Week of August 17 to 23, 2026. AI didn't get smarter last week. It moved further in. And it became obvious that almost nobody knows how to measure what it says about them.
Here's the week, and at the end, the one thing worth doing.
1. You asked twice and got two different answers. It's not you.
An analysis of 14,472 citations reviewed 14,472 AI citations across 1,487 local queries in 50 U.S. metro areas and ten service categories. It's the first study of that size on how AI answers pick who to cite when somebody looks for a local business.
The numbers
Google repeats nine times out of ten. Gemini, seven times out of a hundred. These are not the same kind of system, and they don't get measured the same way.
Our take
The industry is going to sell fear with this. "AI is a slot machine, hire us." We're using it the other way, because it takes the drama out of something that is ruining a lot of people's week.
Every business owner we know does the same thing. Opens ChatGPT, types "best dentist near me," doesn't show up, closes the laptop, and spends the next month convinced they're invisible. And the other version, just as common: they do show up, feel enormous relief, and don't check again for six months.
Both reactions are wrong for the same reason. They measured once.
What to do today, in five minutes
Before you measure anything, make sure there's something to measure. Open your services page and check that these three things are there in plain text - not inside an image, not in a PDF, not only behind a form:
2. AI moved onto your desktop, and now it wants to see your screen
Meta released a beta Meta AI desktop app for Mac, aimed at creators and small businesses that run their own pages. The headline feature is screen sharing, so the AI can see what you have open and answer with that context. It adds voice dictation across apps, a Google Workspace connection, and document generation. The basics are free.
And there's one line in the original coverage worth reading twice: Meta uses interactions with its AI features to train its models.
Our take
Everyone is going to share this as "free new tool for your social media." The angle almost nobody is taking:
It's the same surface we worked through in hacking your business now costs $50: what exposes a small business is almost never a sophisticated attack. It's a door somebody inside opened without thinking.
What to do, today, in two minutes
3. Seven cases that don't exist, across eight filings, with a big firm's name on them
The firm defending State Farm admitted in court to filing briefs containing seven AI-invented cases spread across eight documents, plus fabricated titles and quotes.
Our take
The industry reflex is to use this to scare people. We're not doing that.
The mistake wasn't using AI. The mistake was that eight documents went out and nobody read them.
The AI invented seven cases because that's what these models do when they don't know: they fill the gap confidently. That isn't a surprising failure. It's documented behavior.
What's actually news is the rest of it. This didn't happen to somebody rushing at midnight. It happened at a firm defending one of the largest insurers in the country, with senior attorneys and a budget. If your process is "I skim it before it goes out," that was their process too.
4. Google gave students a free year of AI, and you're the one paying for it
Google is giving U.S. college students a full year of Google AI Pro, a plan that normally runs $200 a year. It includes Gemini inside Gmail, Docs and Workspace, 5 TB of storage, and access to its agentic platform.
It renews automatically at full price when the year is up.
Our take
The obvious read is "Google is buying young users." The read that matters to a professional services business is different:
Two things, and the second takes ten seconds: write out the three complete questions somebody would use to find you (questions, not keywords - you'll use them in the protocol below), and write down the renewal date on every free trial you have open. Today.
5. The rules are being written down now, and the platforms went first
OpenAI published, for the first time, the list of what its model must not do with a minor: no romantic language, no feeding emotional dependency, no implying it has feelings.
And Anthropic opened up its internal AI fluency training for free. The useful part isn't the course. It's the rule it teaches: verify in proportion to what's at stake.
Our take
These look like big companies talking to each other. They aren't. If the platforms are writing their own manual, the business that never wrote one is running on somebody else's. Your WhatsApp assistant already has rules: the ones your vendor set, or the ones the model inferred on its own. The only open question is whose they are.
A badly checked caption costs you a few minutes. A document that goes out with your signature costs something else entirely.
6. A third of the web already reads like AI, and that works in your favor
Pew Research Data Labs reported on August 20 that more than a third of English-language pages published since ChatGPT launched show signs of AI authorship, based on their July 2026 snapshot.
Our take
That number always gets shared as a complaint. It's the opposite.
Where this fits inside the whole job is in the five disciplines map. Copy that anyone could have written has already been written by anyone, thirty-three times over.
The one thing to do: the 3x3x2 protocol
Three questions, three different days, two engines. Eighteen checks.
Three settings first - they change the result
The three questions
Written the way a customer would write them, not the way an expert would:
- "I need a [your service] in [your city]"
- "Who's the best [your profession] near [your area]"
- "[Your typical problem], who do I call in [your city]"
How to score it
Six boxes, one per question-and-engine pair. Three observations in each.
Don't add the eighteen into one number. What reads clearly is the pattern:
If the same competitor keeps showing up in your spot: open their site and look for the three plain-text answers from section 1. They almost always have them and you don't. And they may be showing up there without ranking on Google at all: ChatGPT cites pages Google won't rank.
AI didn't get smarter last week. It moved further in. Into your screen, into your documents, into the way your next customer is learning to ask. And at the same time it became clear that the way almost all of us check our presence there doesn't measure anything at all.
Run the 3x3x2 this week. Signed out, over three days, in both engines.
If the boxes come back at zero and you don't know where to start, we'll look at it together on a free call.
Keep reading
- AI doesn't guess, it skips you - why they skip past you when they can't find what they need
- ChatGPT cites pages Google won't rank
- 44% of LLM citations come from the first 30% of your page
- A ChatGPT citation won't pay the bills - what to do when you do show up and the phone still doesn't ring
- AI search has no page two
- The five disciplines map - the whole framework
- Hacking your business now costs $50
- Last week's recap
Sources
- Analysis of 14,472 AI citations in local search - published August 6, 2026, data collected July 30, updated August 20 - single-vendor study, no independent replication - https://searchengineland.com/business-websites-gemini-citations-local-ai-search-study-485506
- Meta AI desktop app for Mac, MacRumors, August 19, 2026 - https://www.macrumors.com/2026/08/19/meta-ai-mac-app/
- AI-fabricated cases in the State Farm litigation, CalMatters, 2026 - https://calmatters.org/economy/2026/08/state-farm-lawsuit-ai-hallucinations/
- Free year of Google AI Pro for college students, Google blog, August 19, 2026 - https://blog.google/innovation-and-ai/products/gemini-app/student-offer-google-ai/
- OpenAI Model Spec update with teen protections, August 18, 2026 - https://openai.com/index/updating-model-spec-with-teen-protections/
- Anthropic approach to teaching and learning AI, August 21, 2026 - https://claude.com/blog/anthropics-approach-to-teaching-and-learning-ai
- Pew Research Data Labs, how much of the internet is written with AI, August 20, 2026 - https://www.pewresearch.org/data-labs/2026/08/20/how-much-of-the-internet-is-written-with-ai/
- Eight Oh Two, 2026, on 37% of searches starting in AI
