ChatGPT, Perplexity and Claude for law firms

ChatGPT answers most questions without searching at all

Over three months Profound examined around 730,000 conversations of US ChatGPT users with at least one web citation¹.

A lot of this searching happened the same time the user was asking their first question. The first turn had 2.5x more chance of citations than the tenth. That's nearly 4x more than the twentieth¹. So the opening turn is the more cited and the more important at the same time.

The needs of the 18% and the 82% are almost diametrically opposed to each other.

For the 18%: ordinary SEO

Your page will be considered for a citation if you rank, get retrieved and get your sentences selected .

For the other 82%: nothing of yours is read

Say somebody merely asks for the best injury lawyers in Tampa. The answer in a chat with no search though, comes from what happened while the model was trained. The content of that answer depends on how much consistent coverage your firm is getting across the web.

When Dan Petrovic studied this effect, he called it primary bias. It takes about 3 to 6 months to shift that bias with almost no help from anything you publish (it moves when the model is retrained)².

When ChatGPT does search, your Google ranking mostly decides it

The way AirOps tested ChatGPT was actually matching 43,233 of its original and follow-up queries against Google's top 20 and see what got cited³.

• 55.8% of cited pages ranking in Google's top 20 for at least one of those queries
• 43.2% citation rate for pages ranking first in Google, roughly 3.5 times the rate for pages outside the top 20

But wait it gets steeper, if you look inside ChatGPT's own retrieval order⁴³.

The finding that favours a smaller firm

ChatGPT takes a question, splits it up into a bunch of searches, and merges the ranked lists of those searches into one. The scoring rule is one, divided by your position (plus a constant) on each list. You can see how all these scores really add up in the aggregate. The biggest winner is consistent ranking across the whole set of searches.

Using that formula as a reasonable reference Metehan Yesilyurt calculated that, if you rank around sixth for thirty related queries your score is roughly ten times that of ranking first for three⁵. Like almost all multiples, take it with a grain of salt. It really depends on the constant you use in the formula.

Perplexity is close to a mirror of Google

So Perplexity often ends up doing not much more than citing what Google already ranks.

The only catch is that you cannot read much into the visibility. An academic study scoring pages against a quality framework showed that Perplexity cited lower-quality pages on average than Google or Brave, with a 45% citation rate against 72% and 78%⁷. I'm not surprised Perplexity cites weaker pages, as it casts a wider net than Google or Brave would.

Claude waits for sources to agree

In a useful experiment, Lily Ray put out a source (a new article with specific claims) and see which systems repeated it. Ray, the SEO, said the lack of other sources may make Claude and Perplexity less willing to take a single new claim at face value⁸.

What they cite when somebody asks about your firm by name

If you think about it, when you ask about a firm by name, you get a very different mix of sources⁹.

• Social citations up from 5.4% to 15%
• Reddit alone at about 8% of ChatGPT's citations on those questions

Oddly enough, ChatGPT's social citations in the US (57%) and every other country measured are dominated by Reddit. Low end is 51% of them, upper is 76%¹⁰.

Let's say you get a prospective client checking your firm out in ChatGPT. What strangers say about you in public isn't the entire answer, but it's a big part of it. But that certainly does not mean it's all out of your control. The solution is fairly straightforward, don't let the details drift, use the same verifiable facts everywhere that these systems weigh it all against.

Review platforms carry unusual weight here

These have a legal equivalent and I'm talking about the likes of Avvo, Justia, Martindale, Super Lawyers, Best Lawyers, your state bar's own directory and your Google Business Profile.

So, a half-finished profile itself isn't a good source. Since two profiles with different details become a conflict, this can easily weaken the entity that ties them together unless some sort of checks are in place. Most firms don't even complete their profiles. I don't think a firm needs more than an afternoon setting all of these up.

How quickly all of this changes

Semrush kept track for three months of which domains each platform cited most¹².

• ChatGPT's Reddit citations down from close to 60% of responses to around 10% in about six weeks
• Wikipedia on the same platform down from roughly 55% to under 20%
• Both steady on AI Mode and Perplexity over the same period

ChatGPT's sourcing can change rapidly, without warning in unexpected ways. Any fixed playbook for getting cited presents an extremely incomplete picture of AI search. For one it's a constantly moving target. Please keep in mind that nothing here is a tactic for one engine.

First of all, a single check has been run once and means nothing. Answers change between runs, different users get different results, the sources behind them change constantly, and there's no fixed baseline. More extensive sampling and testing (across runs and engines, per academic work on measuring AI visibility) would be required to validate any number¹³.

What I would do

1. Rank for the practice-area questions. In every system measured here, ranking is the biggest factor.
2. Cover the question set broadly. Consistency (in ranking across many related queries) plus breadth really do make the merge scoring work for you⁵.
3. Complete and align every third-party profile. At the very least, complete your bar listing and start fixing the legal directories. In SE Ranking's data review-platform domains got about 3 times as much ChatGPT source use as the rest¹¹.
4. Make the same claims everywhere. As long as you're consistent about what your site, your profiles and any press say, the systems that weigh agreement can get it confirmed⁸.
5. Check a fixed question list monthly, on at least two engines. Care about the share of questions that name your firm and keep track of which engines are naming you and which ones are not. Method details are on the measuring page.

What I would not spend money on

• ChatGPT-specific optimisation. ChatGPT's sourcing in general moves fast, but a fixed optimisation package is painfully slow¹².
• Fixes for what the model "knows". These packages promise to fix what the model "knows" about your firm. I think there will always be a model view of your firm, but what that view looks like will change over months of retraining².
• Perplexity-specific work. There's no need to ever pay for Perplexity work⁶.