI workflow is faster and good enough, and where it falls short of what manual research would have caught.
In our experience, the first parallel run convinces most teams immediately. The second and third runs are when the habits stick and the prompts get refined. By the fourth or fifth run, the workflow is natural and the only remaining question is how to keep improving the prompt library.
If you want to think through how this kind of workflow could work for your team’s specific research needs, the contact page is the right place to start.