Preliminary: third-reviewer rulings on form length, Spanish option, medical sources and privacy link are pending; see corrections.
What does the benchmark measure?
The benchmark records what 30 birth injury practice pages showed on 2026-11-03; it is not a ranking, compliance audit or judgment of legal work. PI Marketing sells birth injury marketing to law firms, so it has an interest in these results.
Who scored the pages?
AI agent sessions scored every page; no person has re-scored any. The first scorer scored all 30: 15 from a phone render, as shown on an emulated 390 x 844 phone, and 15 from page source, the HTML the server sends before any script runs. A same-day render pass added three phone-screen items to those 15. A second scorer scored 10 pages from its own renders. Both re-scored past results on all 30; a third AI agent session settled their one difference. Spencer Moeller, PI Marketing’s founder, approved the article outline on 2026-11-03 and publication on 2026-11-04.
How were the firms and pages picked?
By a fixed draw from 20 searches, run once each on 2026-11-03 in an AI agent’s web search tool (US, organic results only). Each was “birth injury lawyer” plus, in order: New York, Chicago, Los Angeles, Houston Texas, Philadelphia, Atlanta Georgia, Miami Florida, Boston Massachusetts, Baltimore Maryland, Detroit Michigan, Seattle Washington, Phoenix Arizona, Cleveland Ohio, Denver Colorado, Minneapolis Minnesota, St. Louis Missouri, New Jersey, Dallas Texas, Washington DC Virginia and Charlotte North Carolina.
- Eligible: the firm’s own page, loading normally (HTTP 200), with “birth injury”, “birth trauma” or “birthing injury”, or a plural, in its title or main heading. Directories, lead-generation sites and other publishers were excluded.
- Drawn: 3 multi-market firms (own pages in 3 or more searches), then, in search order, the first eligible new firm per search (20) and the next for the first 7 searches (7).
- Passed over: 5 firms whose sites blocked the fetch (4 in the draw, 1 checked during design). The draw rule missed two well-known birth injury firms.
How was each page observed?
All 30 pages were observed and scored on 2026-11-03, US Pacific time; one later failed to load at the past-results re-score, so that item is not observable for it. The first screen is the view 5 seconds after load. Page text is what is visible after scrolling down and back. No scorer accepted a banner, opened a chat, submitted a form or called.
How many pages were scored from source?
On 15 of the 30 pages, the first scorer read source; only the phone-on-first-screen, live chat and sideways-scroll items (rendered about 25 minutes later) and past results came from a render. Source can hold text a visitor never sees and miss text that scripts add. Of the 6 such pages the second scorer rendered, 2 differed: a form hidden on phones and a menu-only Spanish link.
What exactly was scored?
Twenty items, each scored yes, no, not applicable, not observable or a number (past results has its own scale). Yes is better unless noted; this is not a legal standard.
Intake
- Phone on the first screen
- A full number, or a “Call” button that dials, legible on the first screen; icon-only is no.
- Tap-to-call link
- A
tel:link, which opens the phone’s dialer. - Form length
- Visible fields in the first step of the first contact form (checkboxes count, CAPTCHA does not). Fewer is not scored better.
- Live chat
- A chat launcher visible at phone width within 15 seconds. Chat code alone is no.
- After-hours promise
- A phrase about the firm’s availability, not medical care: 24/7 (or 24 / 7), 24 hours a day, 24-hour, available 24, around the clock, nights and weekends, evenings and weekends, 365 days, day or night.
- Spanish option
- A visible Spanish link or button, a Spanish version declared in the code, or “Hablamos español” or “Se habla español”. A closed translation or site menu is no.
- Response-time promise
- A specific response time, such as “within 24 hours”, outside testimonials. Recorded as a promise, not as kept.
Content
- Condition pages
- How many of 10 conditions link by name to a working firm page with the condition in its title or main heading: cerebral palsy; HIE, birth asphyxia or oxygen deprivation; Erb’s palsy; brachial plexus injury or Klumpke’s palsy; shoulder dystocia; kernicterus or newborn jaundice; neonatal seizures; brain bleed (intracranial or intraventricular hemorrhage); periventricular leukomalacia; skull fracture or cephalohematoma. Yes if 3 or more.
- Medical sources
- Distinct qualifying sources in the main content: CDC, NIH including PubMed, MedlinePlus, FDA, HHS, AHRQ, the WHO, ACOG, AAP, March of Dimes, peer-reviewed journals, academic medical centers and children’s hospitals. An unlinked reference counts if it names the journal or organization and the article title or year. Wikipedia, other law firms, directories, news and lead-generation sites do not. Yes if 1 or more; accuracy is not assessed.
- Author or reviewer
- A named person with a byline cue (“By”, “Legally reviewed by”). A bio box, or a name only in code, is no.
- Visible date
- A labelled month and year for the page (“Updated”). A copyright year is no.
Trust: placement observed, not compliance
- Advertising label
- “Attorney Advertising”, “Lawyer Advertising”, “Advertising Material”, or a sentence calling the site or page advertising. A disclaimer link alone is no.
- Past-results statement
- Defined below.
- Guarantee language
- A “guarant-” word asserting a guarantee; denials do not count. No is better.
- Privacy link near the form
- A privacy link inside the form or the box that holds it, or within 300 pixels below submit. Footer-only is no.
Technical
- HTTPS with redirect
- Loads over HTTPS, and the
http://address redirects to it. - Viewport, no sideways scroll
- A viewport tag, code telling phones to fit the page to the screen, and no sideways scroll at 390 px.
- HTML document size
- Uncompressed bytes of the HTML document the server returns to a desktop-browser fetch, not a browser “Save as”. Not a speed measure.
- Structured data
- Machine-readable labels for search engines that include a
LegalServiceorAttorneyitem. - robots.txt and sitemap
- Both load:
robots.txt, which tells crawlers what to fetch, and a sitemap listing the site’s pages.
How is the past-results statement scored?
From the phone render only: is a past-results statement visible within one phone screen of every block of dollar results?
- Visible: in page text, inside the screen width, with no tap or swipe; collapsed accordions, closed menus and hidden slides are absent.
- Dollar result: money the firm says it obtained for clients, reviews included; not fees, costs, caps, hypotheticals or averages.
- Block: one results-only container, one paragraph, or one standalone figure.
- Qualifying statement: a sentence or footnote saying in its own words that past results do not predict an outcome, that they depended on their facts, or that every case differs. An inferred meaning is a near miss. Qualifying means this study’s wording test, not any state’s rule.
- Next to: within one phone screen; linked disclaimer and terms pages never make a block pass. Footer region: the last footer and anything after the main content.
Values, first that applies (only Yes counts):
- Not applicable: no dollar result.
- Yes: every block passes.
- Partial: some blocks pass.
- Not adjacent: a statement outside the footer region.
- Footer only: statements only in the footer region or a linked page.
- Not observable: a linked page never rendered, or the page did not load.
- No: no statement anywhere.
How is each count calculated?
Counts are “x of n”, n being the pages where the item applied and could be observed.
- Left out of n: not applicable and not observable. The past-results n is 17: 11 pages showed no dollar result and 2 could not be observed. The privacy link’s n is 29: 1 page had no form.
- Measured items (form length, condition pages, medical sources, HTML size) get a median and range; three also yield a yes (a form, 3 or more conditions, 1 or more source). A form hidden on a phone has 0 visible fields.
How closely did the scorers agree?
On the 10 pages both scored, the two AI scorers agreed on 192 of 200 values (96%), counting past results under its first definition. Excluding 18 values from the render pass rather than the first scorer, it is 174 of 182 (96%).
Kappa discounts chance agreement (1.00 is perfect) but is unstable on 10 pages: one difference gave 0.62. No page in the 10 passed response time, byline or visible date, so those passes went untested.
Where the scorers differed (10 of 30 pages, 2026-11-03)
| Criterion | Agreement | Kappa | The difference |
|---|---|---|---|
| Form length | 9 of 10 | 0.88 | Hidden form: 4 fields in source, none rendered |
| Spanish option | 9 of 10 | 0.80 | A link only in a closed menu |
| Medical sources | 9 of 10 | 0.85 | 1 source against 5; both yes |
| Past-results statement | 8 of 10 | 0.73 | Rewritten; see below |
| Privacy link near form | 9 of 10 | 0.62 | Follows from the hidden form |
| HTML size | 8 of 10 | None (byte count) | 4 and 202 bytes apart |
What happened to the past-results item?
It was rewritten, as any item below 90% is; other differences go to a third reviewer. Both scorers re-scored all 30 pages from their own renders under the rewritten definition: 29 of 30 agree (97%), kappa 0.96. The one difference was a sentence saying cases can vary widely. A third AI agent session ruled it a near miss. The count of 1 does not depend on that ruling: the scorers agreed on whether each of the 17 scored pages passed.
How much could a count move?
Most counts could move by one or two pages if every borderline, near-miss or disputed score went the other way. Structured data could rise from 20 to 25 of 30, because 5 of its 10 “no” pages were checked only in source. Past results stays at 1.
How far borderline, disputed or source-only scores could move each count (30 pages, 2026-11-03; preliminary)
| Criterion | Reported | Range | Why |
|---|---|---|---|
| Form present | 29 of 30 | 28 to 29 | The disputed hidden form |
| Live chat | 13 of 30 | up to 15 | Tap-to-text; a launcher after a scroll |
| After-hours promise | 15 of 30 | up to 16 | A “24 Hours” header |
| Spanish option | 15 of 30 | down to 13 | A call-link label; a menu-only link (disputed) |
| Response-time promise | 3 of 30 | 2 to 4 | Two borderline phrases |
| Medical sources | 10 of 30 | up to 11 | A state children’s services link |
| Visible date | 5 of 30 | down to 4 | A date with no label |
| Advertising label | 9 of 30 | down to 8 | Unlisted footer wording |
| Past-results statement | 1 of 17 | 1 of 16 to 1 of 17 | One page could drop out |
| Privacy link near form | 6 of 29 | 4 to 6 of 29 | A CAPTCHA policy; an embedded form. The hidden form would make n 28 |
| Viewport, no sideways scroll | 26 of 30 | up to 27 | Scroll suppressed on a wide page |
| Structured data | 20 of 30 | up to 25 | A ceiling: source-only checks |
The ranges exclude the unchecked visibility risk on 9 of the 15 source-scored pages, so “down to 13” is not a floor.
What can the counts not tell you?
Whether features work, pages comply, or firms outside the sample differ.
- Not a random sample: one search tool, one day and one emulated phone, with no Google results, ads or map listings; it favors firms visible in organic search and misses those that block a plain fetch.
- Presence is not performance: nothing was tested against calls, signed cases or kept promises. See the intake capacity ceiling.
- Placement is not compliance: attorney advertising rules vary by state, so ask your own counsel. A diagnosis is not by itself evidence of negligence; whether page copy implies otherwise was not assessed.
- AI scorers: sessions in one environment can share blind spots; their agreement is weaker evidence than independent human review.
- Thirty pages: one page moves a share by about 3 points, or about 6 on the past-results count (of 17); treat one or two pages as noise.
Where did the study depart from its method?
- 15 pages were scored from source, not rendered as planned.
- The second scorer’s subset, every third page, started with the third, not the first; its fetches came about 40 minutes later, so pages may have changed.
- Per-page evidence folders were not created; values rest on scorers’ notes and observed addresses, with screenshots later archived with checksums.
- The scorers’ files sat side by side, so independence rests on instructions, not separation.
- The scoring forms had no box for two new past-results values (Not adjacent, not observable); they came from the scorers’ notes (4 pages for one, 5 for the other). The boxes alone would have hidden the one real difference.
- HTML size, below 90% on exact bytes, was not rewritten: the method sets no tolerance for measured values.
What is withheld, and why?
Anything identifying a firm or vendor, because the study is about practice, not firms or products: names, firm-level values, page addresses, screenshots, identifying quotations, the search tool and the study’s software. The full rubric names firms, so it stays internal. The cost: with results changing and the tool unnamed, a re-run will draw a different sample; without firm-level values, counts cannot be rechecked page by page.
What is still open, and how are errors corrected?
Third-reviewer rulings on four items at 90% agreement (form length, Spanish option, medical sources, privacy link) may move those counts within the ranges above; the medical-sources ruling can change only the range of sources per page (0 to 3, or 0 to 5). Changes will be dated here; report errors through the contact page.
- 2026-11-03. Sample drawn and observed; past-results item rewritten and re-scored.
To check your own page, use the benchmark’s self-check.