01 The consumer data
Fewer visitors showed up, and better ones stayed
One agency opened its whole client roster: traffic down 10.5%, conversions up.
Brainlabs pulled Google Analytics data covering all 54 of its clients and gave it to Digiday, published Tuesday. Across 14 months from January 2025 through April 2026, organic search sessions fell 10.5%, from 140.1 million to 125.4 million. Forty-six of the 54 clients saw a drop.
The inflection was September 2025, when AI Overviews began appearing on at least 30% of United States results. Referral sessions from AI platforms rose 163% over the same window, reaching roughly 200,000 a month against organic’s fall from about 20 million to under 15 million. The replacement is nowhere near one for one.
The second number is the one worth a meeting. Key events, which count actions like a purchase or a newsletter signup, rose 335% from AI referrals, and visitors arriving from ChatGPT, Copilot, Gemini and Perplexity completed one at roughly 1.5 times the rate of organic search visitors. Fewer people walk in the door, and more of them arrive with the decision half made.
Read the limits before you quote it. This is one agency’s roster, weighted toward larger advertisers, with 29 of the 54 employing at least 1,000 staff across 19 sectors and three markets. Brainlabs’ own SEO manager Sam Paez says the AI referral counts are probably conservative, because some AI visits land as direct traffic when a user copies a link into another window. And a smaller pool of higher-intent visitors is not automatically good news for a business whose economics depend on volume at the top of the funnel.
What to do about it
Split your reporting in two before your next quarterly review. One line for organic search sessions, one line for referrals from ChatGPT, Perplexity, Gemini and Copilot, and a conversion rate on each. If you only look at the combined total, a channel mix that is losing volume while gaining quality reads identically to a plain decline.
For a law firm, define the event as a booked consultation rather than a form view. Set that up now, because the comparison only becomes useful once you have two quarters of it.
Source
Digiday, “In Graphic Detail: How AI search has impacted the web traffic of over 50 advertisers,” by Sam Bradley, August 25, 2026. Method: Brainlabs first-party Google Analytics data across 54 clients (24 United States, 24 United Kingdom, six other markets; 29 with 1,000 or more staff), 19 sectors, January 2025 through April 2026 · digiday.com
02 The evidence
In legal, being cited never became being named
A study of 1,458 brands found the payoff for broad coverage stops at your industry.
Semrush, working with Kevin Indig of Growth Memo, published an analysis earlier this month that tracked 1,094 United States categories in ChatGPT monthly from January through June 2026, five prompt variants each. Behind it sit 283,215 domain-category citation observations, 76,493 brand-mention observations, and 45,578 category-expansion appearances by 1,458 mapped brands.
The study separates two things most vendor dashboards blend together. A citation is a source ChatGPT links to as evidence. A mention is your name written into the answer text. In categories close to a brand’s core expertise, brands were cited in 74% of appearances and named in 44%. In distant categories that fell to 50% cited and 25% named.
Then the finding for this audience. In legal and healthcare, even full coverage across all five of a category’s prompts left brand-mention share trending negative. Finance and real estate got citation leverage from breadth. Legal and healthcare got the citations without the name. A citation is a footnote; a mention is a referral.
Depth beat breadth on the mention side across the whole sample. A brand appearing in only one of a category’s five prompts was associated with a drop in mention share, and that penalty did not clear until the brand reached three of five. Citations showed no equivalent spread-thin effect.
Now the caveats, which the authors state themselves. This is observational, it is Semrush’s own toolkit measuring a market Semrush sells into, and the write-up says “associated with” throughout rather than claiming cause. It also stops in June, before this month’s AI Mode changes. Treat it as a well-documented pattern, not a law.
What to do about it
Pick one practice area and go deep before you go wide. A personal injury firm publishing one thin page each on estate planning, immigration and employment is the exact shape this study associates with losing name recognition rather than gaining it.
Ask any vendor reporting your AI visibility which of the two numbers they are showing you. If the dashboard cannot separate a citation from a named mention, it cannot tell you whether a prospective client would ever hear your firm’s name.
Source
Semrush, “How topical authority spreads (and where it doesn’t) in ChatGPT [Study],” by Margarita Loktionova with Kevin Indig of Growth Memo, August 3, 2026. Method: Semrush AI Visibility Toolkit, 1,094 United States categories tracked monthly in ChatGPT from January through June 2026, five prompt variants per category; 283,215 citation observations and 76,493 brand-mention observations · semrush.com · Reporting: Search Engine Land
03 The data deals
A signed contract moved the citation numbers
Publishers with an OpenAI deal earned 48% more ChatGPT citations per page.
Press Ranger and OtterlyAI released a joint study on August 20 covering 129.3 million citations across seven AI platforms in June 2026: ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini and Claude. The central comparison is that news pages from publishers with an OpenAI licensing deal averaged 10.2 ChatGPT citations each, against 6.9 for pages from publishers without one, a 48% gap.
The effect was not confined to one platform. Across all seven, the OpenAI-licensed cohort earned 46% more citations per cited page. Publishers licensed only by OpenAI showed a 112% advantage on ChatGPT specifically, while comparable home-platform effects did not appear for Google or Perplexity deals.
The part you can act on is about size, not contracts. Trade and niche publications generated most news citations in 15 of 16 United States industries, and received 213% more AI citations than mainstream media. For a firm choosing where to place a bylined article, the legal trade press is not the consolation prize for missing a national paper. On this data it is the better target.
Be careful with the causal reading. Publishers who sign licensing deals tend also to be publishers with large, well-structured archives, so the contract and the citations may share a cause rather than one producing the other. Both firms behind the study sell AI-visibility and PR tooling. A licensing deal behaves here like a press pass: it does not write the story, it gets you into the room.
What to do about it
Redraw your publicity list around trade publications in your practice area, and weight a bylined piece in a legal trade outlet above a passing mention in a general-interest paper. That is a defensible reading of this data, and it happens to be the cheaper option.
Nothing here is a route to a licensing deal of your own; those are negotiated between platforms and publishers. What it tells you is whose pages your name is most likely to be read out of when a model answers a question about your field.
Source
Press Ranger and OtterlyAI, “Study Showing Publishers With OpenAI Deals Earn 48% More AI Citations on ChatGPT,” August 20, 2026. Method: 129.3 million citations measured across seven AI platforms during June 2026, comparing citations per cited page between licensed and unlicensed publisher cohorts · globenewswire.com · Reporting: Media Copilot
04 The operational change
A server rule knocked the IRS off Google
Federal sites lost rankings, favicons and snippets, and Google did not do it.
Over several days last week the IRS website and other federal government sites dropped sharply in Google’s results. Lily Ray posted a Sistrix chart showing irs.gov losing positions on queries it had ranked first for a month earlier, among them tax brackets, free tax filing and identity theft protection.
The symptoms named the cause. Google’s listing for the IRS showed a broken favicon and no description snippet, which is what happens when Googlebot cannot fetch the page it is listing. Search Engine Roundtable’s Barry Schwartz attributed it to bot-blocking applied too aggressively, and Ray reported she could not load the site herself on desktop. By August 28 the listing had largely returned.
This is the least glamorous failure in search and the most common one at professional firms. A firewall rule or a rate limit gets tightened for good security reasons, the crawler is caught in it, and rankings decay over days with no penalty, no notice and nothing in Search Console. It is a locked service entrance: nothing is wrong with the building, the delivery just cannot get in.
The AI version is worse, because no ranking chart moves. If GPTBot, ClaudeBot, PerplexityBot or Google-Extended is blocked at the edge, the model does not report an error. Your firm simply stops being available as a source, and you find out when somebody else’s name comes back instead.
Check this week
Ask whoever manages your site to send you the current robots.txt and the bot rules in your firewall or CDN, in writing. Confirm Googlebot is allowed, then make a deliberate decision about the AI crawlers. Blocking them and allowing them are both defensible; not knowing which one you are doing is not.
Then check the cheap symptom yourself. Search your firm’s name, and look at whether your favicon renders next to the result and whether a description appears beneath it. Both missing is the same signature the IRS listing showed.
Source
Search Engine Roundtable, “IRS Website Drops Heavily In Google Search,” by Barry Schwartz, August 27, 2026, updated August 28, 2026, citing Sistrix visibility data posted by Lily Ray and analysis by Glenn Gabe · seroundtable.com
05 The money
ChatGPT switched on ads in its largest user market
Self-serve access opens September 4 with a daily budget floor near nine dollars.
OpenAI began showing ads to logged-in adults on ChatGPT’s Free and Go tiers in India on August 27, reported by TechCrunch. Go costs ₹399 a month there, and the paid Plus, Pro, Business, Enterprise and Education tiers stay ad-free. Ads sit below the response, labelled and visually separated from the answer.
The scale is the point. OpenAI said in February it has more than 100 million weekly active ChatGPT users in India, most of them on free or low-cost tiers. The rollout started with roughly 50 brands, with WPP and Omnicom as agency partners.
The date worth writing down is the second one. Self-serve access to ChatGPT Ads Manager opens in India on September 4, with daily budgets starting around ₹725, roughly nine dollars. That is not an enterprise product, and it is not priced like one.
None of this is buyable from a United States firm today, and there is no announced date for one. It matters as a read on pace. The gap between a market being switched on and any business in it being able to buy a placement was eight days, and the format that ships in India is the format most likely to ship here.
What to do about it
Put September 4 in your calendar as a reading date rather than a spending one. Look at what the ad unit actually is when the self-serve tools open, because that is the cheapest look you will get at the format before it reaches your market.
Settle the compliance question before the channel exists. State bar advertising rules apply to an ad inside a chat answer exactly as they apply to a billboard, which means comparative claims, results, testimonials and required disclaimers all carry over. Decide that with your reviewer while it is hypothetical.
Source
TechCrunch, “OpenAI to start showing ads on ChatGPT’s free and Go tiers in India,” August 27, 2026 · techcrunch.com · Reporting: CNBC
