01 The market data
One in six top-ten pages simply vanished
Google’s spam update removed rankings at nearly twice the usual rate, and the data cannot say which pages or why.
SE Ranking gave Search Engine Land its measurements from the August 2026 spam update, and the number that carries the story is a disappearance rate. During the update, 16.71% of URLs that had ranked in Google’s top 10 fell beyond position 100 for the same keyword. Across a comparable five-day window in July with no confirmed update running, the same measurement came in at 9.2%.
That works out to an 82% increase over the baseline, which SE Ranking flagged as its most interesting figure, and it makes a top-10 URL roughly 1.8 times more likely to fall off the first hundred results entirely. Losing a few positions is weather. Losing the entire first hundred is closer to the building being condemned.
The churn ran both ways. SE Ranking also recorded a 12% increase in the share of URLs ranking top three after the update that had not previously cracked the top 20.
No niche was spared: top-10 volatility ranged from 74.64% in real estate to 85.55% in fashion and beauty. The categories Google treats as your-money-or-your-life, which include real estate and healthcare, were the steadiest of a badly shaken set.
Read the limits before you read the meaning. SE Ranking ran no URL- or domain-level analysis of the pages that lost rankings or the pages that replaced them, and it tracked only positions 1 through 100, so a page that vanished may have been deindexed or may be sitting quietly at 104. The study measures how much moved. It does not claim to know what Google was aiming at.
What to do about it
Pull your own before-and-after rather than reading the industry average. Compare your tracked keywords on August 17 against August 22, the same two dates SE Ranking used, and separate pages that slipped a few spots from pages that left the first hundred. Those are different problems with different causes.
If a practice-area page is among the second group, resist the urge to rewrite it this week. A spam update is not a content-quality signal you can read directly, and rebuilding a page against a guess is how firms lose the rankings they still have.
Source
Search Engine Land, “Google’s August 2026 spam update hit rankings harder than normal,” by Danny Goodwin, August 27, 2026, reporting SE Ranking data. Method: 100,000 U.S. organic keywords across 20 industries; rankings compared August 17 against August 22, with a July 26 to July 31 baseline · searchengineland.com · Secondary: the update’s rollout confirmation
02 The research
The model was just as sure when it made the name up
An audit of four AI recommenders found up to 38.7% of the local businesses they named were not in the catalog at all.
Srijith Ravikumar posted an audit to arXiv on August 7 and revised it on August 12. Four zero-shot recommenders from four independent vendors, Mistral Large, Llama-3.3-70B, GPT-OSS-120B and Claude Sonnet 4.6, were asked for ten recommendations each across three catalogs, one of which was the Yelp Open Dataset. Zero-shot means the models answered from memory, with no live retrieval and no fine-tuning.
How often they named something absent from the target catalog depended almost entirely on the catalog. The rate ran 0.6% to 2.7% on a movie dataset, 11.6% to 38.7% on Yelp, and 49.3% to 61.0% on Amazon toys. Local businesses land in the middle, which is the least useful place to be: wrong often enough to matter, right often enough to be believed.
The finding that should stop you is the second one. Each model held a near-constant confidence level that barely responded to the catalog, while the actual hit rate swung 60 points across those catalogs. Ravikumar reads this as an elicitation mismatch: ask a model how sure it is and you get a generic quality rating, not a probability that the business exists.
He is careful about his own instrument, which is rarer than it should be. On identical model outputs, the reported hallucination rate moves by an order of magnitude depending on which string matcher checks catalog membership, so he validated the chosen matcher against 201 human judgments. And a conformal abstention threshold applied to the models’ stated confidence changed the hallucination rate by at most 1.65 percentage points. You cannot filter this out by telling the model to be careful.
Two caveats travel with it. This is a preprint, not peer reviewed, and it tests ungrounded models, while the assistants your clients use increasingly search the web first, which should push these rates down. What grounding does not fix is the confidence channel, and that is the part worth carrying.
What to do about it
Stop treating a model’s certainty as information. If an assistant names your firm alongside four competitors, the fluency of the answer tells you nothing about whether those four are real, current, or licensed in your state.
The practical countermeasure is to be verifiable rather than memorable. Consistent name, address and practice details across your own site, your Business Profile, bar directories and legal directories give a grounded model something to match against, which is the mechanism the paper suggests actually separates a real listing from a plausible one.
Source
arXiv:2608.10008, “Do LLM Recommenders Know When They’re Hallucinating? Auditing Confidence Calibration in Catalog Faithfulness,” Srijith Ravikumar, submitted August 7, 2026, revised August 12, 2026. Preprint, not peer reviewed. Method: four zero-shot LLM recommenders, top-10 recommendations, three catalogs (MovieLens-25M, Amazon Reviews 2023 Toys, Yelp Open Dataset), stratified by item popularity; matcher validated against 201 human judgments · arxiv.org
03 The courtroom
A judge called the publisher bargain really unfair
Google described AI Overviews as a product improvement. The judge asked whose work it was built on.
Google asked the court to dismiss Penske Media’s antitrust suit over AI Overviews, and the hearing on that motion took place Tuesday before Judge Amit Mehta, the same judge who found Google to hold an illegal monopoly in search. Jason Kint of Digital Content Next attended and posted a running account, which Search Engine Journal reported.
By that account, Google’s attorneys repeatedly characterized AI Overviews as a “product improvement” and described the long-standing traffic-for-crawling arrangement with publishers as “a vague historical course of dealing.” Mehta pushed back that the situation “all seems really unfair,” said the improvement was built “on the backs of the publishers,” and noted that product improvements “are not immune from antitrust scrutiny.”
He also observed that a market is visibly forming for generative AI inputs, with Penske pointing to OpenAI and Perplexity paying for content Google obtains at no cost.
Now the discipline. Mehta has not ruled, and questions from the bench are not a holding. Google’s position has a factual leg too: publishers can opt out of AI training, and can now opt out of AI Mode and AI Overviews as well.
The bargain was never written down, which is why it took a courtroom to start pricing it. A firm should care because the question on the table is whether search dominance can buy content for an answer that replaces the visit, and your practice-area pages are content in precisely that sense.
What to do about it
Nothing to change today, and that is the honest answer. What is worth doing is knowing where your own opt-out controls sit, because a firm that has never looked cannot say whether it is currently supplying answer text at no charge or withholding it.
Watch for the ruling rather than the commentary. If the motion is denied, the case moves to discovery and the publisher-traffic evidence becomes public, which will tell small firms considerably more about AI Overviews than any vendor study this year.
Source
Search Engine Journal, “Judge Says Google/Publisher AI Situation ‘Seems Really Unfair’,” by Roger Montti, published August 26, updated August 27, 2026, reporting Jason Kint’s account of the August 25 hearing before Judge Amit Mehta · searchenginejournal.com · Secondary: case docket, Penske Media Corporation v. Google LLC
04 The deadline
On Tuesday Google rewrites your ad settings for you
Search campaigns using broad match or automatically created assets move to AI Max starting September 1.
Beginning September 1, Google automatically upgrades Search campaigns that use Automatically Created Assets or the campaign-level broad match setting to AI Max for Search campaigns. Affected advertisers were emailed. There is no opt-in and no manual migration, and the rollout proceeds gradually through the month rather than all at once.
The two groups get different defaults. Campaigns using Automatically Created Assets are upgraded with search term matching and text customization enabled by default. Campaigns using the campaign-level broad match setting receive AI Max with search term matching enabled. Google says it is applying settings that mirror current configurations to keep behaviour as close to unchanged as it can.
Text customization is the setting to look at. It permits Google to rewrite headline and description text to fit the query. For a law firm that is not a formatting preference. Every state bar advertising rule that governs your ad copy governs copy a model wrote on your behalf, including comparative claims, results language and required disclaimers, and you are the one who defends it.
This was signposted rather than sprung. Google stopped allowing new campaign-level broad match and legacy automatically created assets across the interface, Ads Editor and API on August 3. The difference between hiring a copywriter and finding one already at your desk is consent, and the second arrives Tuesday.
Do this before Tuesday
Open Google Ads today and filter your Search campaigns for campaign-level broad match and for Automatically Created Assets. Write down which ones appear. That list is what changes on September 1.
For any campaign on the list, decide now whether text customization stays on. If your ad copy was reviewed for bar compliance, an automated rewrite puts it back outside that review, so the safe default for a regulated practice is to turn text customization off and re-enable it deliberately if you want it. Either way, screenshot the settings before Tuesday so you can tell what changed.
Source
Search Engine Land, “Google to auto-upgrade automatically created assets to AI Max,” by Anu Adegbola, August 6, 2026 · searchengineland.com · Secondary: Search Engine Roundtable
05 The money
Your next lead may open in a chat thread
Demand Gen is testing conversations that start from a YouTube ad and never touch your website.
Google’s August Demand Gen Drop, reported August 27, adds a test that lets someone start a conversation with a business through a messaging app directly from a Demand Gen ad on YouTube. The stated purpose is to shorten the journey for high-intent prospects so a viewer moves into a conversation rather than through a landing page and a form.
The same release makes Multimodal Video Creation in Asset Studio generally available, carrying an advertiser from storyboard to finished assets in one workflow, and expands travel placements to surface local activities, events and real-time offers.
Google’s own figure is that the hundreds of Demand Gen improvements shipped in the second half of 2025 are driving an average 30% increase in conversions or conversion value. No sample, no comparison group and no methodology accompanies it. That is a vendor number about a vendor product, and it belongs in the claim column.
The pattern is the one running through the whole issue. The booking, the answer, the ad copy and now the opening conversation are all moving onto surfaces you rent. The ad used to be a billboard pointing at your office; it is becoming the waiting room. A message thread that opens from a video ad is a client intake conversation, and the confidentiality and advertising rules that govern intake do not care which app it happened in.
What to do about it
If you run YouTube ads, decide who answers a message before you enable the format, and how fast. An intake channel nobody watches is worse than one you never opened, because the prospect has already been told you are reachable there.
Write the messaging script the way you would write a phone-intake script, with the same conflict check, the same no-legal-advice language and the same retention rule for the transcript. If your practice cannot retain those threads on your own systems, that is a reason to leave the format off rather than a detail to sort out later.
Source
Search Engine Land, “Google expands Demand Gen with messaging ads, travel tools and AI video creation,” by Anu Adegbola, August 27, 2026 · searchengineland.com · Primary: Google’s August Demand Gen Drop
