Edition 001 · Revision 1 · 2026-Q3

The Demand Score Index

The Demand Score rates how attractive a software category looks for building a new vertical product, combining existing buyer adoption, competitive whitespace and buyer discoverability into a 0-100 score. This edition covers 68 categories serving 94 US job titles. Of these, 14 score 70 or above and 11 also carry enough evidence to be called BUILD, led by laboratory information management system software at 77.

The baseline

This is edition one, so there is no trend to report. These numbers are the baseline every future edition is measured against.

Distribution of Demand Scores across 68 categories, in ten-point buckets. Most of the index sits between 40 and 70, which is the point: on this evidence most categories are not worth building for.

The scorecard

A category’s score is the highest of the occupations it serves, not the mean. Reach sample is the number of postings the reach component was actually computed from, capped at 25 and taken from the one occupation that supplies the score. It is not the category’s total posting count, which is larger and would overstate the evidence. Anything under 10 is flagged ⚠ and cannot be labelled BUILD, however high it scores.

18 of 68 categories have a reach sample under 10, marked ⚠, and are not eligible for BUILD. 11 of those fall under 5, where reach cannot be computed at all and is set to 0.

3 categories score 70 or above but are held below BUILD by the evidence gate: Market research software (73, n=6), Child care software (70, n=7), Law practice management software (70, n=5). Their scores stand; what we will not do is call them BUILD on that much evidence. Finding 5 shows why.

68 scored here, 58 with a page on the site. Every category in this report is scored and included in the dataset. 10 of them have no per-category breakdown page, because we could not gather enough about what the job involves or which tools serve it to fill one: law enforcement, logistics, real estate transaction management, IT management, membership management, jail management, lawn care, financial management, financial planning and real estate software. 9 of those 10 are also in the flagged thin-sample group above. They are held back from the site, not from the index: their scores are in the table above and in the CSV.

What we found

1. We excluded 55% of our own complaint evidence

Of 1,786 complaints collected about the software these jobs use, 1,065 were excluded because they referred to software outside the category: patient-booking apps filed under electronic medical records, a consumer credit app under loan servicing, business magazines under business management. 721 remain. 17 of 68 categories lost every complaint they had.

Unit: complaint records. Window: collected July and August 2026. Rule: a record is excluded when the product it reviews is not the software a practitioner in that category uses for that work. Classification was manual, at the product level rather than the record level, across 149 product/category pairs. Reversible: the exclusion is a flag, not a delete.

This is why complaints do not feed the score. They are shown as evidence on category pages and they carry no weight in the number.

2. Star ratings cannot tell you whether software is bad

Every one of the 68 categories averages between 4.25 and 4.71 stars, a coefficient of variation of 0.021. Ratings are compressed at the top in this dataset. Review solicitation is one plausible contributor, but these data do not identify the cause. Either way, a central judgement for a prospective founder, whether an incumbent is weak enough to displace, is one this dataset cannot support.

Unit: category. Denominator: 68 categories. Source: Capterra products with 50+ reviews. We tested star ratings as a scoring component and rejected them for exactly this reason.

3. Job postings almost never name the software

41 mentions across 1,808 job postings. 60 of 68 categories never name their own incumbent tool once, even where that tool has thousands of reviews.

Unit: category, and separately job posting. Method: match Capterra product names with 50+ reviews against the requirements field of each posting. An earlier draft of this finding said "63 of 74 categories"; 74 was the count of occupations, not categories, and the two universes had been mixed. Corrected before publication.

4. Observed advertiser cost-per-click, by category

From $7.68 to $219.27 a click, mean $55.34. The most expensive buyer in the index is compensation management software.

These are Google advertiser estimates, not an acquisition-cost model. They are the only modelled figures in this report, they are not part of the Demand Score, and they should not be read as what it would cost you to acquire a customer.

5. A small sample does not just add noise to reach, it inflates it

Reach is distinct employers divided by postings. More postings means more chances to see the same employer twice, so that ratio falls as the sample grows even when the underlying hiring spread is identical. Across the index the observed ratio drops from 0.861 at samples of 5 to 9, to 0.760 at a full 25. Gini-Simpson, which measures the same idea but does not drift with sample size, moves only from 0.942 to 0.928: the hiring spread is not really different, only the estimator is.

Within the 15 categories holding a full 25-posting sample where every posting names its employer, exact rarefaction estimates the upward bias introduced by scoring a smaller subsample: how would those same categories have scored on less data?

reach points gained purely from a smaller sample, out of 30

  n=5   +4.9      n=13  +2.1
  n=7   +4.0      n=20  +0.8
  n=10  +3.0      n=25   0.0

A category with 5 postings collects roughly 4.9 of 30 reach points for free. That is why BUILD now requires a reach sample of 10 or more. It is a bias with a known direction, not an error bar, and it cannot be fixed by widening a confidence interval.

Method: exact Hurlbert rarefaction, the expected number of distinct employers in a subsample of m drawn from the observed n. No resampling. A naive bootstrap with replacement was tried first and was unsuitable here: it manufactures duplicate-employer draws, which biases a distinct-count downward hard enough that the intervals it produced did not contain their own point estimates. Rarefaction is exact for this quantity, so we used it instead.

6. 11 of 68 categories clear both the threshold and the evidence gate

14 categories score 70 or above. 3 of them are held back by the evidence gate from finding 5, leaving 11. Led by Laboratory information management system software at 77, then Fire department software and Procurement software at 76 and 76.

BUILD is a label on a continuum, not a binary truth. On score alone it would be 31 categories at threshold 65; 14 at 70; 5 at 75. We publish the whole curve so you can set your own bar, but we apply the evidence gate at every threshold.

7. Where the work is still manual

Spreadsheet, data-entry and manual-process mentions in job postings range from 0.76 for First-Line Supervisors of Production and Operating Workers down to 0.00 in 15 categories.

Published with a caveat, because we tested it and it does not mean what it looks like. It separates desk work from clinical work, not good incumbents from bad. A dentist does not mention spreadsheets because their hands are busy, not because their practice software is good. We rejected it as a scoring component for that reason and include it here as an observation only.

How the Demand Score is calculated

Three capped components summing to 100. Each is a count of stored records, normalised against fixed bounds rather than live percentiles, so a score keeps its meaning as the index grows.

Review volume is evidence that eligible products accumulated reviews, which is weaker than proof that people buy. The technical name says the weaker, truer thing.

n = min( 25, postings )

adoption   = round( 35 x clamp01( (ln(max(specific_review_total,1)) - 4.0) / 5.5 ) )
whitespace = round( 35 x ( 1 - min(1, established_tools / 32) ) )
reach      = 0                                                    when n < 5
             round( 30 x clamp01( (distinct_employers/n - 0.2) / 0.76 ) )   otherwise

score      = adoption + whitespace + reach
verdict    = BUILD when score >= 70 AND n >= 10

specific_review_total sums Capterra review counts for products appearing in 3 or fewer categories. Generic products are excluded because unfiltered volume measures how horizontal a category is rather than whether its practitioners buy: QuickBooks carries 20,665 reviews against PioneerRx 92. established_tools counts products with 100+ reviews. reach uses the 25 most recent postings. When fewer than 5 are available, the pre-normalisation employer ratio is set to its floor of 0.2, which normalises to exactly 0: that is the same rule the formula above states as its own case, not a second one. Each component is rounded before summing. A category score is the strongest observed opportunity among the occupations it serves, not the average condition across the category, and categories mapped to more occupations have more chances to surface a high one. The dataset names the occupation that produced each score so you can check which. n is that same occupation’s sample, never the category total. The 10-posting gate applies to the verdict only, never to the score: a gated category keeps its number and its rank, it just does not get the label.

Does the answer survive different weights?

The weights are a declared choice, not a discovered one, so we tested what happens when they move.

The ranking is robust; the binary verdict is not. Rank correlation against equal weights is 0.997, so the order is not an artefact of the weights we picked. But the threshold is a soft cutoff: moving it five points either way changes the count substantially, and a five-point weight shift moves several categories across it. Read the score and the rank; treat BUILD as a label on a continuum. The evidence gate is the one part of the verdict we are not willing to soften, because it corrects a measured bias rather than expressing a preference.

What this data cannot tell you

• Whether an incumbent is weak enough to displace. A central judgement for anyone deciding what to build, and star ratings cannot support it (finding 2). We have no clean measure and we are not going to invent one.
• Market size or willingness to pay. The score is a comparative proxy. A 77 is higher than a 62; it is not a dollar figure.
• Anything about your ability to execute. No score knows whether you can build it, reach the buyer, or want to.
• Reach on thin evidence. 18 categories have a reach sample under 10, flagged ⚠ and barred from BUILD. Their reach number is still published and still inflated (finding 5); the gate stops it becoming a verdict, it does not repair the estimate.
• Confidence on the other two components. Adoption and whitespace are counts of every eligible record, not samples, so there is no sampling interval to quote. Their uncertainty is classification error instead, which is what finding 1 is about.
• Anything longitudinal. This is edition one. There is no trend, and we will not manufacture one.

Cite this report

Changelog

• Edition 001, 6 August 2026. First edition. Establishes the baseline: mean 59.2, range 12 to 77, 14 of 68 scoring 70 or above and 11 called BUILD. The score moved from five components to three on 5 August 2026, so earlier figures are not comparable.
• Evidence gate added, 7 August 2026. BUILD now requires a reach sample of 10 or more as well as a score of 70. Rarefaction showed the reach component is inflated by roughly 4.9 of 30 points at a sample of 5 (finding 5), so 3 categories that score 70+ are no longer labelled BUILD. No score changed. The scorecard now reports the reach sample rather than the category posting total, which was larger and overstated the evidence.