How an AI answer gets built, and where a law firm fits

Nobody can see the pipeline from outside Google, so we're all working from evidence. I will give you the sources for each step rather than give you the description on trust.

The man who worked out the sequence below, Dan Petrovic, runs an agency called Dejan. Let me assure you that his findings actually come from experiments rather than speculation. So he reads the grounding data, which is the text that is assembled and handed to the model writing the answer. If you read the grounding data for a query, you'll see that it only holds text that survived the earlier steps¹.

In his write-up, he describes four steps:

1. Your question, broken into several single-intent sub-questions.
2. Each sub-question searched, returning roughly five to twenty ranked sources.
3. Each source cut into sentences, with every sentence scored against the question.
4. The highest-scoring sentences assembled into snippets, the only part that reaches the model.

Step one: your one question becomes several

Google themselves call this query fan-out and describe it as "issuing multiple related searches across subtopics and data sources"². Google's documentation does say this plainly.

How many sub-questions does it take to answer one question? One of AirOps's studies found that ChatGPT generated two or more follow-up searches on 89.6% of 15,000 prompts when they're run through it. In their larger run AirOps most commonly saw exactly two sub-queries (88.6% of queries), and some complicated comparisons can get to four or more (2.5%)³⁴.

For example, somebody could search for this and find some answer full of advice that's stitched together from a handful of pages:

"should I get a lawyer after a car accident"

The phrase is pretty light on details, but it appears the system is attempting to answer the questions behind it: how long do I have to file, what will a lawyer cost, do I have to talk to the insurance adjuster, what is my claim worth, what happens if I was partly at fault. A page is judged by how it answers each of those sub-questions.

Now here's where law firm keyword research gets interesting.

They are real questions and nearly every one of them shows zero searches in a keyword tool. Maybe your "car accident lawyer Houston" page can just target that phrase and carry a contact form, and nothing else. And it answers none of the questions that the system asked.

Fan-out is essentially probabilistic and personalised, so expect ten different sets from ten runs instead of one fixed list. You can collate them from many runs and simply answer the recurring ones directly on your website⁹.

Step two: retrieval, where your ranking decides if you are considered

Each sub-question has a set of ranked results to draw on, coming from a search run in the background. Making that candidate list comes down to ranking, and only a small fraction of pages make it. seoClarity's analysis of 362,000 AI Overview searches showed that 94% of the answers cited at least one page from the normal top 20⁵.

The difference between position 1 and position 20 is huge⁵:

• Position 1, cited 43% of the time
• Position 3, 31%
• Position 20, 7%

Same thing goes for ChatGPT, measured against Google's rankings. The AirOps report showed that 55.8% of the pages ChatGPT cited ranked in Google's top 20 for at least one of the queries involved³. In both AI Overviews and ChatGPT, the page in position one is the most cited, drawing a citation 43% and 43.2% of the time respectively⁵³.

The part that favours a smaller firm

Each sub-question's list is ranked separately from the others. The standard way to merge the lists is to use a reciprocal rank fusion formula that adds up one divided by a page's position in each list.

So a page that sits in middling positions across many lists instead of first in a single list will end up with the higher score.

Treat that multiple just like an illustration of the arithmetic. It obviously depends on the constant used in the formula, but from his working nothing about the direction sounds wrong.

Step three: being retrieved is not being cited

Position inside the retrieval set matters nearly as much as position in Google. ChatGPT also cites the top retrieved result 58% of the time and the tenth one 14%⁴.

That is not a comfortable finding. I assume your page gets found through ordinary search, but the system could read it in full and drop it before the answer, and no one would know. Neither the act of finding, reading or dropping that page is recorded in Google Analytics or Search Console.

The typical "AI visibility dashboard" is glossing over this step. If you want to see more of this, then you can check out the measuring page, which covers that in more detail.

Step four: grounding, and the 540 words that get used

The last step is a new idea to most people, and one that probably calls for a change in how you write. Petrovic used 7,060 queries and 883,262 grounding snippets when trying to figure out how much text actually goes into an answer.

The budget is roughly 2,000 words per question, divided by rank (and shared across every source that made it this far)⁷.

The top source typically gets 531 words, and the fifth gets about half that. Unfortunately, grounding is a zero-sum game⁷.

Here is the same data from the side of the page:

Specifically, the middle column is how many words get used, which is the whole argument against long pages.

Where on the page the words come from

Google finds most of the grounding at the top of the page. In Petrovic's examples it ended up skipping all the unrelated sections and kept the opening paragraphs almost whole¹.

A separate analysis of 42,971 citations matched to their source pages showed that the median cited sentence sat 31.2% of the way down in the page. Three quarters of the cited sentences were in the page's first half⁸.

The extraction works by looking for your best-scoring sentences and lifts them as written¹.

While a sentence that leans on the one before will work just fine in context on the page, the same sentence lifted into an answer usually doesn't work on its own.

Useless alone: "This deadline is shorter."

Survives the trip: "Claims against an Ohio city or county must be filed within two years, and some notice requirements run in months rather than years."

What gets thrown away, and what survives

Here is the gist of what Petrovic's annotated grounding package shows¹.

Thrown away: navigation, headers and footers, time-limited promotional lines, legal boilerplate, and most long customer review text.

Kept: the core service statement, specific process and how-to detail, pricing and specifications, and concrete facts.

What Google confirms, and what is inference

People confuse what Google has confirmed (which is a short list, I admit) with third-party findings such as the ones above.

Confirmed by Google: query fan-out, and that a page must be indexed and eligible to show a snippet to be used in AI features².

Everything else above: the grounding budget, the rank-weighted shares, the sentence-level scoring, the retrieval discard rate. All of these come from (entirely independent) reverse-engineering and third-party measurement. Cyrus Shepard's review scores 55 such studies against each other, which is useful, but it doesn't count as documentation¹⁰.

Anybody who tells you Google confirmed any step beyond fan-out is someone who doesn't read Google's documentation.

What I would do with a practice area page

This is likely an afternoon per page, rather than a project.

1. Collect the real fan-out questions. Try typing your main question into Google's AI Mode and into ChatGPT and see which sub-questions the answers and their cited sources cover. Fan-out is just too dynamic to try and hope to grab every sub-question at one time⁹. The useful ones are the ones that recur over three or four runs.
2. Check Bing Webmaster Tools. Bing Webmaster Tools gives you the grounding queries behind citations of your pages, which gives you real fan-out data to work from instead of the guesses you would otherwise make. It is the only free source that I can possibly point you to for real fan-out queries⁹.
3. Put the direct answer in the first two sentences. So many firms put the firm history and the call to action in the first paragraph so the page is an automatic miss. Most openings are extracted close to wholesale¹, so give them a bit of substance for the extraction to take.
4. Turn each recurring sub-question into a heading, with its answer directly underneath. The heading is central to the match with the question.
5. Rewrite every claim so it stands alone. To test each sentence on the page, quote it on its own. If it doesn't stand alone, try folding it into the sentence before.
6. Put the quotable facts in. Statute number, deadline, filing fee, the court, the fee percentage and what comes out of it¹.
7. Cut back to about 1,500 words. Move the rest to one-question pages and link those up with each other. That spread of pages is useful later during the merge scoring⁶.
8. Leave the trust material on the page for humans. The awards and testimonials may not be grounding fuel, but they are regularly what convinces the person reading. I just put them back below the answer for the people who scroll.

In my experience, the best arguments for content arise when working through the first item on this list. Once attorneys see their own clients' questions and a website answering two of nine, the content argument is over.

There is a page on the sentence-level craft: writing pages AI cites. Much of how the engines diverge is on the ChatGPT and Perplexity page.