Legal marketing runs on secondhand information. The average article about SEO for law firms cites other articles about SEO for law firms, which cite others still, and somewhere at the bottom of that stack there is usually no primary source at all, just an early claim that got repeated until repetition started functioning as evidence. I decided several years ago that everything I publish would be built the other way: from primary documents up. This piece is the actual method, in enough detail that you could run it yourself, because the method is also the fastest way to evaluate anyone who wants to advise your firm.
Every claim that enters my work gets classified before it gets used, and the classification decides what I am allowed to do with it.
Tier one is the primary document: the patent as granted, the court opinion, the statute, the regulator's own filing, the government dataset, Google's own documentation, or my own counted data with the spreadsheet to show. Tier-one claims can carry numbers and dates.
Tier two is the reputable secondhand report: a serious publication describing a primary document. Tier two is a POINTER, not a source; it tells me a tier-one document exists and where to dig. I cite the document it points to, never the pointer.
Tier three is industry content: vendor blogs, marketing articles, conference decks, competitor pages. Tier three is useful for exactly one thing, mapping what the industry believes and covers, and is banned from being a source of facts in anything I ship. A claim that exists only in tier three is, by my working definition, folklore until proven otherwise.
That single rule, competitor content tells you about coverage and never about truth, eliminates most of the errors in this industry on its own, because the industry's standard practice is the opposite: read the top ten ranking articles and synthesize them, which is a machine for laundering one unsourced claim through ten retellings.
Primary documents rarely rank. They are PDFs on government subdomains, filings in dockets, spreadsheets on statistics portals, and patents in databases nobody reads for fun, which is precisely why content built from them contains things the ranking articles do not.
So the hunting looks different from normal searching. I query by file type and by domain class, filetype PDF against government and university sites, because agencies publish their best data as documents, not as pages. I read dockets and regulatory filings when a legal-industry claim needs a date and an amount. I pull state court statistics, transportation department crash data, and census tables directly rather than accepting any article's summary of them. And for search mechanics specifically, I work from a semantic index of Google's own patent filings, hundreds of thousands of documents, searchable by concept, so when I say the machinery reads review text or schedules crawling by a formula, I can put a patent number next to the sentence and anyone can pull the document and check me.
Two practical notes from years of this. First, government sites block and break constantly, so the discipline includes archiving what you verify the day you verify it. Second, the gap between what agencies publish and what the industry repeats is enormous; most of the "well-known statistics" in legal marketing cannot be traced into any primary dataset at all.
The most valuable research act in a folklore-heavy industry is negative: taking a widely repeated claim and following its citation trail to the end. The trail ends one of three ways, at a primary document that confirms it, keep it, with its real date; at a primary document that says something DIFFERENT, which is the jackpot, because now you know a thing your whole industry has wrong; or at a circle of blogs citing each other, quarantine it forever.
Real examples from my own builds, each one chased personally. A penalty figure for fake-review violations circulated across the industry at one number; the Federal Register document says 53,088 dollars, and the repeated figure was an out-of-date amount that kept getting copied from old articles. A claim that review stars in search results work a certain way turned out to trace to a Google change announced in 2019, still being described as current years later. FAQ rich results get recommended in legal SEO articles to this day; Google removed them for all sites in May 2026, with the sitelinks search box retired back in 2024. "Deadliest highway" claims about specific interstates, a staple of injury-firm content pages, kept failing against the actual state transportation data when I pulled the county tables. A supposed limit on how many times a business can appeal a removed review appears in dozens of guides and in no Google document I could find.
None of those catches required special access. They required the one thing the content industry refuses to spend: the hour it takes to find the primary document instead of the three minutes it takes to cite another article.
The second half of the method is making primary data where none exists. The questions I most wanted answered, what share of page-one PI firms have complete profiles, what their review footprints actually look like, how their schema is really built, whose attorneys exist as verifiable entities, had no dataset. So I built them: defined a population, pulled every record, counted by hand and script, wrote down the method, and published the result with the denominators attached, the largest being an audit of 1,005 personal injury firm profiles. The studies now live free on my research page with their methods stated, which means anyone, including a skeptic, can check the counting.
Publishing the method matters more than publishing the findings. A number with a visible denominator and a reproducible procedure invites correction, and invited correction is what separates research from content. It is also, frankly, a trust strategy: in an industry where nobody shows their work, showing yours is a moat that costs only honesty.