How reliable are market size estimates? As reliable as the source's answers to four questions: what exactly the estimate counts, which year, whose spending, and why it was published. It must then survive one hard number, such as company revenues, trade data or transaction counts. A figure that can't answer the first question is a scenario, not a fact.
How reliable are market size estimates in general?
As single points, not very. As bands with a stated basis, reliable enough to underwrite. Caspr's Studies of three markets found the lowest and highest published 2025 estimates between 2.9x and 5.4x apart, and in each case the definition explained most of the gap, not a disagreement about demand.
| Market | Lowest published | Highest published | Spread | What explains most of it |
|---|---|---|---|---|
| AI clinical documentation, global | $0.79bn (Fortune Business Insights) | $4.3bn (The Business Research Company) | 5.4x | Generative AI tools only, against a category that adds coding, compliance and consulting services |
| Industrial valves, Saudi Arabia | $725.0M (IMARC) | $2.09bn (Ken Research) | 2.9x | Industrial valves only, against a wider valves perimeter |
| B2B payments, India | $7.84bn (Mordor Intelligence, real-time payments) | $41.9bn (IMARC) | 5.3x | One payment rail's economics, against the whole B2B payments industry |
This means a single published figure tells you more about its publisher's definition than about the market. Credible is a narrower claim than correct: a credible figure is one whose basis can be stated in a sentence and which survives a count. When two figures count different things, the work is to reconcile market size estimates by definition, not to split the difference between them.
The screen runs in a fixed order, because each answer depends on the one before it.
- Open the original publisher's page, not a reseller listing, a press release or an article quoting it, and note the edition.
- Write down what the figure counts: the product perimeter, and whether it is revenue, transaction value or volume.
- Write down the base year and the forecast window, separately from the date on the page.
- Write down whose spending it covers: geography, buyer segment and channel.
- Note who produced the figure and what it was published to sell.
- Test it against one hard number: a filing, a trade statistic, a transaction count or an adoption survey.
- Carry it forward as a band with its basis and base year, or drop it.
What does the number actually count?
Two choices decide what a market size counts: the perimeter, meaning which products and services are in, and the unit, meaning revenue earned by sellers, value flowing through the market, or volume. Publishers state the perimeter more often than the unit, and the unit is what breaks a model first.
India's B2B payments market shows why. IMARC puts it at $41.9bn for 2025 and $84.2bn by 2034; its page defines B2B payments as transactions between businesses but does not say whether the dollar figure is provider revenue or the value of those transactions. Mordor Intelligence sizes India's real-time payments market at $7.84bn for the same year. NPCI recorded 23.2 billion UPI transactions in May 2026 alone. Caspr's Study of the market read the published figures as revenue pools rather than flows, since the flows are vastly larger, and kept throughput, industry value and monetization pools as three separate numbers for three separate uses.
This means the unit sets the revenue ceiling in the model. Read transaction value as revenue and the addressable pool is overstated by the inverse of the take rate: 100x at a 1% take rate. Read revenue as transaction value and the target's implied share looks implausibly large.
Start from the publisher's own page, not a reseller listing or an article quoting it. A short list of market data you can cite saves most of the search; the rest is reading the definitions before the headline.
Which year does it describe, and how old is too old?
A market report page can carry three dates, and the update date is the least informative. Fortune Business Insights' generative AI clinical documentation page reads "Last Updated: September 28, 2026"; its base year is 2025 and its historical series runs from 2021 to 2024. Mordor Intelligence's page on the same category was last updated April 27, 2026, with figures current "as of January 2026" and a forecast window of 2026 to 2031.
Growth rates carry their own window. Mordor's 21.46% CAGR runs 2026–2031; Fortune's 33.3% runs 2026–2034 on a narrower generative AI subset. A rate quoted without its window and base year cannot be compared with another rate, or with the target's own growth.
Age matters in proportion to growth. At IMARC's 4.96% forecast rate for Saudi industrial valves, a four-year-old figure understates today's market by about 18%. At Mordor's 21.46% for AI clinical documentation, it understates it by about 54%.
54%
This means there is no fixed shelf life for a source. In a slow market, an old figure can be rolled forward at a stated rate and labeled as such. In a fast one it has to be replaced, because the roll-forward assumption becomes the number.
Whose spending does it cover?
A market is somebody's spending, and a credible page says whose: which countries, which buyers, which channels. The headline is almost always the widest version.
Fortune's $0.79bn for generative AI clinical documentation in 2025 is a global figure; the same page gives North America 48.1% of it, or about $0.38bn. A US-focused vendor screened against the global number appears to have roughly twice the room it has.
Buyer segment works the same way. Caspr's Study of the category matched each published estimate to a decision: narrow AI-only figures for screening pure-play vendors, broader adjacent figures for planning by EHR vendors and diversified healthcare IT firms. The same number can be right for one decision and wrong for the next.
In the market section of an investment memo, that means leading with the spending the target can reach, the serviceable market, and showing the global headline beside it for scale rather than as the opportunity.
Why was it published?
Every market figure was produced by someone for a reason, and the reason predicts the direction of error. Motive does not make a number wrong. It tells you how much independent support to ask for before the number goes into a memo.
A syndicated report's product page exists to sell the report. Mordor Intelligence's page for AI clinical documentation says its figures come from a "proprietary estimation framework" and lists a research methodology chapter that the page itself does not show; Fortune Business Insights' page links to its methodology rather than stating it. That is the format's norm rather than one publisher's habit: the method sits inside the product, and the product sits behind the purchase.
Evidence from companies selling the product deserves the same reading. The AI clinical documentation Study found that several of the most detailed deployment studies were written by organizations building or deploying the tools, including a 2026 validation study whose authors were all employees of a single vendor. The US Government Accountability Office made the same point from the other side in July 2026: "There are few independent studies evaluating the accuracy of these tools."
A seller's teaser slide, written to support a price, and a supplier's category claim, written to win shelf space, belong to the same class. Neither is disqualifying. Both need one figure from a party with no stake in the answer.
Does it survive one hard number?
An estimate is a model; a hard number is a count. Company revenues, customs records, transaction totals and adoption surveys bound what a market can plausibly be, and a credible estimate sits on the right side of them.
Saudi industrial valves show how quickly the test works. The narrowest published estimate is IMARC's $725.0M for 2025. World Bank trade data show Saudi Arabia imported $239.6M of safety and relief valves in 2024, and the Saudi valves Study cites a further $142.8M of check valves: $382.4M across two product lines, at customs value, a year earlier. Customs codes classify valves by type, not by end use, so either the narrow estimate leaves out a large share of what customs records, or it sits low. The page has to say which, and that is question one again.
$382.4M
A figure that fails is rarely useless. Reclassify it as a scenario bound, replace it with one whose basis is stated, or rebuild the market from the counts up. Passing the screen is what lets you defend a market size estimate when the committee pushes back, because every answer is already written down.
| What you see | What it usually means | What to do |
|---|---|---|
| No statement of what is counted | The figure cannot be placed in a model | Use it as a scenario bound, or replace it |
| Two listings of one publisher's figure disagree | Different editions, or a copying error downstream | Go to the publisher's own page and edition |
| A growth rate to two decimals, no method shown | Model output, not measurement | Treat the rate as a scenario |
| A recent update date over an older base year | The update date is not the data date | Roll forward at a stated rate, or replace |
| A global figure for a regional business | Headroom overstated | Rebuild the serviceable market for the geography |
| Found only in articles quoting other articles | Provenance lost | Find the original, or drop it |
Worked example
Is a $3.05bn forecast for AI clinical documentation credible?
Mordor Intelligence sizes AI in clinical documentation at $0.98bn in 2025 and $1.15bn in 2026, reaching $3.05bn by 2031 at a 21.46% CAGR. Caspr's Study of the category, completed in August 2026, asked whether that forecast is credible. Its findings map onto the four questions. What it counts: Fortune Business Insights sizes a narrower generative AI subset at $0.79bn for 2025; The Business Research Company sizes clinical documentation improvement at $4.3bn, a category that includes coding support, compliance monitoring and consulting services. Two resale listings of that publisher's figures carry different 2026 values, $4.65bn and $5.16bn: a reason to find the edition, not to use either. Which year: Mordor's window is 2026–2031. Fortune's broader ambient documentation forecast runs from $1.46bn in 2026 to $7.00bn in 2034 at 21.6%. Different windows and labels, the same low-20s growth. Whose spending: provider organizations buying ambient documentation, increasingly through the EHR itself after Epic rolled out AI Charting on February 4, 2026; bundled EHR distribution makes that spending harder to count precisely. Why published: the forecasts sell reports, and much of the adoption evidence is vendor-authored; the GAO noted in July 2026 that few independent accuracy studies exist. The hard number: KLAS's 2025 AI update, drawn from 1,742 healthcare organizations, found 79% using ambient speech: the demand is real even where the revenue boundary is not. The verdict: credible as a band, not as a point. The Study put the defensible 2026 range for pure ambient documentation at roughly $1.0bn to $1.5bn, treated estimates above $4bn as an adjacent market, and told investors to underwrite base cases to the narrow pool at low-20s growth, adding upside only where coding or chart-review revenue is already evidenced.
The published figures
| Publisher | Figure | Period | What it counts | Published | Last verified |
|---|---|---|---|---|---|
| Mordor Intelligence | 0.98 (USD bn) | 2025 | AI in clinical documentation, global revenue | 2026-04-27 (page last updated) | October 17, 2026 |
| Mordor Intelligence | 3.05 (USD bn) | 2031 (21.46% CAGR, 2026–2031) | AI in clinical documentation, global revenue, forecast | 2026-04-27 (page last updated) | October 17, 2026 |
| Fortune Business Insights | 0.79 (USD bn) | 2025 (base year) | Generative AI for clinical documentation, global | 2026-09-28 (page last updated) | October 17, 2026 |
| The Business Research Company (listed by Research and Markets) | 4.3 (USD bn) | 2025 | Clinical documentation improvement: software and consulting services, including coding support and compliance monitoring | 2026-02 | October 17, 2026 |
| TechTarget, reporting KLAS Research | 79 (% of organizations) | 2025 survey | Healthcare organizations using ambient speech AI, KLAS sample of 1,742 organizations | January 20, 2026 | October 17, 2026 |
Why the figures differ
- definitional Mordor Intelligence $0.98bn and Fortune Business Insights $0.79bn (2025) Fortune sizes generative AI documentation tools only, a subset of what Mordor counts as AI in clinical documentation.
- definitional Mordor Intelligence $0.98bn and The Business Research Company $4.3bn (2025) Clinical documentation improvement adds coding support, compliance monitoring and consulting services to the software, so it is a wider market, not a higher estimate of the same one.
Frequently asked
How old can a market size source be in a pitch deck?
A pitch deck's market size source has no fixed age limit; growth sets it. At 5% a year, a four-year-old figure understates today's market by about 18%, which a stated roll-forward can correct. At 21%, it's understated by more than half: replace it. Always show the base year beside the number, not the date you found it.
How do I tell whether a Statista number has a real methodology behind it?
Read a Statista number's source block, not the chart. Statista's pages show the last update separately from the survey period; one statistic on 2023 market research spending was last updated March 10, 2025. If the source names an outside publisher, open the original. If it names Statista's own research department, it's Statista's estimate, so find its method.
How do I check a supplier's market claims independently?
Ask the supplier for the original source behind its market claim, then run four questions on it: what it counts, which year, whose spending, and who paid for it. Set it against one figure the supplier didn't produce, like your own category sales or a government statistic. A claim that survives both is ready for the board.
How do I avoid made-up numbers in an MBA assignment?
Avoid made-up numbers: use only figures whose original publisher page you've opened. AI-generated answers and secondary articles can carry unsourced numbers, or cite a source that doesn't contain them. Cite the publisher, title, date and what the number counts, and give a range where publishers disagree. A band with a basis beats a precise number with none.