You do not know from a single percentage. “Market standard” is a base rate — the share of agreements in some defined corpus that carry a term — and that number moves with the corpus far more than it moves with the market. The same indemnity cap reads as ordinary against SEC-filed material contracts, aggressive against your own signed paper, and unheard-of against a vendor panel built from unnegotiated order forms. Four things have to be on the table before a benchmark changes a position: the corpus definition, the denominator for that specific clause, whether the agreements in it were negotiated, and the width of the confidence interval. This page is that test, with 2026 numbers to calibrate against.
When to use this
Use it when someone hands you a percentage and asks you to concede, hold, or escalate on it — a counterparty citing “market”, a benchmarking feature inside your drafting tool, an annual deal-points study, a partner’s recollection of the last twelve deals. It is a test for the number, not for the clause. Whether a term is acceptable is a risk question your own playbook and counsel answer; this page only establishes whether the number in front of you describes a market at all.
The three corpora, and what each one actually measures
| Corpus | What it measures | What it cannot tell you |
|---|---|---|
| Your own executed agreements | Outcomes you have accepted, under your bargaining position, counterparty mix, and playbook | What anyone outside your counterparty set agrees to |
| Public filings (EDGAR) | Non-ordinary-course contracts material enough that a public registrant had to file them | Anything about routine commercial paper, or unredacted economics |
| Vendor aggregate panels | The distribution across that vendor’s customer base and their counterparties | Whether that customer base resembles your segment |
Your own paper is the corpus most teams reach for first, and it is the one that answers a different question than the one asked. It lives in the CLM — Ironclad, LinkSquares — and it reports what you have signed. If you have conceded uncapped IP indemnity forty times, your internal base rate is 40/40, and the honest label for that finding is “our position,” not “the market.”
Public filings carry a structural exclusion that most readers never see. Item 601(b)(10)(i) of Regulation S-K defines a material contract as one made outside the ordinary course of business and material to the registrant; a contract that ordinarily accompanies the registrant’s business is deemed ordinary-course and does not have to be filed unless it falls into one of the enumerated categories. Since the 2019 FAST Act amendments, registrants can also omit information from a filed exhibit without a confidential treatment request where that information is immaterial and would likely cause competitive harm. Your standard MSA, DPA, and order form are excluded by definition, and the caps and pricing inside what remains are the parts most likely to be redacted. DraftWise pairs a firm’s own precedent bank with an EDGAR integration precisely because neither side is sufficient alone.
Vendor panels measure the vendor’s book of business. Spellbook shipped Compare to Market to general availability in January 2026, publishing 270+ clause benchmarks across 13 contract types drawn from hundreds of thousands of contracts across 30 countries, assembled on a stated give-to-get model in which customers share anonymous aggregate statistics in exchange for access. That is a real corpus and a defensible one. It is also, by construction, a census of Spellbook’s customers rather than of the market, and the two coincide only where your segment matches theirs.
Test 1 — what is the denominator for this clause?
Corpus size is not clause size. A panel of 200,000 agreements that contains 300 data processing addenda gives you a DPA benchmark with n = 300, and the vendor’s marketing number is the one you will be shown. Guard: require the per-clause n before acting. A benchmarking feature that reports a percentage without a denominator is reporting an opinion in the format of a statistic.
Test 2 — negotiated paper, or form paper?
TermScout’s February 2026 termination analysis is unusually honest about this and worth borrowing as a template: it discloses a corpus of 786 contracts split into 591 vendor forms, 101 customer forms, and 94 negotiated agreements. Vendor forms are 75% of the corpus and negotiated agreements are 12%.
A base rate computed across all 786 is therefore mostly a distribution of opening asks. It tells you what vendors put on the table, which is useful for predicting a counterparty’s first draft and useless for defending a landed position. Only the negotiated subset describes where deals actually close, and that subset is 94 contracts.
Guard: for any position you intend to hold, ask for the negotiated-only cut. If the vendor cannot produce one, the number describes drafting convention, not market outcome, and it goes in the “context” column rather than the “evidence” column.
Test 3 — how wide is the interval?
This is the part that gets skipped, and it is arithmetic, not judgment. The 95% Wilson interval for a proportion near 50% — the worst case, and the case where “is this standard?” is genuinely contested — behaves like this:
| Clause-level n | 95% interval at a 50% finding | Half-width |
|---|---|---|
| 12 | 25.4%–74.6% | ±24.6 pp |
| 30 | 33.2%–66.8% | ±16.8 pp |
| 50 | 36.6%–63.4% | ±13.4 pp |
| 94 | 40.1%–59.9% | ±9.9 pp |
| 139 | 42.2%–58.5% | ±8.2 pp |
| 400 | 45.1%–54.9% | ±4.9 pp |
| 786 | 46.5%–53.5% | ±3.5 pp |
Two thresholds fall out of that curve, and they are the calibrated values worth memorising: n = 93 is where the half-width reaches ±10 points, and n = 381 is where it reaches ±5 points. Findings far from 50% tighten — an 80% finding at n = 139 runs 72.4%–85.7%, ±6.6 points — so treat the table as the ceiling on your error, not the estimate.
The operating bands:
- Under 30 — anecdote. Use it to form a hypothesis, never to concede a term.
- 30 to 92 — direction only. Enough to know which way the market leans, not enough to quote.
- 93 to 380 — enough to hold a position, with the interval stated alongside the point estimate.
- 381 and up — a number you can put in a memo without an asterisk.
Test 4 — did the inclusion criteria move?
Year-over-year benchmark comparisons assume a stable population. Study sponsors change their sampling frames, and the change is disclosed in the methodology note that nobody reads.
The worked example: what “earnouts fell” is worth
The ABA Business Law Section’s Private Target M&A Deal Points Study is the strongest public benchmark in its category, and it makes the point better than a weak one would. The 2025 edition, announced in December 2025, covers 139 definitive acquisition agreements for transactions executed or completed during calendar 2024 or Q1 2025 — public buyer, private target, material enough to require SEC disclosure, purchase prices from $25 million to $900 million with a majority under $200 million. It reported earnouts in 18% of deals, down from 26% in the 2023 study. The 2023 edition covered 108 agreements from 2022 and Q1 2023, at purchase prices from $30 million to $750 million.
Run the arithmetic on the five headline moves. Treating the two editions as independent samples, the two-proportion test gives:
| Deal point | 2023 (n = 108) | 2025 (n = 139) | Difference | 95% interval on the difference | p |
|---|---|---|---|---|---|
| Earnouts | 26% | 18% | −7.9 pp | −18.4 to +2.5 pp | 0.13 |
| RWI referenced | 55% | 63% | +8.7 pp | −3.7 to +21.0 pp | 0.17 |
| No survival of reps | 30% | 41% | +11.4 pp | −0.5 to +23.3 pp | 0.065 |
| Alleged-breach indemnity | 17% | 27% | +10.7 pp | +0.5 to +20.9 pp | 0.047 |
| Double materiality scrape | 69% | 82% | +12.6 pp | +1.8 to +23.4 pp | 0.021 |
Three of the five moves are indistinguishable from sampling noise at conventional thresholds. Two clear an unadjusted 5% threshold — and with five comparisons drawn from one study, a Bonferroni correction sets the bar at 0.010, which neither reaches. That is before accounting for the sampling frame moving from $30–750 million to $25–900 million between editions, which makes even the independent-samples treatment generous.
None of this makes the study bad. It is the best available instrument for mid-market private M&A, and n = 139 is close to the practical ceiling for hand-coded deal points. It makes the headline bad. “Earnouts are down eight points” is a sentence about 139 agreements that supports no change to any negotiating position, and the study’s own committee is more careful about this than the client alerts summarising it.
Watch-outs, each with its guard
Your own base rate presented as the market. An internal corpus reflects your bargaining position and your counterparty mix, so it will always describe you. Guard: store internal base rates in a separate field from external ones, labelled “our position,” and never let a single field hold both. Clause library design covers the schema.
A vendor panel that does not match your segment. A panel weighted toward $20K-ACV SaaS says nothing about a $4M enterprise deal, and the aggregate hides it. Guard: demand the cut by contract value, industry, and governing law; if the vendor cannot cut it, downgrade the finding to direction-only regardless of corpus size.
Quoting an economic term from public filings. The dollar figures inside filed exhibits are the ones registrants are most entitled to omit under the 2019 amendments, so the filed set skews toward deals where the number was not worth protecting. Guard: use public filings for structural terms — presence, scope, survival, carve-outs — and treat any filed cap or pricing distribution as a lower bound on real dispersion.
Treating “market” as a legal standard. A base rate is descriptive. It does not establish that a term is enforceable, reasonable, or adequate for your risk posture, and in a regulated context it establishes nothing at all. Guard: keep the benchmark and the risk assessment in separate columns of the playbook, and route jurisdiction-specific questions to counsel rather than to the percentage.
Silent recomputation between refreshes. Vendors expand and reweight panels continuously, so the same clause can move four points because the corpus grew. Guard: record the corpus date and per-clause n alongside every benchmark you write into a playbook, and treat an undated benchmark as expired.
When this framework breaks down
It breaks on new clause types. AI training and model-improvement restrictions, output-ownership terms, and zero-retention fallbacks entered mainstream vendor paper too recently for any 2026 corpus to hold a stable base rate, and a benchmark drawn from agreements signed before the term was common will report scarcity that has already reversed. It breaks where governing law dominates the drafting convention, since a global aggregate averages across regimes that do not average. And it breaks when you are large enough to set the terms in your own supply chain, where the base rate you should care about is the one your own paper is creating.
In all three cases the substitute is the same: build the corpus you need out of your own executed agreements plus the counterparty drafts you have received, extract the clause systematically rather than by memory, and hold yourself to the same denominator and interval discipline you would demand of a vendor. The clause extraction Skill does the extraction pass, and playbook from precedent turns the resulting base rates into positions with fallbacks attached. For how those positions get used at the table, see the MSA negotiation prompt pack; for how the same discipline applies to a vendor’s accuracy claims rather than its market data, see AI contract review accuracy and contract risk scoring.