Industry Research and Benchmarking

The first real decision in a benchmark is not which metrics to pull; it is which companies belong in the comparison at all. A peer set is a small group of companies close enough in business model, end markets, and scale that setting them side by side is meaningful. Pick companies that make similar products for similar customers at a similar size, and the comparison describes something real. Pull in whatever a keyword search returns, and the table can look complete while it quietly compares a niche specialist to a diversified conglomerate.

Peer selection is a judgment the analyst owns and should be ready to defend. A common practice is to start from the company's own description of its competitors, the operating segments it reports, and the industry classification it files under, then narrow to the handful that share its economics rather than just its label. Size matters because scale changes cost structure and margin; end market matters because a components maker selling into aerospace faces different economics than one selling into consumer goods. The peer set is where a benchmark most often goes wrong before a single figure is pulled, so it belongs on paper, chosen and justified, up front.

Once the peer set is fixed, each metric needs one shared definition applied across each company. Gross margin, revenue growth, and research intensity each sound self-evident, yet each can be computed more than one way, and a benchmark is only honest when the same formula runs on inputs that mean the same thing. This is the analyst's version of a principle the FASB names directly: comparability is one of the enhancing qualitative characteristics of useful financial information in Concepts Statement No. 8, the quality that lets a user identify genuine similarities and differences rather than artifacts of how each company chose to report.

The catch is that comparability is a property of the disclosures, not something a formula can manufacture. If one company breaks out research and development on its own line and another folds it into a broader cost-of-sales or engineering line, the two R&D-intensity figures are built from differently defined inputs, and dividing by revenue does not make them comparable. The disciplined analyst reads each filing closely enough to see whether a line means the same thing across the set, applies a caveat where it does not, and keeps a cited figure (a fact) separate from a reasoned reading of what the figures imply (an inference). Comparable peers, shared definitions, and honesty about where a disclosure does not line up are what make an analysis trustworthy.

The shape of this work is source-bound reading and repeatable arithmetic, with the judgment concentrated in a few decisions made up front. That profile is exactly where AI can help, and also where a fluent tool can quietly do harm, by inventing a peer, filling a figure that sits in no filing, or treating a bundled line as if it were comparable. Part Two is about drawing that line: what to hand the tool, and what to keep for yourself.

The destination is a one-page benchmark memo where each figure is cited and each claim is tagged. What anchors "good" here is the metric definition sheet plus, ideally, one prior benchmark you trust as a worked example. The definitions pin down exactly how gross margin, revenue growth, and R&D intensity are computed, so the model is not free to improvise a formula. With those in the folder, you are not asking the tool to guess what a benchmark should contain; you are asking it to fill a known shape from a known source.

Mapping the journey also means naming the peer set explicitly at the start. Four companies chosen for a reason (similar products, similar scale) make a defensible comparison; four companies pulled in because a search returned them do not. The peer set is the first place a benchmark can go wrong, so it belongs in the map, written down, before extraction begins.

Research work carries two splits worth separating. The first is the familiar one. Computing a metric from disclosed figures is deterministic: for a given revenue and gross profit there is one gross margin, and for a given R&D expense and revenue there is one R&D intensity. That arithmetic belongs in a formula or template, computed once and validated once, rather than produced by a language model in prose. The model's job is the non-deterministic part: reading the pattern across companies and writing the comparison in a measured voice.

The second split is specific to citing sources, and it is the heart of this module. Each statement in a research memo is either a fact or an inference. A fact is a figure or quotation traceable to a source line: Halbrook's revenue was 340.0 million dollars, cited to its income statement. An inference is a reasoned claim the analyst builds on top of the facts: that Halbrook's lower R&D intensity may reflect a more mature product line. The discipline of fact versus inference is to tag each claim as one or the other, so a reader can see where the filing ends and the analyst's reasoning begins. Blurring the two is how a plausible guess ends up quoted as if a company had disclosed it.

The lab folder holds two things: the excerpted peer filings and the metric definitions. Nothing else. In research, the folder is the set of sources the memo is allowed to cite. If a figure is not in the folder, it is not a fact the memo can cite, which keeps the tool from wandering onto the open web and pulling in a number no reviewer can trace. Minimum context is again the fuel, and here it doubles as a control on what counts as evidence.

This is why the extract-verify-cite pattern starts inside a scoped folder. When the citable universe is small and known, verification is fast: each figure either resolves to a line in the folder or it does not. Widen the folder to the whole internet and verification becomes the slow, error-prone part, which is exactly where fabricated citations tend to survive.

Make the two splits concrete with the R&D column. Halbrook discloses research on its own line: R&D of 17.0 million on revenue of 340.0 million. R&D intensity is a deterministic calculation, so there is one answer: 17.0 divided by 340.0 is 0.050, or 5.0%. That figure is a fact, cited to Halbrook's income statement, and it is directly comparable to any peer that also breaks research out on a separate line.

Meridian looks like it offers the same calculation, and that is the trap. Its filing shows 8.6 million of research, but the number sits inside a broader "Engineering and product costs" line that also carries non-research spending. Run the same arithmetic and you get 8.6 divided by 214.0, or 4.0%, a figure that reads as if Meridian reinvests less than Halbrook's 5.0%. The 4.0% is real arithmetic on a number the filing discloses, which is exactly what makes it dangerous: it is not a research-only intensity, because the numerator bundles research with other costs, so the true research share could sit above or below Halbrook's 5.0% and the filing gives no way to separate it.

The comparability caveat is concrete here: it is the difference between reporting 4.0% and flagging the cell. The honest benchmark shows Halbrook at 5.0% as a comparable fact and marks Meridian's R&D intensity as not directly comparable, with a one-line note that its research is bundled. Two numbers that come out of the same formula are not comparable when they are built from differently defined inputs, and that single judgment decides whether the R&D column informs the partner or quietly misleads them.

The three moves become repeatable when you capture them once as scaffolding rather than retyping them each quarter. Two artifacts carry most of the weight. The first is a reusable prompt or skill: a short, saved instruction that names the peer set, points the tool at the scoped folder, fixes each metric definition, and states the rules of the house (compute nothing in prose, cite each figure to a source line, tag each claim as fact or inference, and flag any metric that is not directly comparable rather than estimating it). Written once, it makes each future benchmark start from the same disciplined shape instead of a blank box.

The second is a verification checklist the analyst runs on the draft, because the prompt sets the intent while the checklist confirms the result. A workable one is short: does each figure resolve to a line in the folder when you open it; was each metric computed on the stated definition; is each inference labeled as an inference rather than dressed up as a disclosure; is each comparability caveat surfaced rather than filled with an estimate; and does the memo lead with a takeaway a reader can act on. The prompt and the checklist are the two ends of the same loop, and Parts Three through Five turn them into the pattern, the lab, and the validation step you will actually run.

Download the folder and run the extract-verify-cite pattern in whatever AI you use. The folder holds excerpted figures for Meridian and three fictional peers, plus a metric definition sheet that fixes how gross margin, revenue growth, and R&D intensity are computed. Build the benchmark, cite each figure to a source line, tag each claim as fact or inference, and then come back to verify. One peer bundles its research spending into a larger line and another does not disclose it separately, so part of the exercise is deciding where a metric is comparable and where it is not. The companies and the numbers are fictional.

The heart of validation is verifying that each citation resolves. Take each figure in the benchmark and open the source line it points to; the number should be there, unchanged. A single figure that does not resolve is enough to hold the whole memo back, because it means the extract-verify-cite loop was not closed. When time is short, verify in order of consequence: open the figures that carry the takeaway first, since a wrong number there changes the reader's decision, while a wrong number in a supporting cell mostly costs credibility. Confirm too that each metric was computed on the stated definition. Each inference should be labeled as an inference rather than dressed up as a disclosure, and any comparability caveat surfaced rather than papered over with an estimate. Work the checklist below against your draft before you would call it done.

The short memo below is what you would put in front of the partner. It leads with the takeaway, tags each interpretive claim as an inference, and surfaces the comparability caveat instead of hiding it in a footnote.

Takeaway. Across the four-company set, revenue growth clustered in the mid-single digits, from 4.6% at Trellis to 7.4% at Cardan (fact), and gross margins ranged from 30.0% at Trellis to 36.0% at Halbrook (fact), with Meridian and Cardan both at 33.0% (fact). The growth spread is narrow, so the companies look broadly similar on that measure (inference).

Growth and margin. Cardan grew fastest at 7.4% (505.0 million from 470.0 million) and Trellis slowest at 4.6% (158.0 million from 151.0 million) (fact). Halbrook carried the widest gross margin at 36.0%, six points above Trellis at 30.0% (fact). Halbrook's margin lead may reflect scale or product mix (inference), which the filings in the folder do not settle.

R&D intensity, where it is comparable. Among the two peers that disclose research separately, Trellis reinvests a higher share of revenue than Halbrook: 6.0% against 5.0% (fact, then a direct comparison of two disclosed figures). Meridian's research spending sits inside "Engineering and product costs" at 8.6 million, so its intensity is not isolable and is flagged rather than reported (fact about the disclosure; the caveat is the analyst's call). Cardan does not disclose research separately, so its intensity is left blank rather than estimated. On the figures that are comparable, Trellis looks like the most research-intensive of the set (inference, limited to the two peers that disclose it).

The lab is tidy by design: four companies, one definition sheet, and disclosure differences that are easy to spot. Real peer sets are messier, and the comparability caveat has to work harder. Three failure modes recur:

• Fiscal years that do not line up. One peer closes its year in December and another in June, so each company's "latest year" covers a different twelve months; lining up their growth rates compares two windows that a seasonal quarter or a macro shock can move independently.
• A metric defined three ways across a sector. "Gross margin" for a company that books distribution and warranty inside cost of sales is not the same margin as a peer that books those in operating expense, so two identical-looking 33.0% figures can describe different cost structures.
• Adjusted figures sitting next to reported ones. A tool pulling from earnings releases can grab a company's non-GAAP "adjusted" margin for one peer and the GAAP figure for another, and the two read as the same metric while excluding different items; the same trap appears when a peer restates a prior period, so last year's number in this year's filing no longer matches what that company reported a year ago.

The comparability caveat is not a one-time flag; it is a habit of asking, for each cell, whether these two numbers were built the same way. The workflow does not remove that judgment, and a fluent tool can make it harder by presenting mismatched figures as a clean row. That is why verification opens the source rather than trusting the citation, and why an undisclosed figure is flagged rather than filled. The pattern speeds up the extraction and the arithmetic so the analyst's attention is free for the question that decides the memo's quality: are these companies, and these numbers, genuinely comparable?

Rate your own benchmark against the rubric below. An honest score against the rubric shows which parts of the extract-verify-cite pattern are solid and which still need practice. Your scores roll up to the workflow maturity dashboard on the course hub.