4.3 Secondary Market Research

In Brief

Description

Secondary market research tests a market hypothesis you already hold — about its size, segment, competition, or trajectory — against existing published material: government statistics, analyst reports, academic research, regulatory filings, trade-association data, public-company disclosures, founder benchmarks, and competitor data. You come in with a stated belief (“this segment is large enough and moving in our favor”) and use desk research to confirm, resize, or kill your idea using evidence that already exists. The goal is a defensible TAM / SAM / SOM — total addressable market (everyone who could buy), serviceable addressable market (the slice you can reach), and serviceable obtainable market (the slice you can realistically win) — and a trend map that together validate, or refute, the hypothesis before you commit primary-research budget.

Approaches to secondary market research vary, but two principles hold throughout.

The first is triangulated and bottom-up by default. A single TAM number from a single analyst press release is a guess. A defensible TAM rests on at least two independent estimation paths — a top-down estimate from analyst or government data and a bottom-up build-up from named-segment counts multiplied by annual contract value (ACV, the revenue one customer generates per year) — that agree to within roughly 3–5×. When the two disagree by more than that, the assumptions need work, not the number. A polished $50B TAM pulled from one report is usually less credible than a $200M TAM with airtight bottom-up math, explicit segment definitions, and visible sensitivity to its assumptions. The same triangulation rule applies to macro trends: every trend claim should rest on at least two independent perspectives — analyst forecast plus government data, or trade press plus a public-company 10-K, not two restatements of the same primary source.

The second is citation-backed claims with mandatory fact-check. AI assembles secondary-research material quickly, but current models fabricate report titles, statistics, and citation URLs at rates between 18% and 94% depending on the task. The most pernicious failure mode is real URL plus fabricated claim — the citation looks defensible because the URL resolves, but the cited page does not actually contain the cited number. Every numeric claim and every named source must trace to a verifiable URL, and a fact-check pass — by a human or a verification subagent that re-fetches each URL and confirms the cited claim is on the cited page — is part of the method.

For query-level demand signal — what people are typing into Google and asking AI assistants right now — use Search Trend Analysis. This page is about how big a market is and where it is heading according to published material; that one is about what the audience is actively searching for. The two methods are complementary and often run in parallel during the same week.

How to

Prep AI Prompt

1. Frame the decision question, not the topic. “Is this market big enough?” is too vague to research. “How many US mid-market fintechs (50–500 employees) would pay $30K+/yr for compliance automation today, and how do the EU AI Act + state privacy laws shift that number over the next two years?” is researchable and decision-anchored. Limit yourself to 2–3 such questions per research session — anything more and the work expands indefinitely.

2. Define the segment slice you can actually reach. Not “businesses” but “B2B SaaS companies with 10–50 employees in the US.” Not “consumers” but “homeowners aged 30–50 with household income above $100K.” Every later number — TAM, SAM, SOM, willingness-to-pay benchmarks, channel-cost estimates — is anchored to this slice. Skip this step and you produce a headline industry number that looks finished but tells you nothing about the market you can actually sell to.

3. Identify candidate sources across multiple buckets. A single source is a guess; multiple sources from different buckets is a triangulation. Cross-reference at least three of these source-pool buckets before locking your reading list:

   • Government data — Census Business Builder, Bureau of Labor Statistics, SBA.gov, EU Eurostat, country-level statistics offices. Free, large samples, slow to update.
   • Library-accessible analyst databases — Most US public libraries and university libraries carry at least one major consumer- or industry-research database; ask the reference desk before paying.
   • Free analyst summaries — McKinsey Global Institute, Bain Insights, BCG Insights, HBR, World Economic Forum, OECD. Full reports are paywalled; the free-tier summaries often carry the headline number you need.
   • Founder benchmarks — annual benchmark reports published by SaaS lenders, VC firms, and payment platforms. Free, recent, founder-relevant.
   • Regulatory filings — SEC 10-Ks for public competitors and adjacent players, FDA / EMA pipelines for regulated categories, EU AI Act compliance filings, state-level privacy filings. Underused; often the only place segment-level numbers are disclosed.
   • Trade associations — IAB for ad tech, ICMA for capital markets, AHIP for health insurance, RIAA for music — often publish unit-economics and channel-cost data not available elsewhere.
   • Trade press and beat reporting — sector-specific publications and dedicated beat coverage. Useful for recency, but always trace the cited primary source.

   If a number shows up in only one bucket, treat it as directional. If it shows up in two or more independent buckets, it earns a slot in your sizing.

4. Sketch your sizing approach and trend dimensions before pulling data. Decide your TAM / SAM / SOM build-up plan in advance:

   • Top-down — start from a published industry total, multiply by the share that matches your segment slice (geography × company size × use case).
   • Bottom-up — start from a named count of segment members (Census counts, funding databases, professional networks) × ACV × time period.
   • Benchmark-anchored — anchor an unknown number to a comparable category’s known benchmark (e.g., scheduling-software ACV ≃ project-management-software ACV at the same segment size).

   Plan to run at least two paths and triangulate. For macro trends, pick the dimensions that matter to your specific segment from the PESTEL frame — Political, Economic, Social, Technological, Environmental, Legal — and drop the dimensions that do not. For most B2B software the dimensions that matter most are Legal (compliance regimes), Technological (adoption curves of upstream platforms), and Economic (segment-level capital-spending shifts); Environmental rarely applies. Stub the trend map with the dimensions you will fill, before collecting data.

5. Tag claims as decision-grade or directional. Some claims drive decisions (TAM in the 1–5× range you are committing to, willingness-to-pay benchmarks for the segment you are pricing against, regulatory pipeline that affects your go-to-market). Others are background context (rough industry growth rate, generic technology-adoption color). Tag each candidate claim decision-grade (must be verified against a primary source and re-fetched by a fact-check pass) or directional (a roundup or AI summary is fine). Without this triage, fact-checking either takes forever or doesn’t happen.

Execution AI Prompt

The verified reading list from Prep is now the input to structured collection. The goal here is not to write essays on each source; it is to fill a fixed dual schema — sizing facts and trend facts — so the synthesis stage has clean inputs.

1. Use a fixed dual schema, not free-form notes. Every source produces two structured records:

   Sizing record — segment definition (geography × company size or demographic × use case), headline number, methodology used by the source to derive it, sample or sample size, year of underlying data, who funded the report (if applicable), and any caveats the source itself flagged. Inline [1], [2] citations on every numeric claim.

   Trend record — PESTEL dimension(s) the source addresses, the specific shift being claimed, the time horizon, the methodology behind the projection, the named affected segment, and any countervailing forces the source acknowledged. Inline citations on every projection.

   A Sources list at the end of each record resolves every marker to a working URL with the date the source was retrieved.

2. Collect each field from its canonical source.

   • Segment counts and demographics → government statistical data, funding databases, and professional-network counts → bottom-up sizing.
   • Top-down industry totals → analyst reports and library-accessible databases → top-down sanity check.
   • ACV, willingness-to-pay, retention, CAC benchmarks → founder benchmark reports, public-company 10-Ks for adjacent players, trade-association unit-economics reports → bottom-up multipliers.
   • Regulatory pipeline → primary regulatory filings (SEC EDGAR for public companies, FDA / EMA pipelines, EU AI Act timeline documents, state-privacy law trackers like IAPP) → trend map (Legal).
   • Technology-adoption curves and platform shifts → analyst adoption forecasts, public-company 10-Ks of upstream platforms, World Economic Forum and OECD reports → trend map (Technological).
   • Demographic and economic trends → Census projections, BLS labor projections, IMF / OECD forecasts → trend map (Social, Economic).

3. Build top-down and bottom-up in parallel. Run both estimation paths from the start, not sequentially. The point is not to pick one — the point is to see whether they triangulate. If your bottom-up is $200M and your top-down is $50B, one of the two is reading the wrong segment and you need to find out which before you write anything.

4. AI does the first pass; humans verify decision-grade claims. A capable LLM with web access can fill 70–80% of the dual schema in hours. The remaining 20–30% — and any claim you tagged as decision-grade in Prep — must be confirmed by visiting the cited source yourself or routing through a fact-check subagent. Pricing pages change weekly; analyst-report headline numbers shift between editions; “X% growth” claims often hide a different segment definition than yours.

5. Flag everything gated as MANUAL. Anything behind a login, paywall, library-only access, or paid analyst report gets flagged as MANUAL — [what would resolve it] in the record. AI cannot fetch behind authentication, and a confidently-stated number from a source the agent could not actually retrieve is the most common failure mode of AI-generated market sizing.