Competitor SEC Filing Analysis Prompts

A ten-K runs to a hundred thousand words and most of them are carried forward from last year. Whole risk factors are industry standard text. Paragraphs of it could be filed by any software company in the world with one name swapped.

That is a problem for a summariser specifically. Standard text is what a language model produces most fluently. So a summary of a filing reads well and sounds authoritative. It also says nothing that would be false of a competitor you have never heard of.

The right instrument is the one a competitive SWOT already uses: swap the company name and see whether the claim survives. Risk factors are where that test earns its keep most. Run it and most of the section disappears, which is the right outcome. The few survivors are the reason for reading at all.

Two passages in an annual filing are not standard text, and both are unusually valuable. The first is the sentence in which a company says what it is. The second is the passage where it says who it competes with.

Both are written by lawyers and signed by executives who carry liability for them. Nothing else a competitor publishes is produced under that constraint. A marketing page can claim anything; this passage has to be defensible.

The constraint pushes the competition passage in a particular direction, and knowing which direction is the whole skill. Under-disclosing a threat is the expensive mistake, so the list comes out deliberately wide. Categories appear that the company barely meets in a deal. This is the mirror image of the skew on a customer page, where a logo wall is a selection built to flatter.

Read the list as a boundary, not a roster

A company that adds a category to that list has decided something is now close enough to be worth naming. That decision is often made a year before anybody outside notices.

The reverse is worth knowing too. A category that has sat in the list for five years is not evidence of a live threat. It is evidence that removing it would be harder to defend than leaving it in, which is true of most sentences in a filing.

Three passes. The first extracts, the second compares, and the third throws most of the document away. Run them on named sections rather than on a whole filing.

One line in the second prompt is doing more work than the rest of it. Work from the older filing outwards. Comparison naturally runs the other way and finds what is in the new version, and a sentence that has been removed leaves nothing to find.

Two inputs, and the second is the method rather than an extra. Filings are the one competitor source where structure is fixed by rule, so last year’s version of any section is guaranteed to exist and to be in the same place.

Label each paste with its item number. It costs a second and it is what lets somebody sceptical find the sentence. A quote nobody can locate gets treated as a paraphrase within about two forwards of an email.

We asked a straightforward question with nothing supplied, then ran the comparison properly on two consecutive years of one filing. Both filings are public and free to open.

Nine plausible brands, none of them in the document

The first answer is easy to check, and it fails. It names nine companies. The filing names none. It lists twelve categories of competitor and leaves every company unnamed, which is a normal drafting choice and a surprising one if you have not read a filing before.

What makes the answer dangerous is that the brands are the right brands. Anybody who knows the market will read that list, recognise it, and treat the filing as having confirmed something. The document confirmed nothing. Two of the twelve categories it does use are new this year, and neither of them would have occurred to the model.

The sentence nobody compares

The second run found something in the place people skip. A company’s business description opens with one sentence saying what it is, and that sentence had been rewritten. A platform became an agentic platform. The audience narrowed from businesses in general to marketing, sales and customer service teams.

Read that beside the same year’s additions to the competitive factors list, where agents, a unified data platform and answer engine optimisation all appear for the first time. The filing was published in February. The products matching those phrases shipped in the autumn. A weekly competitor digest would have reported the launches; the filing had already reported the intent.

An annual filing arrives once a year, which makes this the slowest source in competitive work. The wording you just compared describes a company as its lawyers saw it on a date eight months ago.

That is not a reason to skip it. It is a reason to read it as the confirmation rather than the alert. A phrase in a filing has usually been visible somewhere faster first, on a pricing page or in a set of job adverts, and the filing is where you find out the company meant it.

Two limits do not move. Filings exist only for companies obliged to make them, so a private competitor has none. And a single reported revenue line covering six products says nothing about the one you meet in deals, however carefully it was audited.

Flares tracks how competitors describe themselves in public and dates every change. When a filing appears in February, the sentences in it can be read against the week each phrase first showed up, which is usually months earlier and on a page nobody was watching.