A managing partner asked us why his firm’s best page never shows up in AI answers. It is a 2,400 word explainer on comparative negligence, it holds the top organic position in his market, and it reads well. Nothing is wrong with the page. The problem is that the page was never the unit of competition in the first place.
Retrieval systems do not read your article the way a prospect does. They split documents into passages, convert each passage into a vector, rank passages against the question, and hand a small set of surviving passages to a language model that writes the visible answer. What gets quoted is a passage, not a page. A paragraph that only makes sense after reading the four paragraphs above it arrives at the model stripped of that context, and it loses to a paragraph from some other site that stood on its own.
This is not a new or speculative mechanic. The retrieval research has treated the passage as the working unit for years, going back to work like Dense Passage Retrieval for Open Domain Question Answering, which showed that chopping documents into short passages and matching them individually outperformed document level keyword retrieval by a wide margin. Every consumer AI search product built since inherited that shape. If you want to understand the surrounding landscape first, we covered the mechanics of how law firms get cited by ChatGPT and LLMs and the SERP side in AI Overviews and SGE for law firm SEO. This piece is narrower. It is about how the text itself has to be built.
What a Retrieval System Actually Grabs
The chunking step is the part most legal marketers have never pictured. In production systems the common configuration is a splitter that cuts on structural boundaries first, headings and paragraphs, then falls back to a character or token limit somewhere in the range of 200 to 800 tokens per chunk, often with a small overlap of 10 to 20 percent so a sentence cut in half survives in one of the two neighbors. Open source splitters that most teams have used at some point, such as the recursive character splitter shipped in LangChain, default to exactly this behavior.
Two consequences follow, and they are the whole argument of this article.
- Your 2,400 word explainer is not competing as one document. It is competing as roughly eight to fifteen independent fragments, each of which will be judged alone.
- Only a handful of fragments, commonly three to eight across all sources, make it into the context the model writes from. Everything else on your page is invisible for that answer.
So the practical question stops being how good is this article and becomes which single block of 150 words on this page would I be happy to see quoted with no other sentence attached. If you cannot point to one, the page will not be quoted, however well it ranks.
Why the First Paragraph Under Every Heading Has to Answer
Legal writing is trained in the opposite direction. Lawyers build to a conclusion because that is how a brief persuades. Answer first content inverts it. Under every heading, the first two or three sentences state the answer plainly, and the reasoning, the exceptions, and the statutory detail follow.
Compare two openings under a heading about the filing deadline for a car accident claim. The first runs, the statute of limitations is one of the most misunderstood parts of personal injury practice, and clients are often surprised to learn how it works. The second runs, in Texas, most car accident lawsuits must be filed within two years of the crash date under Section 16.003 of the Civil Practice and Remedies Code. Missing that deadline usually ends the claim regardless of how strong the facts are.
The first opening is not wrong. It is simply unquotable. It contains no answer, so no retrieval system has any reason to lift it, and a reader scanning on a phone learns nothing from it either. The second opening is a complete answer with a jurisdiction, a number, a statutory anchor, and the consequence, in two sentences. That block can be extracted, cited, and shown to a stranger without a single sentence of surrounding page.
The tradeoff is real and worth naming. Answer first prose reads flatter. It sacrifices some of the narrative build that good legal writing enjoys, and partners who write well often dislike the first draft. The compromise we use is that the first block under each heading is written for extraction and the rest of the section is written for the human reader. Nobody has to give up voice past sentence three.
What Makes a Section Self Contained
Self contained means the passage does not depend on anything above it to be correct and understandable. In practice, four habits carry almost all the weight.
- Repeat the subject instead of using pronouns across boundaries. A section that opens with the deadline is shorter for these claims is broken the moment it is lifted. Name the claim type again. Redundancy that feels clumsy in a linear read is what makes a fragment survive on its own.
- Restate the jurisdiction in the block. Not once at the top of the article. In each section where the rule is state specific.
- Keep the answer above any list. Bulleted lists get chunked cleanly and are quoted often, but a list with no preceding sentence of context is orphaned. One framing sentence, then the list.
- Cap each section at roughly 150 to 300 words before the next heading. That range tends to land inside a single chunk. Sections that run 800 words get split at an arbitrary point, and the back half becomes a fragment with no anchor.
A useful test costs nothing. Copy any single section out of your draft into a blank document, read it as if you found it on a page you have never seen, and ask whether it answers a real question correctly. Sections that fail that test are the ones being ignored.
Why Jurisdiction and Date Belong Inside the Sentence
Legal answers are false without a jurisdiction, and the two places firms usually put it, the page title and a sidebar, are the two places a chunk cannot see. If the deadline is two years in Texas and three years in Maryland, a passage that says the deadline is two years is not a partial answer, it is a wrong one, and it is exactly the kind of passage a careful retrieval system learns to distrust.
The same logic applies to time. Statutory amounts, filing fees, damage caps, and mileage rates change. Write the operative date into the prose, as in, as of the 2026 legislative session, the cap stands at, rather than relying on a byline date the chunk will never carry. Pair that with a visible last reviewed line at the top of the page, and, where the page qualifies, structured data that states the same facts in machine form, which we walk through in FAQ and HowTo schema for law firm pages.
There is a maintenance cost attached and firms should budget for it honestly. Putting explicit dates and dollar figures in prose means somebody has to review those pages on a schedule. We plan on a twice yearly review pass for statutory pages and an immediate pass whenever a client alerts us to a legislative change. A firm unwilling to fund that review is better off writing rules qualitatively than publishing a specific number that quietly goes stale, because a confidently wrong quoted passage is worse than no citation at all.
Should Headings Be Phrased as the Question
Mostly yes, and for a mechanical reason rather than a stylistic one. Splitters cut on headings, so the heading is usually the first line of the chunk and travels with it. A heading that reads Statute of Limitations Considerations contributes nothing to the match. A heading that reads How Long Do You Have to File a Car Accident Claim in Texas states the query the passage answers, in the words a person would use.
Some discipline keeps this from turning into keyword mush.
- Use the phrasing a client would actually type or speak, not the phrasing that appears in a pleading.
- One question per heading. Compound headings produce compound chunks that half answer two things.
- Do not force every heading into a question. Process and comparison sections often read better as noun phrases, and a page of twelve consecutive questions looks like a template.
- Keep the wording distinct across the site. Two pages with near identical question headings compete with each other for the same passage slot.
How to Retrofit Pages You Already Have
Nobody needs to rewrite a content library. The realistic scope for a firm with 80 to 150 published pages is a first pass on the 15 to 25 pages that answer genuine questions and already earn impressions.
- Pull every page with its H2 and H3 text using a crawler such as Screaming Frog, and export it to a sheet.
- Cross reference Search Console for the queries each page already receives, filtered to question phrasings and to any query with more than a handful of impressions.
- Flag headings that state a topic rather than a question, and flag sections longer than about 400 words.
- For each flagged section, rewrite the heading as the question and write a two to three sentence direct answer as the new opening block, with the jurisdiction named.
- Split anything oversized into two sections with their own headings rather than leaving one long block.
- Add the operative date and a last reviewed line, then request reindexing.
Budget roughly 30 to 45 minutes per page for a real editorial pass by someone who understands the law, not a find and replace. On a 20 page batch that is somewhere between two and three working days. It is one of the few content projects where the work is bounded, the changes are visible in the draft, and the attorney review burden is light because you are restructuring existing approved language rather than making new claims.
What This Changes and What It Does Not
Be clear about the mechanism. Structure decides whether your passage can be used once your page is already in the retrieval candidate set. It does not put you in that set. Nothing in this article substitutes for the things that get a page considered at all, which are still topical depth, a technically sound and fast site, real citations and mentions elsewhere, and a firm that is genuinely known for the subject.
Two more limits worth stating plainly. First, answer first structure does not make you the source an assistant prefers when three firms all did the work well. At that point the tie breaks on authority signals you influence slowly, not on formatting. Second, none of this is trackable the way rankings are. AI assistants do not send referrer data consistently, answers vary between users and sessions, and there is no impression report for being quoted. The honest measurement approach is a fixed panel of 25 to 40 real client questions, tested manually across the major assistants on a monthly cadence, recording whether your firm appears and what it appears next to. That is a directional read, not a metric, and anyone selling you a precise number for AI citation share is selling you a model of a thing nobody currently measures directly.
What you do get is durable. Passages that stand alone also read better on a phone, convert better because the answer arrives before the reader loses patience, and hold up as retrieval implementations change, because every version of this technology has to cut documents into pieces. The firms that will be quoted in two years are the ones writing text today that survives being cut.
Rubiks rebuilds law firm explainers for extraction as part of content work, starting with the pages that already earn impressions rather than a full library rewrite. Talk to Rubiks and we will show you which of your existing pages are one editing pass away from being quotable, and which ones are not worth the hour.