How to Check an AI Market-Size Estimate Before Sharing It With a Small-Business Client

Direct answer: Treat an AI-generated market-size figure as a draft hypothesis, not a verified fact. Before sharing it, reconstruct the definitions, source inputs, arithmetic, time period, geography, currency, and denominator; compare the result with an independent primary dataset; and label what is observed, calculated, assumed, or uncertain.

That workflow matters because a polished answer can hide a mismatched definition or an unsupported leap. The Federal Trade Commission says advertising claims must be truthful, not deceptive or unfair, and evidence-based.[1] This article is an educational quality-control method, not legal, tax, investment, pricing, or financial advice. For a client’s particular use, consult qualified professionals and current primary rules.

What an AI market-size estimate actually is

A market-size estimate is usually a model built from a definition and a set of inputs. It may describe annual revenue, units sold, number of establishments, number of buyers, or another measure. Those are not interchangeable. “The U.S. market for accounting software” could mean vendor revenue, customer spending, subscriptions, or the count of potential businesses. An AI system may silently switch between them.

TAM, SAM, and SOM are labels for progressively narrower scopes, but the labels do not make a calculation valid. A useful audit asks what each term means in this project, which population it includes, and whether the measures are compatible. The Census Bureau’s Statistics of U.S. Businesses, for example, provides annual data on U.S. business establishments by geography, industry, and enterprise size, while its methodology explains coverage and data construction.[2] That can help test a business-count denominator, but it is not automatically a measure of customer spending.

The six-part audit trail

1. Freeze the question and definitions

Write one sentence that fixes the object being measured: “Estimated annual spending by U.S. nonemployer firms in industry X on service Y during calendar year Z.” Record whether the unit is dollars, customers, transactions, seats, or establishments. Define inclusion and exclusion rules before looking at the result. If the client uses “market size” differently from the analyst, resolve that difference first.

Also separate market size from opportunity for this business. A large category does not establish that a particular offer can reach it, that buyers are accessible, or that demand is proven.

2. Trace every input to its original source

Ask the AI to return a source table rather than only a number. For each input, capture the publisher, exact page or table, publication date, observation period, geography, unit, definition, and access date. Open the original source yourself. Do not treat a search snippet, an uncited summary, or a citation that merely mentions a related topic as verification.

Prefer primary sources such as statistical agencies, regulators, company filings where appropriate, and original survey or methodology documents. The Census Bureau identifies the Economic Census as a comprehensive five-year measure of U.S. businesses and provides multiple business datasets with different scopes.[3] The correct dataset depends on the question; “official” does not mean “automatically suitable.”

3. Check time, geography, and units

Look for silent mismatches. A global forecast may be compared with a U.S. count. A 2030 projection may be presented beside a 2024 observation. Nominal dollars may be mixed with inflation-adjusted dollars. Monthly recurring revenue may be annualized without saying so. A source may report establishments while the model calls them companies or buyers.

Put all conversions in a visible worksheet or notes. State the exchange-rate date, inflation convention, annualization rule, and rounding rule. If the source does not supply enough information to reproduce a conversion, mark that input as unverified rather than filling the gap with a confident-looking assumption.

4. Recalculate the model independently

Rebuild the arithmetic in a spreadsheet or a small, inspectable script. A basic top-down model might be addressable entities × estimated annual spend per entity. A bottom-up model might be number of qualifying entities × adoption rate × annual price. Neither formula is inherently better; the question is whether the inputs are justified and whether the model avoids double counting.

Have the AI show intermediate values, not just the final total. Recalculate with unrounded inputs where possible, then compare the result with the AI’s output. Check signs, decimal places, percentages, and whether a percentage was applied once or twice. If two sources overlap, document the deduplication rule. Keep the original prompt and output in the audit record, but do not let the model’s wording serve as evidence.

5. Run sensitivity and boundary checks

Replace uncertain inputs with a low, central, and high case. This is not a prediction; it is a way to show how dependent the estimate is on assumptions. For example, if the result changes dramatically when the adoption rate moves from 2% to 3%, that assumption deserves more attention than a false impression of precision.

Use boundary checks. Is the estimated number of buyers larger than the total population that could qualify? Is spending per entity implausibly high relative to a known category benchmark? Does the bottom-up total exceed a credible industry total? A failed boundary check does not prove the model is wrong, but it is a stop signal for review.

6. Write uncertainty into the deliverable

Use a claim label such as observed, calculated, assumed, or not verified. State the estimate’s date, scope, unit, and principal limitations near the number. Prefer “under these assumptions, the model produces…” over “the market is…” when the figure is calculated.

NIST’s Artificial Intelligence Risk Management Framework emphasizes that validity and reliability should be demonstrated and that limitations on generalizability beyond tested conditions should be considered.[4] Applied to desk research, that supports documenting the conditions under which an AI-assisted estimate was produced and having a person review the result before it is used.

A practical review checklist

Before sending the estimate, answer “yes,” “no,” or “not applicable” to each question:

• Is the market object defined in one sentence?
• Are TAM, SAM, and SOM operationally defined rather than merely named?
• Does every material input have an accessible original source?
• Do the source period, geography, population, and units match the question?
• Can another reviewer reproduce the arithmetic from the audit trail?
• Have overlapping populations and double counting been tested?
• Have low, central, and high assumptions been shown?
• Have boundary checks been performed?
• Are facts, calculations, assumptions, and unknowns visibly separated?
• Has a human opened the sources and reviewed the final wording?

If any material answer is “no,” hold the number or present it explicitly as an unverified working estimate. Do not improve its appearance by adding extra decimal places.

An original decision tool: the TRACE gate

Use the five-gate TRACE test as a compact handoff tool: Terms fixed, Root sources opened, Arithmetic reproduced, Conditions stress-tested, and Expectations labeled. Give each gate a pass, revise, or hold status. A “hold” at Terms, Root sources, or Arithmetic means the estimate is not ready to share as a factual claim. A “revise” at Conditions or Expectations means it may be shared only with clear limitations and a description of the work still needed.

Keep a dated audit note containing the question, prompt, model version if known, source URLs, extracted values, formulas, assumptions, reviewer, and unresolved issues. This record helps the next reviewer understand what changed when a source is updated. It is also a good way to prevent an AI-generated citation from becoming an accidental unsupported assertion.

Common failure modes

False precision: A result such as $4,287,193 may simply reflect rough inputs. Round to a level the evidence can support.

Forecast laundering: A forecast or vendor estimate is repeated as if it were an observed current total. Preserve the original label and horizon.

Denominator drift: “Businesses,” “employers,” “establishments,” and “buyers” are treated as synonyms. Keep them separate unless a source explicitly supports the conversion.

Source substitution: A secondary article is used when the underlying table or methodology is available. Find and cite the original.

Automation overreach: The analyst assumes that because a model can summarize sources, it has verified them. It has not. Human source inspection remains the checkpoint.

How to communicate the result

Lead with scope, not the headline number: “This is a calculated range for [defined population] in [geography] during [period]. It uses [key inputs], and the largest uncertainties are [items].” Then show the formula, source table, and sensitivity range. Explain what the estimate does not establish, such as demand for a particular offer or a guaranteed business outcome.

The strongest deliverable is often not a single number. It is a reproducible range with a clear evidence trail and a short list of questions requiring client confirmation. That approach makes the research easier to update and less likely to be misunderstood.

Sources and further reading

1. Federal Trade Commission, “Advertising and Marketing Basics.”
2. U.S. Census Bureau, “Statistics of U.S. Businesses.”
3. U.S. Census Bureau, “Business and Economy.”
4. National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework (AI RMF 1.0).”
5. Federal Trade Commission, “FTC Announces Crackdown on Deceptive AI Claims and Schemes.”