A firm can hold a spot on a Google map pack (the three map listings above the regular results), appear in a ranked list, and still be absent from the pages the AI platforms read before they answer. That second list is the one a firm can inspect line by line.
This post covers which websites sit behind AI answers about hiring a lawyer and why so few of them are shared by more than one platform. The How Machines Decide (HMD) research audit, tracked 188 source links that four AI platforms cited while answering ten criminal defense searches in Los Angeles on October 3. Those links came from 78 different websites. Of those, 85% were cited by one platform only, which is 66 of 78. The other 15%, just 12 websites, supplied 43% of all source links, which is 80 of 188.
Behind every AI answer sits a second ranking of source pages
Names are the visible half of an AI answer. Each answer also rests on sources, and those sources form a second ranking that runs underneath the first.
The second ranking is spread thin. Perplexity returned 114 source links on these searches, and they traced back to only 41 websites, an average of 2.8 links per website. Across all four platforms the average fell to 2.4 links per website.
A low average with a few exceptions is the signature of a long tail. A small group of addresses carries the weight and a large group appears once.
That is what the data shows.
Of the 78 websites, 48 were cited exactly once in the entire collection.
Another 18 were cited more than once, always by the same platform. Add those to the 48 and you reach the 66 that no second platform touched. Each time a platform cited a website counts as one citation, and those 66 websites supplied 108 of them, an average of 1.6 each.
The 12 shared websites supplied 80 citations, an average of 6.7 each. Per website, that is roughly four times the rate of the single-platform group. Eleven were cited by two platforms, one by three, and none by all four.
Five of the 12 were directories or legal reference sites, and together they supplied 45 citations. Four were law firm websites, which supplied 26. The last three were court, police and civil liberties sites, which supplied 9.
Twelve websites covered 80 citations, and the other 66 covered 108.
What it means when two platforms read the same website
A shared website is one that more than one platform chose to read for the same kind of question. In this data those websites are the closest thing to a common reference point across the platforms.
A website that lands on more than one platform matters for two reasons.
The first is reach. A profile or page on a shared website is available to several platforms on the same day, while a page on a single-platform website is available to one.
The second is consistency. When a directory profile lists one phone number, one practice area list or one office address, and the firm's own site lists another, the platforms have two versions to read. This post has no data on how a platform resolves that conflict, so treat it as a place to look rather than a finding.
Four firm websites appeared on more than one platform, and they were cited 26 times together. Those sites answered the question on the page itself, in plain language, which is the same work SEO has rewarded for years. Clean loading, clear page titles and a direct answer to one question per page are all things a search specialist already knows how to build.
The link between that work and the citations is inference. The data records what got cited on one date and cannot say why.
Checking one platform leaves the other three unread
If a firm checks which sources one platform cites for its core searches, it will learn about that platform. Whatever the other three cite stays out of view.
Take the Perplexity result. Its 41 websites are a short, readable list. Yet 66 of the 78 websites across the collection were seen by one platform alone, so a firm that reads only that list has a partial map no matter how carefully it reads.
As an example rather than data, a personal injury firm running the same check would build its own list of shared websites, and that list could look very different from this one.
A fair comparison across platforms needs the same searches, run in the same week, with the same wording. HMD holds those constants so the counts can be compared. A firm doing this alone can hold them too.
Which of the shared websites carries the heaviest load for your searches?
Directories, firm sites and courts played different roles
The 12 shared websites fall into three groups, and each group behaved differently.
The five directories and legal reference sites supplied 45 citations, an average of 9.0 each. They are built to answer broad questions, and the platforms returned to them more often than to anything else in the collection.
The four firm websites supplied 26 citations, an average of 6.5 each. They were the only business-owned pages that two or more platforms agreed on.
The three court, police and civil liberties sites supplied 9 citations, an average of 3.0 each. They covered the procedural answers, the kind a person needs after an arrest rather than before hiring anyone.
Directories were 5 of the 12 shared websites and supplied 56% of the shared citations. Firm websites were 4 of the 12 and supplied roughly a third.
A firm has no control over the court and police sites. It has direct influence over its own website and some influence over the directories, through the accuracy of its own profile.
The 18 websites a single platform kept returning to
One group sits between the 12 and the 48.
Eighteen websites were cited more than once, always by a single platform. They supplied 60 citations in total, an average of 3.3 each. One platform kept returning to them, which says more about how that platform reads than about the market.
This group matters because it shows a platform's habits. A directory that one platform cited nine times and the other three never cited is a reading preference of that platform on that date. A firm that looks only at the websites all platforms share would miss it.
The sensible treatment is a watch list. A website that one platform leans on repeatedly is worth a look, even though it has not yet landed on more than one platform. A website that moves from one platform to two has crossed the line that matters. A watch-list website also gives you a first place to look when a platform's answers change.
A names-only report misses the part a firm can change
A report on AEO, the work of getting a firm named in AI answers, that stops at names and positions records only who was recommended and where.
That record is real data, and HMD collects it too. It describes the visible part of the answer. A report with no source layer cannot tell a firm which pages sat behind those names, and the source layer is where a firm finds concrete options, such as correcting a profile, adding a page or updating a listing.
The two layers also move separately. A firm can gain a position without any change in the pages cited, and a page can be cited without a firm being named. The 188 citations in this collection are a measure of reading. Names are a measure of recommending. A report built on names alone would have shown none of the 66 single-platform websites. Both belong in an audit, and the source layer is the one a firm can inspect line by line.
Four checks to map the websites behind your own searches
Each one traces to a figure above.
Run the five or six searches a prospective client would type, on each platform, and write down every website cited next to the platform that cited it.
Mark the websites that show up on two or more platforms. That is your version of the 12.
A simple grid works well, with websites down the side and platforms across the top, so a shared website shows up as a row with two or more marks.
Open each profile your firm has on those shared websites and read it as a platform would, starting with the firm name, practice areas, address and phone number. Compare it line by line to the firm site, since the platforms can read both.
Add every website that only one platform keeps returning to onto a watch list, and run the same searches again next quarter to see which ones move.
None of this promises a citation. It shows what the platforms had in front of them on the day you looked.
Run the count for your own market
These figures describe ten searches in one metro on a single date, and platform sources can change between runs. The next run may produce a different 12.
Pick a practice area, run the searches and count the websites that appear on two or more platforms.
Start with that count.