01 The demand signal
Your client checked your advice against a chatbot
A sixth of people in a legal dispute asked an AI, and some of them were auditing their own lawyer.
JUSTICE, a UK legal charity, and the Administrative Fairness Lab surveyed 1,428 people who had been involved in a legal dispute in the previous two years. About a sixth used a chatbot such as ChatGPT during it. Among respondents aged 18 to 24 that figure was 26%. Among 55 to 64 year olds it was 10%.
The list of what they asked for is where this stops being a story about young people. Respondents used chatbots to explain legal jargon, to draft complaints and emails, to get emotional reassurance, and to check advice they had already received from a lawyer. Housing, employment, debt and consumer disputes came up most often, which are the problems that fall outside legal aid.
Only 6% used a chatbot and nothing else, so for most people this supplements professional help rather than replacing it. That is the reassuring number. The other one is that a client who runs your letter past ChatGPT and gets a confident contradiction has no way of telling which of you is wrong.
The researchers flagged something worth sitting with. JUSTICE notes evidence that chatbots built to sound warm are more likely to reinforce a user’s existing beliefs, and that the most vulnerable-sounding users get the least accurate answers. The client most likely to argue with you is the one who was frightened when they typed the question.
Two limits on the data. The sample is British, and fieldwork ran from late 2025 into January 2026, so this describes the winter rather than this week. The direction is solid. Treat the exact percentages as a floor.
What to do about it
Put the reasoning into client letters, not just the conclusion. A chatbot handed a paragraph of bare conclusions will fill the gap with its own; a paragraph that shows why survives the audit.
When a client comes back with a chatbot’s version, ask to see the exact prompt before you argue with the answer. Most of these disagreements turn out to be the model answering a slightly different question.
Source
JUSTICE and the Administrative Fairness Lab, “What AI Chatbots Can Teach Us About Unmet Legal Needs,” 2026 · justice.org.uk · reporting: Legal Futures
02 The evidence
The file your plugin wrote blocks nothing
584,107 llms.txt files, read line by line, and the opt-outs inside them are decorative.
Common Crawl published a content analysis of 584,107 llms.txt files on August 31, drawn from its July 2026 crawl. If the term is new to you, llms.txt is a proposed convention: a small text file at the root of a site offering AI systems a curated map of it. Earlier studies counted who publishes the file and who fetches it. This is the first to read what the files say.
68.27% of the corpus came out of a template, and Wix alone accounts for 41.34% of it. Half the files carry the full structure the specification asks for, which sounds like healthy adoption until you look at what fills that structure. 22.56% contain no links at all, in a format whose entire purpose is a curated list of links, and 2.54% is a parked domain advertising itself for sale to a language model.
The finding that matters to your firm is about blocking. 6.59% of files carry policy language the specification never defined: rate limits, copyright notices, demands to be cited. 1,570 files name a specific crawler and rule on it. Common Crawl checked the 32 sites whose llms.txt denied Common Crawl’s own crawler, and found none of them blocked it in robots.txt, which is the file that actually carries weight. Their conclusion: “Publishers writing opt-outs there have not opted out.”
A file at a predictable path that agents are encouraged to fetch, and that no security tooling inspects, is also an interesting place to leave instructions. The analysis found 3,793 files carrying mild steering and 102 telling the model to recommend a product or avoid naming competitors.
Common Crawl is careful about what this cannot tell you. It is one crawl and one snapshot, so nothing here says whether any of it is growing, and the sample covers only hosts Common Crawl fetches successfully.
What to do about it
Type your domain followed by /llms.txt into a browser. If a file loads and you did not write it, your SEO plugin did, and it is currently speaking for your firm to every agent that asks.
If anyone has told you that file keeps AI systems away from your content, that is wrong. Crawler directives belong in robots.txt, and even there they are a request that well-behaved crawlers honour rather than a lock on the door.
Source
Common Crawl, “A Content Analysis of llms.txt Files from the July 2026 Crawl Archive,” by Malte Ostendorff, August 31, 2026 · commoncrawl.org
03 The platform change
Claude now signs its own writing, invisibly
A statistical fingerprint in the word choices, switched on worldwide to satisfy a European rule.
Anthropic has begun watermarking the text its newer Claude models produce, and updated its explainer on September 1 with details of the detection tool. The trigger is the EU AI Act. Around 190 organisations signed the EU Code of Practice on Transparency of AI-Generated Content in July, and the marking requirement took effect on August 2.
Nothing is inserted into the text. There are no hidden characters and no invisible Unicode, which is what most people picture. The model biases a long series of coin-flip word choices using a secret key, so the passage carries a statistical pattern that a holder of that key can test for.
What the watermark does not do is the part worth knowing. It carries no identifying information and cannot be traced to a person, a firm, or a conversation. It cannot establish that a human wrote something, only that Claude probably did. It fails on short passages, thins out on factual writing where the wording is forced, and barely registers when Claude has only proofread your draft. Anthropic states plainly that it changes nothing about ownership or legal responsibility.
Anthropic applied it globally rather than only in Europe, saying it has no durable way to scope the change by region yet. The detection API is in private preview, limited for now to regulators, law enforcement, media, fact-checkers, researchers and companies with their own obligations under the Act. Other providers signed the same code and are building their own versions with their own keys.
The near-term exposure for a firm is not that anyone can prove your blog post was drafted by a machine. It is that within a year a bar regulator, an opposing party or a client may be able to ask and get a probability back. Anthropic announced this on August 14, so it has been running for three weeks already.
What to do about it
Document the human review step on anything you publish. Not because a watermark creates a new duty, but because the question of whether a person checked it is about to be answerable by someone other than you.
Do not buy an AI-detection tool on the strength of this. Detection software guesses at writing tells and has no access to any watermark key; the two get confused constantly, and the guessing kind produces false accusations against people who write cleanly.
Source
Anthropic, “How Claude’s text watermark works,” August 14, 2026, updated September 1, 2026 · anthropic.com · analysis: Search Engine Land
04 The deadline
Your search campaigns are being migrated this month
Google starts moving Broad Match and auto-generated assets into AI Max, in place, without asking.
Beginning September 1 and running through the month, Google is automatically upgrading Search campaigns that use campaign-level Broad Match or standalone Automatically Created Assets to AI Max. Google says campaigns move in place using equivalent settings, and that existing brand inclusions and exclusions carry over.
The gate closed a month ago. On August 3 Google stopped allowing anyone to create the legacy structures across the Ads interface, Ads Editor and the API. The migration is the second half of a decision that was already made.
In plain terms, AI Max is a bundle of settings that lets Google match your ads to searches you never listed and write ad text from your landing pages. Dynamic Search Ads get a reprieve until February 2027, with in-account warnings starting this month.
One detail is easy to miss and expensive to discover late. Google’s own help documentation now notes that broad keywords are not treated as exact match in AI Max prioritisation the way they behaved when you switched the broad match setting off. If your account relied on that, an unchanged keyword list will behave differently after the migration than before it.
For a firm buying legal search terms this lands harder than it does for a retailer. Loose matching on legal intent puts your budget against searches you would never have chosen, and the ad copy Google generates from your pages is subject to your state bar’s advertising rules exactly as copy you wrote yourself is.
Do this before the end of September
Check this week whether any campaign still runs campaign-level Broad Match or standalone Automatically Created Assets. If it does, it will be migrated before the month is out.
Before it moves, export a 90-day search terms report and save your negative keyword list somewhere you can restore it from. After the migration, read the ad text Google generated from your pages against your bar’s advertising rules in the first week, not the fourth.
Source
Search Engine Land, “Google sets AI Max migration timeline for Search campaigns,” by Anu Adegbola, August 14, 2026 · searchengineland.com · primary: Google Ads Developer Blog
05 The money
ChatGPT ads reached a billion in two hundred days
And as of Sunday, a small firm in Europe can buy them without going through an agency.
OpenAI says its advertising business is running at a $1 billion annualised rate, six months after launch. Read that number carefully. It is one month’s ad revenue multiplied by twelve, a snapshot of today’s pace rather than money collected. Digiday puts actual booked revenue across the first eight months nearer $330 million.
The change that reaches you is about distribution. On August 31 OpenAI opened beta self-serve access to its Ads Manager across the 31 European markets where ads launched earlier in the month. Until then, buying meant going through an agency or a tech partner. Now a two-person business can build and manage a campaign in the app directly, subject to category approval and a policy check.
Ads run against the free tier and the Go subscription plan, carry labels, and OpenAI says they are kept separate from how the model composes its answers. Advertisers cannot see what individual users discussed.
The company is targeting $2.5 billion in recognised revenue for 2026. Reaching that from a roughly $83 million month means the fourth quarter has to be violent, which tells you how hard the pressure on inventory is about to get. Expect the ad load inside ChatGPT to rise before it settles.
None of which is a reason to buy yet. There is no independent measurement of this inventory, no third-party verification, and no published benchmark for what a click out of an AI conversation is worth next to a Google search click. Anyone quoting you a cost per lead for this channel is quoting you a guess.
What to do about it
If you test it, cap the spend at a number you would not notice losing and run it for a full month. A channel with no benchmarks needs your own baseline more than it needs a strategy.
Route the creative through the same compliance review as any other advertisement. A new surface does not create an exemption, and the disclaimers your state bar requires travel with the message rather than with the medium.
Source
Digiday, “OpenAI’s ChatGPT ads business hits $1 billion run rate as Europe gets self-serve access,” by Seb Joseph and Krystal Scanlon, August 31, 2026 · digiday.com
