The Law Firm AI Visibility Index.

Source: KKV Labs, Google AI Mode answers captured September 9, 2026 and September 11, 2026 in 60 cities.

Page builders are read on WordPress sites and reported as a share of every site we counted: Gutenberg blocks 74%, Elementor 14%, WPBakery 5%, Divi 4%, Beaver Builder 3%. A site can carry more than one.

A legal website vendor is named on 8% of the sites we counted. This is one of the checks that reads the other way round: 5% of the firms the answer named against 10% of everyone else. A vendor is counted only when the homepage carries a platform tell, so a profile link on a directory is not read as a website build.

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

How the homepage takes the enquiry

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

This is the widest pair of separations on the page that a firm can act on the same afternoon. A case results page is linked from 63% of the firms the answer named and 44% of everyone else, and a figure on the homepage itself from 48% against 33%.

What the homepage offers a hurt person

Whether the page prints a number a reader can check

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

Named firms are the 187 crawled sites Google AI Mode listed in their own city answer. Everyone else is the other 498. Most of this table is a list of the kinds of case a homepage says it takes, and a page of results, which is to say it is about what the site writes down rather than what it runs.

The checks that read the other way round

These are the most useful rows on the page, because they are the ones the category would bet against. Nothing here runs backwards by much, so the honest reading is that these checks do not sort the market rather than that they hurt. An analytics tag, a licence line and a website vendor are all worth having. On September 11, 2026 they did not mark out the firms the answer named.

The checks that did not tell the two groups apart

The two schema checks that say what the business is are the closest rows on the page: law firm or local business markup sits on 66% of the named firms and 68% of everyone else, and the LegalService or Attorney type on 62% against 61%. That makes markup open ground rather than catching up.

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

94% of the homepages we counted carry structured data, and 26% carry only the website record the theme ships, with no business type on it. The named firms sit on 30% of that last line against 24%, so naming the business type is open to everyone here. It is a one line change on most sites.

A robots.txt was reachable on 99% of the sites we counted, which is the base for the rows above. About half of those files disallow some path for every crawler, which is what a default admin rule looks like. We did not keep the rule text, so we report only the outright blocks and leave the rest alone.

46% of the sites we counted serve an llms.txt file, at 314 sites. The shape of that number is the finding: every Duda site we read has one, 37 of 37. That is 37 of the 314 llms.txt files in this market, written by the builder rather than chosen by the firm.

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

What the homepage links before anyone calls

Words and weight are counted from the HTML our crawler downloaded, not from a browser that ran the page. A site that builds its text with JavaScript therefore reads thinner here than it looks to a visitor. Both halves of the named comparison were measured the same way on the same day, which is what makes that gap the useful part.

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

The kinds of case that separate the two groups furthest

The road accident lines are where the distance is. It is a reading on one day and not a cause, and the cheapest thing it suggests is naming every kind of case the firm already takes.

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

Readiness runs from Connecticut at 68 down to South Dakota at 22, across 38 states with at least 8 sites counted. Several of those states rest on 8 to 10 sites and most are a single city, so a state row here is often really a city row. Read the row before you read the ranking.

Source: KKV Labs crawl of 685 personal injury firm homepages, of 875 opened, September 11, 2026.

Source: KKV Labs screening and Google AI Mode answers captured September 9, 2026 and September 11, 2026. 60 cities, 297 firm names. Every city returned three or more names, so the answer column is the count itself rather than a label. Download the city data (CSV). What we do for law firms.

The city sample

60 US cities across 39 states. We found 861 personal injury firm websites in them and screened 701. This is not a random sample of all US injury firms: it is the markets we were looking at, which lean toward small and mid size cities. A firm with offices in two of these cities is listed once, in the city we found it in first.

The AI answers

One question per city, asked in Google AI Mode from a fresh logged out session, with the firms recorded in the order they were given. The answers were captured in two batches: 27 cities recorded September 9, 2026 and 33 cities recorded September 11, 2026. Answers change. This is a dated snapshot, not a ranking and not an endorsement of any firm on it.

The technology crawl

On September 11, 2026 we opened 875 personal injury firm homepages and read 709 of them. 130 answered and turned the crawler away, and 36 did not answer at all. Detection reads the homepage HTML for known scripts, links and markup, so a firm can use a tool without it showing here. Every share is a floor.

The relevance gate

A site is counted only when its own text or one of its own page links shows it does personal injury work. That dropped 24 of the 709 sites we read, which leaves the 685 every share on this page is measured against. A family law or criminal defence firm has practice areas too, so a practice areas link on its own is not treated as proof.

Named versus unnamed

The named set is the 187 crawled sites Google AI Mode listed in their own city answer. The other 498 are everyone else. Comparing the two shows what the named group happened to have that day. It does not show that any of it is why they were named, and we do not claim that anywhere.

How a city was coloured

The answer column is the count itself: 60 cities had three or more firms named, 0 had two and 0 had one. No city was coloured by hand. In this trade the label sorts nothing, because every city fell in the same bucket. Read the firm list, not the label.

What we are not claiming

Nothing here says a firm with better markup will be named, or that a named firm is better at law than one that is not on the list. It says what this market looked like on two days in September 2026, and which of the cheap signals are still open.

A same day comparison, not a cause

The named firms were named for many reasons this crawl cannot see, including reviews, how long they have been open, press coverage and who links to them locally. Nothing on this page is a ranking factor.

Every city is in the same bucket, so the label says little

All 60 cities came back with three or more firms named, so the answer column does not sort this market the way it sorts a market with quiet cities in it. The useful column is the firm list and the count beside it.

56 named firms have no website we could resolve

Google AI Mode gave those firm names with no link, and no first-party record we hold maps the name to a site. They are counted in the city table, which is built from names, and left out of the named versus unnamed comparison, which is built from websites. We resolved 235 sites and read 187 of them, so that comparison rests on 63% of the firms the answer named.

4 websites were thrown out before we published

A guessed hostname can land on a business that shares a name and nothing else. Every match has to pass three tests: a distinctive word of the firm name in the title or the markup, address shaped evidence for the city it was named in, and at least five legal words on the page. 4 rows failed and lost their website, including one employment firm in a different state, two firms in the wrong state entirely, and one server that answered with a page nothing could be read from. Five weak rows were kept with the evidence written down.

130 sites are behind a bot wall, and they are not a random group

130 of the 875 sites we opened, 15%, answered and refused the crawler. A firm spending heavily on its marketing site is more likely to run that kind of front door, and 40 of the 235 named websites sit in that bucket. Every share here is measured on the sites that let us read them.

Six of the sixty markets were searched thinner than the rest

The candidate list was built by search, and the budget ran out before the last few cities got their fourth and fifth query. Those six are Traverse City, MI, Worcester, MA, Reno, NV, Owensboro, KY, Pueblo, CO, Lawton, OK. Their candidate counts are the smallest in the index, so read a per-city site count there as a floor rather than a census.

Homepage only, so every share is a floor

A bar admission on an About page, a settlement on a results page or an intake tool behind a free case review click is not counted. Detection is by known scripts, links and text, so a new or self built tool reads as nothing at all.

Results printed as pictures read as nothing

A recovered amount is counted when a figure sits in the text next to a recovery word. Several firms print their settlements inside images, which a crawler reading HTML cannot see. That is a limit of the measurement, not a statement about the firm.

No speed numbers, on purpose

The crawl ran from India. Response times were recorded and are not reported here, because from that distance they measure the length of the wire rather than the speed of the site.

Small bases on some rows

The named side is 187 sites, so several rows in the comparison rest on a handful of firms, and the site counts are printed next to every share. Quote the counts. Never say twice as likely.

The state table runs on small bases too

38 states had at least 8 sites counted. Several sit on 8 to 10, and most are a single city, so a state figure there is really a city figure.

A partial robots.txt rule is not an AI block

About half of the robots.txt files we read disallow some path for every crawler, which is what a default admin rule looks like. We did not keep the rule text, so we report only the outright blocks and leave the rest alone.

Two capture days, not one

27 cities were recorded September 9, 2026 and 33 cities were recorded September 11, 2026. Two cities two days apart were not asked in the same session, so read each row against its own date.