01 The consumer data
One page in ten was written by a machine
Pew scored 490,000 pages for AI authorship. The tells it measured are the ones this brief is written to avoid.
Pew Research Center published the count on Thursday. Using Common Crawl snapshots from January 2021 through July 2026, its data-labs team sampled about 490,000 English-language pages and scored them with Open Pangram, an open-weight detection model. In a random sample of 10,000 pages collected in July 2026, 10% showed significant signs of AI authorship.
That headline number understates the trend, because most of the web is old and could not have been machine-written. Filter to pages published after ChatGPT’s November 2022 release and the signs appear in over one third of them.
The distribution is lopsided. Around one in ten .com pages carry the markers, against 4.6% of .org pages and roughly 1% of .edu and .gov pages. Commercial writing is where the machine arrived first, which means your competitor’s practice-area page is likelier to be machine-written than almost anything else you will read this week.
Then comes the part that reads like an audit of this newsletter’s own style sheet. Pew measured specific tells against a 2023 baseline. Em dashes appear about twice as often, Oxford commas are up 63%, and words the models reach for, including “delve” and “testament,” have more than doubled.
The “it’s not X, it’s Y” construction has nearly tripled, though it stays rare in absolute terms.
Document-level detection stays unreliable and Pew says so without hedging. These patterns hold across very large collections, not on any single page. They are fingerprints, not a confession.
What to do about it
If your firm publishes AI-assisted content, the human review step is not cosmetic. Read for the markers Pew measured and cut them, because the patterns that mark a page as machine-written are the same ones a reader notices before they can name why.
Your advantage is the material a model cannot invent. The matter you actually handled and the local rule you actually argued belong on the page, in specifics.
Source
Pew Research Center, “How Much of the Internet Is Written With AI?,” Bestvater, Smith, TerBush, Baronavski & Chavda, August 20, 2026 · pewresearch.org · Reporting: TechCrunch, August 20, 2026
02 The research & evidence
Pages built to be cited are now countable
A new detector puts GEO-optimized content at 8.90% of the web, and 16.36% of pages modified this year.
Generative engine optimization, or GEO, means editing a page so that a chatbot is likelier to quote it. Until now nobody could say how much of it was happening. A team led by Junjie Chu posted a paper to arXiv on August 17 that tries to count it.
They built a benchmark called GEOFlagBench, 3,200 web pages spanning 400 queries, four subject domains and eight families of GEO tools, then tested whether existing detectors could separate an optimized page from an ordinary one. The strongest existing method reached an F1 score of 0.880, which sounds respectable until the authors break it down and find it partly keying off whether text merely looks AI-written.
Their own approach trains a detector on paired before-and-after versions of the same page. That lifted F1 to 0.944 and worst-group accuracy from 0.725 to 0.883. They then pointed the finished pipeline at live results: Google Search and Gemini-grounded retrieval for 1,000 real user queries, across 10,095 pages.
The estimate is 8.90% GEO prevalence overall, and 16.36% among pages modified in 2026. Roughly one page in six of this year’s crop has been dressed for the interview rather than qualified for the job. The authors’ stated worry is that optimization can hand a page visibility out of proportion to its authority, and can make thin claims look well supported.
The limits are worth stating plainly. This is a preprint and has not been peer reviewed, the prevalence figure rests on a detector with a known error rate, and the query set is general rather than legal. Treat the direction as solid and the decimal places as provisional.
What to do about it
The finding that should change your behavior is not about your own website. It is about the pages a model reads before it answers a question about your practice area, a growing share of which were built to be quoted.
When you check what an assistant says about your firm, open the sources it names. A cited page with no author, no date and no primary document behind it is still shaping the answer.
Source
arXiv, “GEO-Flag: Detecting and Measuring GEO-Optimized Web Content,” Chu, Leng, Li, Shen, Shen & Zhang, submitted August 17, 2026 · arxiv.org
03 The platform shift
Google will run your ads before the badge clears
Local Services Ads now serve during onboarding, ranked below every provider who finished.
Google added pre-badge ads to Local Services Ads this week. A business that clears preliminary checks can start receiving leads while the rest of onboarding is still running, instead of waiting for the Google Screened or Google Guaranteed badge to appear.
The preliminary bar is specific and mostly paperwork you already hold. A license or business registration, reviews on the profile, completed billing, and for professional services a headshot. Pre-badge ads are excluded for garage door services, health care and locksmiths.
There is a cost, and Google states it rather than burying it. Pre-badge listings are placed below every provider who has completed onboarding and earned the badge. You are buying earlier presence at the price of position, which is the difference between being introduced and introducing yourself.
Google also rewrote the eligible category list into something far more granular. Google Verified badge exclusions widened to cover verticals in the auto, beauty and dining categories.
For a law firm this matters because Local Services Ads sit above the map pack and above the ordinary search ads. Google told Search Engine Roundtable the migration of Local Services Ads into Google Ads runs through 2027, so the format is being built on rather than wound down.
What to do about it
If your firm has a Local Services profile stalled in onboarding, check whether the pre-badge path is open to you. The gating items are ones you can supply the same day: license documentation, billing details and a current headshot.
Check whether your practice area moved, too. The category list was rewritten this week, and a more granular category changes which searches you appear for.
Source
Search Engine Roundtable, “Google Local Service Ads Pre-Badge Ads & Expansion Of Verticals,” Barry Schwartz, August 21, 2026 · seroundtable.com · Primary documentation: Google Local Services Help
04 The operational change
Your schema markup may have stopped parsing
Googlebot now unescapes HTML once. Characters that used to be repaired for you now break the block.
Google changed how Googlebot extracts JSON-LD, the structured data block that tells Search what your page is about. In Google’s own words: “To bring our parser up to JSON and other standards, we changed our JSON-LD extraction and are now only applying a single pass of HTML unescaping.”
In plain terms, the crawler used to clean up malformed data on your behalf. It no longer does. Double-escaped entities such as & or ✔ will no longer be unrolled, so a block containing them can fail to parse instead of being quietly repaired.
This is a wiring fault behind a wall. Nothing looks wrong until something that depended on it stops working. A firm name with an ampersand in it, emitted by a template that escapes twice, is exactly the case that breaks, and it breaks without an error message anywhere an owner would think to look.
Structured data is how a machine learns your firm’s name, address, practice areas and attorney entities without guessing at them. A block that fails to parse takes all of it out at once, and the rich results built on it go with it. Gary Illyes pointed readers to RFC 8259 section 7 for what correct JSON escaping actually is.
Google announced this in a post rather than a changelog entry, which is a good way for a change to reach a lot of sites late. Nobody has described a grace period.
Do this before Monday
Run your key pages through Google’s Rich Results Test today: the home page, each practice-area page, each attorney bio. It reads the markup the way Googlebot now reads it, so a block that has started failing will fail there too.
If your site runs on a template platform, look specifically at fields holding ampersands, quotation marks or checkmark characters. That is where double-escaping happens, and the fix is standard JSON escapes or Unicode hexadecimal escapes such as \u0026.
Source
Search Engine Roundtable, “JSON-LD Extraction For Googlebot Now Does One Pass Of HTML Unescaping,” Barry Schwartz, August 21, 2026, quoting Google · seroundtable.com · Standard: RFC 8259 §7, IETF
05 The money
Microsoft takes the bid ceiling away October 1
Max CPC will not be available on new standalone-bidding campaigns after that date.
Microsoft Advertising emailed advertisers on August 20 to say that Max CPC “will no longer be available when creating new non-portfolio campaigns” after October 1, 2026. Max CPC is the manual ceiling on what you pay for a single click.
The scope is narrower than the headline suggests. It applies to new campaigns using standalone Maximize Conversions, Maximize Conversion Value and Maximize Clicks bidding. Existing campaigns keep their Max CPC, and portfolio bid strategies keep it as well, along with target impression share and enhanced CPC.
Navah Hopkins, Microsoft’s advertising liaison, gave the reasoning publicly. Max CPC limits override an advertiser’s stated goals and can cause spend pacing irregularities, and advertisers using target cost per acquisition or target return on ad spend hit their goals more easily than those leaning on the older control.
A bid ceiling is a governor on an engine. Removing it makes the target the only thing standing between a new campaign and an expensive week, and a target is only as good as the conversion data feeding it.
Hopkins also said what to do about it, which is unusually direct for a deprecation notice. Run an optimization experiment now that strips Max CPC out of an existing campaign, so you learn how bidding behaves without it while the option is still there to put back.
Do this before October 1
If you plan to launch anything on Microsoft Advertising this autumn, build it before October 1 if you want a manual ceiling on it. After that date, new standalone-bidding campaigns run on targets and budgets.
Better use of the next five weeks: test one existing campaign without Max CPC and watch cost per qualified lead. That number tells you whether the ceiling was doing real work or just making the account feel safer.
Source
Search Engine Land, “Microsoft Advertising removes Max CPC from new standalone bidding campaigns,” Anu Adegbola, August 20, 2026 · searchengineland.com · Secondary: Search Engine Roundtable, August 21, 2026
