A legal page structure that AI assistants can actually quote
An answer engine has to get through seven stages before it will cite you. Most professional content is written to satisfy stage three and fails at stage two and stage seven. This is the page structure that clears all of them.
To be cited by an AI assistant, a page must be crawlable, extractable, understandable, relevant, trustworthy, answer-bearing and independently corroborated. Long-form legal content usually fails on extraction and on the self-contained passage: the answer exists but is spread across three paragraphs that only make sense in order. The fix is structural, not stylistic.
The seven stages
Every retrieval system, whether it powers a search result or a chat answer, runs some version of this sequence. Each stage is a filter: fail one and the later ones never run.
| # | Stage | What decides it |
|---|---|---|
| 1 | Can it be crawled? | robots directives, bot access rules, CDN challenges |
| 2 | Can the content be extracted? | text present in the served HTML, not assembled later in the browser |
| 3 | Can the page be understood? | heading structure, one clear subject, unambiguous scope |
| 4 | Is this source relevant? | match between the page’s subject and the question asked |
| 5 | Is it trustworthy? | named author, verifiable credentials, dates, cited primary sources |
| 6 | Does it answer the question? | a passage that stands alone when lifted out of the page |
| 7 | Are there independent signals? | corroboration from sources that are not you |
Professional content tends to be written entirely for stage three. It is well organised, correctly structured, and comprehensively unquotable, because the answer to any given question is distributed across the page rather than located in it.
The structure
The order below is deliberate. Everything a trust decision depends on sits above the fold; everything a citation depends on sits in a single retrievable block near the top.
- H1 — the question, in the reader’s words
Not the topic label. “ILR refused because of absences — what to do next”, not “Indefinite leave to remain: continuous residence requirements”.
- Byline, credentials and dates — immediately under the H1
Author, professional reviewer with registration number, publication date and last-reviewed date. This is stage five, and it belongs at the top: a passage lifted from the middle of a page often loses whatever sits at the bottom.
- Jurisdiction and date of law
“England & Wales, as at 4 September 2026.” Cross-border subjects are where models most reliably conflate two legal systems, and one line prevents it.
- The answer — 40 to 60 words, fully self-contained
The single most important block on the page. It must make complete sense with no page around it: no “as set out above”, no pronoun whose referent is in the previous section, no dependency on the H1 to be intelligible.
- The rule that governs it
Stated once, plainly, with the provision named and linked inline to the primary source.
- Who it applies to, and who it does not
Scope boundaries are what stop a model applying your answer to a case it does not cover — the failure mode that does real damage in legal content.
- Exceptions
Separately headed. Buried exceptions get lifted as if they were the rule.
- The procedure, with its deadlines
Numbered steps only where order genuinely matters. Deadlines stated as durations and as the event they run from.
- Evidence and cost
What a person needs to produce, what it costs, how long it takes. These are the questions people actually ask, and the answers are rarely on the page.
- Common mistakes
The most cited section in practice, because it maps to how people phrase problems rather than how lawyers phrase topics.
- Official sources, inline as well as listed
Link the primary source next to the claim it supports, not only in a block at the end. A stage-seven signal is worth more attached to the sentence it corroborates.
- Related pages and a single clear next step
One action, named for what happens when the reader takes it.
The four changes that matter most
If a page cannot be rebuilt from scratch, these four move it furthest.
1. Make the answer paragraph survive removal
Write the 40 to 60 word answer, then read it with the rest of the page covered. If it needs anything above it to be intelligible, it is not an answer, it is an introduction. This single test changes more pages than any other item on this list.
2. Move authorship and dates above the answer
Credentials at the foot of a page are credentials that arrive after the decision has been made. A reviewer line reading “reviewed by [named professional], [registration number], on [date]” placed under the H1 does the work of an entire trust section.
3. Check that the text is in the served HTML
Stage two is invisible from a browser, because the browser runs the code that assembles the page. View source rather than inspect the element, and search for a sentence from the middle of the article. If it is not in the source, some retrieval systems will never see it. This is the failure that produces a page that looks perfect and is never cited.
Modern frameworks generally render content server-side by default, so this usually passes — but it fails silently when it fails, and the check takes thirty seconds.
4. Name the jurisdiction and the date of the law in the same line
“The rules changed in April 2024” is useful to a reader and ambiguous to a machine that does not know which country’s rules or when the page was written. “Under the Immigration Rules of England & Wales, as at 4 September 2026” is unambiguous to both, and it is the difference between being quoted accurately and being quoted about the wrong country.
Applied: a settlement refusal page
How the structure looks on a real subject, in outline:
- H1 — Settlement refused because of absences: what to do next
- Byline — writer, reviewing solicitor with registration number, reviewed date
- Jurisdiction line — England & Wales, rules as at the review date
- Answer — 50 words naming the deadline, the review route and the single most common calculation error, standing alone
- The rule — the continuous residence provision, linked inline
- Who it applies to — applicants under the qualifying route; not those on other routes
- Exceptions — absences that do not count, and the evidence for them
- Procedure — day 0 to the deadline, as dated steps
- Evidence — travel records, and how to obtain the official record of them
- Cost and timing — current fees, named as current and dated
- Common mistakes — reapplying before establishing whether the original calculation was wrong
- Sources — the official guidance and the fee schedule, linked inline and listed
Length is not the variable. A 3,000-word guide with no self-contained passage is less quotable than a 900-word page with one.
What this does not do
Structure gets a page through the first six stages. It cannot manufacture stage seven. Independent corroboration — a listing on a professional register, citations from sources that are not you, a verifiable practitioner behind the name — is earned outside the page, and no page structure substitutes for it. Anyone selling a template as a complete answer engine strategy is selling six-sevenths of one.
Basis for this framework
The stage sequence describes how retrieval and answer systems are generally understood to filter candidate sources; it is a working model rather than any single vendor’s published pipeline, and no vendor endorses this formulation of it.
The structural recommendations come from applying this model to multilingual legal content across two jurisdictions, and from auditing pages in that niche against it. The observations about authorship placement, jurisdiction lines and served HTML are drawn from those audits, described in the related case studies below.
No claim is made that adopting this structure produces citations. It removes the reasons a page is disqualified before relevance is ever assessed, which is a different and more defensible claim.
