Everything else will fall into place once we accept how AI answers are built . That is, a page is broken up into sentences and the top-scoring few of those sentences are passed to the model that writes the answer.
You treat sentences as if they are separate answers.
1. Put the answer in the first two sentences
The position on the page is one of the strongest signals in the citation research. Dan Petrovic read the raw grounding data Google passes to the model and found that opening paragraphs were extracted almost wholesale. The unrelated sections were, of course, almost entirely skipped1.
Somebody else measured the effect of position across 42,971 citations matched back to the pages they came from2:
- The median cited sentence sat 31.2% of the way down the page
- The 25th percentile was 18%
- Three quarters of all cited sentences fell in the first half of the page
Their conclusion is unusually direct, which I will quote below:
A page that opens with a clear definition or a direct factual answer has placed that sentence in the highest-probability citation zone on the whole page2.
Now look at the opening of these typical law firm pages.
"For over 30 years, our attorneys have fought tirelessly for injured people throughout Michigan. We understand that an accident can turn your life upside down."
The first sentence doesn't answer a question anybody typed, and neither does the second. The system looks up the question and sees what sentences sit in that top spot, reads them and that is that.
Meanwhile, the typical visitor is wondering how long they have to file. Often these answers are buried in paragraph seven, under a heading like "Understanding Michigan Personal Injury Law".
So here is the rewritten opening.
"In Michigan you have three years from the date of a car accident to file a personal injury claim, under MCL 600.5805. If a government vehicle was involved, you may have as little as 120 days to give notice."
This version is quotable, specific and checkable and answers the visitor's own question. But lower down the page things are different. That's where people weigh up the firm's thirty years, often just before they're ready to call anyone.
2. Make the heading the question
AirOps has traced 353,799 pages through ChatGPT's retrieval over one study, and about 41% of the time the pages with closely matched headings actually got cited. Once ranking was accounted for, matching the heading to the query was more important than breadth of coverage3.
Treat the heading just like a different way of entering the query.
| Usual legal heading | What to write instead |
|---|---|
| Our Approach to Premises Liability | How long do I have to sue after a slip and fall in Ohio? |
| Understanding Comparative Fault | What if the accident was partly my fault? |
| Fee Structure | What does a personal injury lawyer cost in Ohio? |
| Why Choose Our Firm | Leave this one. It is for the reader, not the machine. |
Focus on collecting questions from your intake notes and the sub-questions these systems generate rather than pulling them from a keyword tool. AirOps has found that 95% of the sub-questions driving retrieval have no recorded search volume, with the keyword tools showing nothing for them at all4. Check out this argument from the keyword research page.
3. Be specific enough to be worth quoting
First, AirOps checked out 217,508 pages in its citation data and found that content with five to seven supporting statistics had a 20.3% higher chance of citation5.
As it happens, a Surfer study that compared cited and uncited pages in AI Overviews has come to the same conclusion. Only 24% of a topic's key facts are covered on uncited pages and 31% covered on the cited ones. Even at the median, the cited page covered 62% more key facts than an uncited one6.
The legal equivalent of a statistic is the detail, the statutes, deadlines, dollar figures, counties, form numbers and courts of your practice. Look at what happens to a sentence.
The copy in this right-hand column is illustrative. The statute citations are real but the fees and timelines should be replaced by numbers your firm has verified.
| Unquotable | Quotable |
|---|---|
| "You may have limited time to file a claim." | "Ohio gives you two years from the date of injury to file most personal injury claims, under Ohio Revised Code 2305.10." |
| "We work on a contingency basis, so there are no upfront costs." | "The fee is 33% of the settlement if the case resolves before suit is filed, and 40% after. Case costs such as filing fees and expert reports come out of the settlement, itemised." |
| "Cases vary in how long they take." | "Most rear-end collision claims we handle settle in four to nine months. Cases that reach trial in Franklin County are currently taking around two years from filing." |
The quotable column is a lot harder to write because so many of the numbers that go in it need an attorney to check them and someone to keep them current. That's why so few firms in the US have quotable answers on their practice pages. Those who do are getting the citations from AI search.
4. Write passages that stand on their own
Because each sentence is extracted literally, you can't lean on the one before it1.
Means nothing alone: "This deadline is shorter."
Survives being quoted: "Claims against an Ohio city or county must be filed within two years, and some notice requirements run in months rather than years."
Cyrus Shepard reviewed 55 different studies and scored the qualities that were tied to citation out of 107:
- Factual specificity, 8.3 out of 10
- Explicit phrasing, 8.1
- Self-contained passages, 8.0
Hedged sentences can score badly for the same reason that they can read badly. There is nothing in them to quote.
Read out each sentence as if you were reading it to a stranger. If it doesn't make sense alone, it doesn't get quoted alone.
Much much more often, the real culprit is a "this", "that", "these" or "it" at the start of the sentence. Fix the pronoun or at the very least name what it refers to.
5. Use structure, in moderation
I don't think it's an accident that all three of these findings involve formatting rather than prose258:
- 95.6% of cited pages had at least one list, table or multi-level heading
- Comparison pages carrying three tables earned 25.7% more citations
- Sections of roughly 120 to 180 words between headings drew about 70% more ChatGPT citations than sections under 50 words
Length matters too, section length matters but so does sentence length5.
Nothing in the data on length is a case for ten-word sentences. Seriously, a page of them reads like a machine's first draft. There's a good argument for replacing the paragraph describing options on, say, a practice area page with a table. Breaking down a 60-word sentence with three subordinate clauses into the three facts it contains is not something I would call dumbing down.
6. Judge length by what gets read, not by a word count target
I understand that the studies disagree on length, and there is ample detail on both sides of the disagreement on the research page. This is the measurement that, in my opinion, settles it in practice9.
So go ahead and aim for 1,000 to 2,000 words on a practice area page with the answer at the top, and if you find you are running longer, spend the extra on separate linked pages that each answer one question. You need coverage across many related queries, which is what the retrieval maths is all about10.
The rules you have to write inside
At some point the results and comparisons in your copy come into conflict with attorney advertising rules. But remember that I am the one that told you to add them.
- Model Rule 7.1 sets the ABA's standard, which prohibits lawyers from making false or misleading communications about their services. State versions all vary in how they apply it to past results12.
- In Florida, past results and comparisons of skill or record have to be objectively verifiable. But Florida also requires you to add a disclaimer to testimonials saying prospective clients may not get the same result13.
- New York makes firms use the disclaimer "Prior results do not guarantee a similar outcome" wherever a claim is likely to create an expectation about results14.
It includes the verification trail and the disclaimer in the same breath: "We recovered $2.3 million for a client injured in a rear-end collision in 2024. Prior results do not guarantee a similar outcome."
You can see rule numbers for four jurisdictions on the advertising rules page.
What still has to be true for human readers
None of this replaces Google's own test for whether a page deserves to rank, but I find the questions in it to be very practical11:
- Does the content show first-hand expertise?
- Is it clear who produced it?
- Does it go beyond the obvious?
- Is any use of automation self-evident to visitors?
Many extraction tips are good as a rule of thumb but cause more harm than good when they are blindly applied for a machine. You end up with clipped sentences, a heading every eighty words and no voice at all. The page is unreadable and low quality at the same time. In a field where the reader is deciding whether to trust you with their case, writing for the machine can be a costly mistake.
Write it for the person, put the facts up front and you're done.
What I would do with one page this week
- Pick a practice page to rewrite. Find your most-shown, least-clicked page in the Search Console data. Your best bet early on is the page Google already thinks is nearly relevant.
- Write down the question in a client's words and use it as a foundation for the page. Then check the current copy to see how far down the answer is.
- Move the answer into sentence one or two, and make it specific enough to quote.
- Write every heading in the voice of a client asking a question. Except for those very few headings that are there for the reader.
- Find the facts the page dances around and add them, or get the attorney who handles the work to supply them. Statute and number, deadline, fee percentages and what comes out of them, typical timeline, the court.
- Fix the pronoun openers, then give them the read-aloud test.
- Turn one descriptive paragraph into a simple table. Options, deadlines by claim type, or what happens at each stage.
- Put every results claim through a compliance pass and keep the evidence for yourself. So the claim gets its disclaimer, the comparison gets checked against your state's rule and the page gets published.
- Note the date, wait six weeks and see what Search Console shows for the page.
Knowing which change worked is something you learn one page, one change set at a time. It's tough to learn anything if you rewrite forty pages at once. For most legal sites, you only need to make the top five pages excellent.
Partners worry that if they answer too much too soon the visitor will stop calling.
Counterintuitively, I think the opposite happens. People trust the clearest page, the one that answers the question. Nobody with a real claim is going to risk it on representing themselves because they learned the deadline from your page.
References
- , Selection rate optimisation and grounding snippets The grounding pipeline, extraction behaviour, lead bias, primary bias
- , How Google picks which sentences to cite in AI Mode: 42,971 citations Position of cited sentences, structured element share, text fragment analysis
- , The fan-out effect: what happens between a query and a citation 16,851 queries, 353,799 pages: retrieval position, heading match, coverage, DA correlation
- , The influence of retrieval, fan-out and Google SERPs on ChatGPT citations 548,534 retrieved pages: 15% cited, fan-out volume, title overlap, Google rank effect
- , From retrieved to cited: how commercial content earns citations in AI search 217,508 pages across 7,500 commercial prompts: statistics, tables, sentence length, word count
- , Do pages with more key facts get cited in AI Overviews? 57,000 URLs analysed Fact coverage of cited versus uncited pages
- , AI citation ranking factors: evidence from 54 experiments, patents and case studies (2026-09-11) 23 scored factors and the underlying source collection
- , How to optimise for ChatGPT: 129,000 domains and 216,524 pages Referring domains, domain trust, content length, section length, review profiles, speed
- , How big are Google’s grounding chunks? 7,060 queries, 883,262 snippets: the 2,000-word budget and per-page coverage
- , Fan-out framework: 5 steps to optimise for fan-out queries (2026-09-10) Method for collecting fan-out queries and measuring the result
- , Creating helpful, reliable, people-first content (updated 2025-12-10) Self-assessment questions, E-E-A-T, YMYL weighting, the Who How Why section
- , Model Rules of Professional Conduct Rules 7.1 to 7.3 on communications, advertising and solicitation
- , Quick reference checklist: websites The Bar’s own checklist of what a law firm website must and must not contain, by rule number
- , Rule 7.1: Advertising The "Attorney Advertising" label, the prior results disclaimer, and one-year retention
