Today's theme — The middleman stopped handing off

The referral click became a rounding error

AI answers now resolve most questions without sending anyone anywhere. What still reaches you is a shorter, later, better-informed visit — and it arrives on machine terms.

For twenty years the arrangement was simple. A search engine took a question, pointed at a website, and the website paid for itself with the visit.

The through-line in today's five items is machine mediation. A system now sits between the question and the answer, keeping more of what passes through — the click, the context, and increasingly the judgment about which businesses get named.

What follows is what changed and what it means for a firm whose phone rings because somebody found them. One of the five is a controlled experiment, and it suggests the fix has less to do with writing more and more to do with writing legibly.

01 The consumer data

AI makes the shortlist, Google makes the call

Your firm can be recommended by an AI tool and still lose the client twenty seconds later.

BrightLocal surveyed 1,227 US consumers who had searched for a local business in the previous three months. Twenty-three percent used an AI tool somewhere in that search — but only 18% of them felt ready to contact the business the AI recommended.

The rest went looking for a second opinion. Of consumers who started on an AI tool, 43% went to Google next and 39% went to social media; 2% used AI alone. Compare that with the people who started on Google Search, a third of whom stopped there, satisfied.

The useful way to hold this: AI behaves like a colleague who hands you a name, not like a referral service that closes the loop. A recommendation is an introduction, and the client still runs the background check — 75% used more than one channel before deciding.

One caveat worth carrying: the study measures consumers generally, not people hiring lawyers, where the instinct to verify probably runs stronger still.

What to do about it

Assume every AI mention gets checked within the minute. Open your Google Business Profile, your practice-area page, and your two largest directory listings side by side and confirm that practice areas, address, attorney names, and phone number match exactly. The mismatch loses the client, not the absence of an AI mention.

Source BrightLocal, “Google is #1 for local search, but three-quarters use more than one channel,” July 9, 2026 · n=1,227 US consumers — brightlocal.com · Secondary: BrightLocal, Local Consumer Review Survey — AI trust

02 The research

Same words, different structure, twice the accuracy

A controlled experiment found machines failing on well-written pages for reasons that have nothing to do with the writing.

Two researchers built two versions of the same website. Identical catalog, identical prices, identical copy — the only variable was structure: machine-readable markup, explicit action cues, and signals stating when the information was last true.

They turned three AI browsing agents loose on five tasks, 300 runs in total. The human-oriented version completed 49.3% of tasks correctly; the structured version completed 89.3%. Half-finished attempts fell from 43 to 3, and the steps an agent needed dropped from 9.31 to 6.49.

The finding matters because it isolates structure from content. The prose was identical in both conditions. A page can be accurate, current, well-written, and still unreadable to the thing doing the reading — the way a contract can be flawlessly drafted and still unenforceable because nobody signed the right page.

Hold the size of the gap loosely: the prototype was an e-commerce site, and the authors call the evidence preliminary. The mechanism, though — machines reward legibility over eloquence — runs underneath every AI citation study published this year.

What to do about it

Give your three highest-value practice-area pages four things a machine can find without inferring: a plain-language answer to the page's core question in the first 60 words, a visible last-reviewed date, attorney credentials as selectable text rather than inside an image, and jurisdictions served stated outright. Route anything AI helped draft through an attorney first — bar advertising rules attach to the words regardless of who typed them.

Source arXiv preprint (Elnaffar & Rashidi), “Designing Agent-Ready Websites for AI Web Agents: A Framework for Machine Readability, Actionability, and Decision Reliability,” July 13, 2026 · 300 runs, 3 agent models, 5 tasks — arxiv.org · Secondary: Context: critical survey of generative engine optimization, 2023–2026

03 The structural change

A court just put a price on training data

The inputs side of AI copyright now has a number attached. The outputs side is still wide open.

On July 20, Judge Araceli Martínez-Olguín granted final approval to the $1.5 billion class settlement in Bartz v. Anthropic, the largest known settlement of a US copyright case. Authors receive roughly $3,000 per work — four times the $750 statutory minimum that courts most commonly award.

The scope is narrower than the number suggests, and the narrowness is the story. Class members released only claims about how their works were acquired and copied, through August 25, 2025. Claims based on AI outputs were not released, and neither was future conduct.

For anyone whose visibility depends on these systems, the operative question is what this does to the supply of source material. When pirated corpora carry a nine-figure price and licensed corpora do not, the pool a model will confidently draw on gets smaller, more contractual, and easier to identify.

One court, one defendant, and other cases are still moving. But the order establishes that industrial-scale acquisition of protected text carries a bill — the kind of fact that changes procurement well before it changes law.

What to do about it

Nothing here demands action this week, but it changes what to build toward. Material that is cleanly attributable, dated, and unambiguously yours — your own analysis, bar association contributions, named-author commentary in trade outlets — gets more useful as models narrow toward sources they can defend citing. Anonymous, undated, syndicated content gets less useful at the same rate.

Source The Authors Guild, “Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement,” July 21, 2026 · N.D. Cal., final judgment entered July 20 — authorsguild.org · Secondary: TechCrunch reporting

04 The deadline

In September, Google starts writing your ads

Two Google Ads settings flip to automatic next month, and one of them generates copy.

Google confirmed a split timeline in an update to its April announcement. Dynamic Search Ads got a reprieve until February 2027, but automatically created assets and the campaign-level broad match setting still auto-upgrade to AI Max starting in September 2026, with eligible upgrades expected to conclude by the end of that month.

The detail worth reading twice is what switches on by default. Campaigns using automatically created assets will be upgraded with search term matching and text customization enabled — and text customization means Google generating ad headlines and descriptions from your site content, with no human drafting the line.

For most advertisers that is a performance question; Google reports an average of 7% more conversions or conversion value at similar cost with the full feature set. For a law firm it is also a compliance question. State bar advertising rules apply to machine-written ad copy exactly as they apply to copy an associate wrote — comparative claims, results, testimonials, and required disclaimers all still attach.

There is no opt-out. You can migrate voluntarily now and pick the settings yourself — the difference between configuring a system and inheriting one.

Do this before September

Open Google Ads and filter for campaigns using automatically created assets or the campaign-level broad match setting. Decide now whether text customization stays on, and add brand and text guidelines so generated copy stays inside your firm's approved language. Then put a sample of generated headlines in front of whoever handles your bar advertising compliance — before the auto-upgrade, not after.

Source Google Ads (Brandon Ervin, Director of Product Management), “We're upgrading Dynamic Search Ads to AI Max,” Published April 15, 2026 · updated June 11, 2026 — blog.google

05 The money

Five percent of conversations send anyone anywhere

The referral bargain that paid for the open web is being repriced from the revenue side.

Researchers matched URL-level Comscore US desktop clickstream against ChatGPT and Google information-seeking sessions. ChatGPT produced an outbound click in 5.2% of conversation sessions, far below Google's referral rate.

The clicks that survive are a different animal. They skew toward specialized destinations and away from ad-supported sites — worth more per visit, and far fewer of them. Meanwhile wider ChatGPT Search access cut traditional search use by 9.4%, with losses concentrated in informational queries.

Follow the incentive to the end of the line. A model that answers inside the conversation keeps the attention it used to pass along, and the advertising funding the destination page never gets the impression. The economics behind “publish helpful content, collect the visit” are coming apart at the revenue seam, not the ranking one.

The translation for a firm is uncomfortable and clarifying at once. The measure that still means something is share of voice — how often you are named, in which answers, against which named competitors, tracked over time — rather than a traffic chart quietly measuring a narrowing pipe.

What to do about it

Stop treating organic sessions as your visibility number. Write down ten questions a prospective client would actually ask — “estate planning attorney in [your city],” “how long do I have to file a claim in [your state]” — and record monthly whether your firm is named in answers from ChatGPT, Google's AI Overviews, and Perplexity. That is a share-of-voice baseline, not lead attribution; reading it as such will mislead you.

Source arXiv preprint (Shi, Zhu & Gu), “Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain,” July 8, 2026 · Comscore US desktop clickstream — arxiv.org · Secondary: Related measurement study: Google AI Overviews activation and publisher impact