01 The local data
Gemini’s favorite source about you is you
Six in ten of its local citations point at the business’s own website. ChatGPT, same queries, points there 15.9% of the time.
Ben Fisher of Steady Demand ran 1,487 queries through the Gemini API’s grounding tool on July 27 and 28, covering the fifty largest US metro areas, ten local-service categories and three phrasings of each. Personal injury lawyer was one of the ten categories. The run produced 14,472 citations across 3,611 unique domains.
Gemini cited the business’s own website for 59.9% of those citations. Directories, review platforms and community forums, added together, did not match it. Run the identical queries through ChatGPT’s search mode and the business’s own site falls to 15.9%, while social and community forums climb to 41.7%.
The two engines barely agree on anything. Their cited-domain sets overlap 8.3% on the same query. Where both named a recognizable business, they picked the same top one 4.2% of the time.
Then the control test, which is the finding that should change how you check your own visibility. Fisher ran the same queries five times, five minutes apart, against both Gemini and Google’s ordinary local pack.
Google’s pack returned the same top business in 90.2% of comparisons. Gemini managed 7.9%. A witness whose account changes between tellings is not a witness you build a case on.
Read the caveats, because Fisher published them himself. This is the Gemini API’s grounding tool, not the consumer Gemini app, it is a single snapshot, and its own audit found nine of the fifteen most frequent classification defaults miscategorized before correction. Yext’s much larger citation study also disagrees about how much Reddit matters, and Fisher sells local SEO.
What to do about it
Asking ChatGPT once whether it recommends your firm tells you almost nothing. Ask five times across a week and count how often your name comes back, because a single appearance sits inside the noise.
The durable finding is that on Gemini your own website is the source it reaches for most. That is the asset you control outright, and it is carrying more weight than your directory listings.
Source
Steady Demand, “AI Search Local Search Citation Ledger: Methodology & Study Design,” Ben Fisher, August 6, 2026, dataset collected July 27 and 28, 2026 · steadydemand.com · Reporting: Search Engine Land, Danny Goodwin, August 19, 2026
02 The research & evidence
Someone finally ran the experiment, not the correlation
A preregistered field study of 1,100 people on live Google Search, testing what AI answers do to publisher clicks.
Every figure you have read about AI Overviews costing publishers traffic came from observation. Somebody watched sites with an AI answer above them and compared them to sites without one. That is a correlation, and it cannot separate the AI answer from whatever else made those queries different.
Stephanie Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danaé Metaxa did the other thing. They ran a preregistered field experiment with 1,100 participants on live Google Search, manipulating whether AI Overviews and AI Mode appeared at all. The paper went up on arXiv on August 18.
Removing AI Overviews and AI Mode raised click-through rates to publishers. An AI Mode only experience lowered them, and also eroded how users rated the experience and how far they trusted what they found on Google. Preregistration is what makes this worth reading: the authors committed to what they would measure before they measured it.
Now the honest part. The abstract reports directions, not magnitudes, so anyone citing this as a specific percentage is reading something that is not on the page. It is a preprint, it has not been through peer review, and a study of general search cannot tell you what happens to one practice-area page in Tampa.
What it does give you is a sentence you can defend. AI answers reduce clicks to the sites they draw from, and that is now a causal finding rather than a pattern in a spreadsheet, which is more than the 58% and 39.8% figures in circulation can say for themselves.
What to do about it
Stop quoting a click-loss percentage you cannot trace to an experiment. When a vendor deck puts a number on what AI search costs you, ask which arm of which study produced it.
For your own site, track Search Console impressions against clicks on the queries where an AI answer appears, over months rather than weeks. Anything shorter is noise.
Source
arXiv, “AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence,” Wang, Gleason, Bart, Wilson & Metaxa, August 18, 2026 · arxiv.org
03 The platform shift
Google gave readers a button that picks you
One embed, one click, and your domain is wired into that reader’s AI Overviews.
Google shipped an embeddable Preferred Sources button on Thursday. A publisher places it on their own page, a reader clicks it once, that site is added as a preferred source, and the reader lands back where they were reading. No settings menu, no account to go hunting through.
Preferred Sources feed Top Stories, AI Overviews and AI Mode. That is the part worth sitting with, because until now no lever a publisher controlled reached into an AI answer surface at all. More than 600,000 unique sources have been selected so far.
The limit is real and Google is not hiding it. This is per-reader personalization, not ranking. Nobody who has never heard of your firm sees more of you because you embedded a button; it moves only the people already reading you.
For a firm that publishes on any kind of schedule, that is still worth having. A reader leaving a standing order at your shop counts for more than a stranger walking past the window, and until this week there was no counter to leave it at.
The release also added a control for telling Discover in your own words what you want more of, and tuning for News audio briefings. All of it hands the dial to the reader, which is the direction of the whole announcement.
What to do about it
If your firm publishes a blog or a newsletter, take the embed code from Google’s Search Central documentation and put the button on the pages people read to the end.
Then be clear about what it buys. It hardwires you into the answers of people who already chose you once, and does nothing for the stranger describing their problem to an assistant for the first time.
Source
Google, “Personalize what you see on Google Search, News, Discover,” Mrinalini Loew, August 20, 2026 · blog.google · Reporting: TechCrunch, August 20, 2026
04 The deadline
Google upgrades your search campaigns starting September 1
Campaign-level broad match and automatically created assets get migrated in place, whether or not anyone asked.
On August 3 Google stopped letting anyone create campaign-level broad match or legacy automatically created assets, across the interface, Ads Editor and the API. That was the quiet half of this. Between September 1 and September 30, Google migrates the campaigns already using them into AI Max automatically.
The migration happens in place, with what Google calls equivalent settings, rolling out gradually through the month. Campaigns running automatically created assets arrive in AI Max with both search term matching and text customization switched on. Campaigns running campaign-level broad match arrive with search term matching on.
Text customization is the line a regulated advertiser should stop at. The headline a prospective client reads may not be one anyone at your firm wrote. State bar advertising rules apply to that sentence exactly as to one you drafted, and the tool having generated it has never worked as a defense.
Search term matching is the second one. It reaches past the keyword list you built, which in a legal account is the difference between paying for someone looking for a lawyer and paying for a student looking up a statute.
Dynamic search ads are on the same track one cycle behind: in-account notices in September, reminders on January 15, and automatic migration between February 1 and 28, 2027, after which new dynamic search ad groups cannot be created at all.
Do this before September 1
Open Google Ads and filter for campaigns using campaign-level broad match or automatically created assets. Those are the ones on the September list, and the account will not warn you on the morning it happens.
Before the migration lands, write your brand exclusions and negative keyword list, and decide whether text customization stays on. Whatever it generates runs under your firm’s name, so route it through the same compliance review as any other advertisement and record that the review happened.
Source
Search Engine Land, “Google sets AI Max migration timeline for Search campaigns,” Anu Adegbola, August 14, 2026 · searchengineland.com
05 The money
The ad slot moved next to the client’s problem
OpenAI took ChatGPT Ads into 31 more countries and put a self-serve buyer on the near horizon.
OpenAI expanded ChatGPT Ads into 31 European markets on Tuesday, its largest advertising expansion so far and roughly six months after the US pilot began. A self-service Ads Manager is due later this summer.
The buying controls are the news here, not the geography. The platform already supports conversion optimization, geo-targeting, custom audiences, an OpenAI pixel and a conversions API. Together those are what would let a single-market practice buy anything on it at all.
Ads appear only to people on the Free and Go plans. Paying subscribers do not see them, which is both a real limit on reach and a signal about who OpenAI thinks the inventory is for.
Consider where that inventory sits. A search ad interrupts somebody typing four words into a box. An ad inside an assistant conversation sits beside a person describing their problem in full sentences, which is a good deal closer to an intake call than to a keyword.
That proximity is also the compliance problem. Bar advertising rules follow the message rather than the medium, so comparative claims, past results and required disclaimers all apply the same way. A jurisdiction requiring an advertising label does not stop requiring it because the ad arrived inside a chat.
What to do about it
Nothing new to buy in the US yet, so use the gap. Write the disclaimer language your jurisdiction requires now, while no live campaign is waiting on it.
When the self-service manager opens, treat geo-targeting as the thing that makes this affordable for a local practice, and read the placement in a real conversation before approving spend.
Source
OpenAI, “ChatGPT Ads expands across Europe,” August 18, 2026 · openai.com
