54 Advertisers, 140M Sessions: AI Search Took 10.5% and Gave Back 6%

On 25 August, Digiday published something the AI-search debate has been short of: not a survey, not a sample, but one agency's entire client roster in raw Google Analytics. Marketing services group Brainlabs handed over 54 advertisers across 19 sectors, spanning January 2025 to April 2026 — fourteen months and more than 140 million sessions. The headline reads in one line: organic search sessions fell from 140.1 million to 125.4 million, down 10.5%, while referrals from AI platforms grew 163%, AI-driven "key events" grew 335%, and the key-event rate from AI referrals ran 1.5 times higher than from organic search.

The reason this circulated through the GEO community within 48 hours is not the minus 10.5%. Everyone had priced in the traffic loss. Ahrefs, comparing 300,000 keywords in Search Console data, found top-ranking pages losing up to 58% of their clicks when an AI Overview appears; US zero-click search reached 68% in early 2026. What got shared was the second half: traffic down, quality up. For the entire AEO/GEO tooling category, that sentence is a life raft. The last eighteen months burned a lot of capital here — Profound hit unicorn status at a $1 billion valuation in February 2026, Peec AI closed a $21 million Series A, and both Semrush and Ahrefs bolted AI-visibility modules onto their existing suites — yet the category has never been able to answer the one question every CMO asks: does visibility convert into money? "1.5x" looks like the answer.

Compared with its peers, this dataset is genuinely more honest about sampling. The Ahrefs study is keyword-level CTR and cannot see conversions. Panel-based estimates from the Similarweb school cannot either. Vendor case studies almost always feature the client who won. Brainlabs published the whole roster including the losers — 46 of the 54 clients declined, and it said so. The price of that honesty is a different bias: all 54 are managed by the same agency, 24 in the US, 24 in the UK, six elsewhere, and 29 of them employ at least 1,000 people. This is not a sample of the internet. It is a sample of large brands.

For a small business or a freelance studio, the number that matters is not the 10.5%. It is whether the 1.5x transfers to you. It probably does not, and the reason has nothing to do with effort. What follows: the full figures, what three types of reader should do this week, how the monitoring tools compare, the division this report conspicuously did not perform, and how to reproduce the whole analysis without paying a subscription.

The Numbers in Full

The skeleton of the Brainlabs dataset:

Window: January 2025 to April 2026, fourteen months.
Sample: 54 advertiser clients — 24 US, 24 UK, six other markets; 29 with 1,000+ employees; 19 sectors including 16 retailers.
Inflection: September 2025, when AI Overviews passed 30% coverage of US search result pages. Sam Paez, Brainlabs' senior SEO manager, names that month as the hinge.
Organic: 140.1M to 125.4M sessions, −10.5%; monthly average from roughly 20 million to under 15 million.
AI referrals: up 163%, running around 200,000 sessions a month after the hinge.
Conversion: AI-driven key events up 335%; key-event rate 1.5x organic for ChatGPT, Copilot, Gemini and Perplexity referrals.
Sector spread: fitness, fintech, insurance and CPG fell hardest; retail, beauty and entertainment fell least and gained the most AI referral lift.

Who lost is clear enough: any category whose funnel begins with research. Paez's reading is that AI Overviews are still overwhelmingly triggered by educational, informational and discovery keywords, and have only started leaking into transactional ones. That is why retail held up.

Who gained is murkier. Brainlabs buries a disclaimer: its AI referral estimate is probably conservative, because some users open a separate tab and type the URL rather than clicking through, and GA4 logs those sessions as direct. That has been read almost universally as "so it's actually better than stated." But the defect runs both ways. If part of the AI traffic lands in direct, part of the organic "loss" may be reclassification rather than disappearance.

What Three Types of Reader Should Do

Brand owners and SMB operators

Do not move budget into GEO yet. Open GA4, pull "session primary channel group" for the last fourteen months, and find out what your own organic decline actually is. If it is clearly under 10%, AI Overviews have not reached your category.
Split your keyword set into informational and transactional. Google is internalising the value of the informational half; stop commissioning the tenth "what is X" article. The transactional half is your foundation.
Demand one absolute number from your marketer or agency each month: how many real sales came from AI sources. Not a growth rate — a count. If it is single digits, this is not yet a channel for you.

Marketing and SEO practitioners

Build a custom channel group in GA4 today grouping chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com as "AI Referral." Channel groups do not backfill, so every day you wait is a day of missing baseline.
Audit your key-event list. If it contains scroll, page_view or engaged session, your conversion rate is inflated. Google's own documentation defines a key event as an action "particularly important to the success of your business" — the threshold is yours to set, not Google's.
Track citation rate and referral traffic as two separate metrics. Being cited is not being clicked, and being clicked is not being bought from.

Developers and agencies

Pull server logs and filter User-Agent for ChatGPT-User, PerplexityBot, GPTBot, ClaudeBot and Google-Extended. Crawl frequency is a leading indicator of citation, and it costs nothing.
Convert client FAQs, specifications and pricing into structured data. An LLM can lift a structured field; it cannot lift a price list embedded in a JPEG.
Package an "AI visibility baseline report" as a standalone phase-one deliverable. Three hours of work, separately billable, and a hook for the contract that follows.

AI Visibility Tools Compared

Tool	Position	Public entry price (monthly)	Best for
Profound	Enterprise AI visibility and answer analytics	From ~US$99 (ChatGPT only); realistically US$399+	Dedicated in-house teams needing multi-engine, multi-region splits
Peec AI	Lightweight European monitor, transparent pricing	~€89–199, plus €20–30 per extra engine	Mid-market brands establishing a first baseline
Scrunch AI	Multi-LLM coverage with agency multi-client management	~US$250–300	Agencies running several brands at once
Semrush AI Toolkit	Add-on module to an existing SEO subscription	~US$99 add-on (plan required)	Teams already on Semrush who do not want another vendor
Ahrefs Brand Radar	Built on a corpus of real user prompts	From ~US$199 per engine	Anyone who wants real questions, not synthetic test prompts
DIY (GA4 + server logs)	Referral and crawler monitoring only; no citation rate	One-off setup cost	Businesses under ~100k monthly sessions

Prices are approximate 2026 public tiers; check the vendors directly. The shared limitation matters more than the price gaps: of these six options, only the last one can tell you whether the people arriving from AI actually bought something. The other five mostly stop at whether you were mentioned.

What the Report Does Not Tell You

First, the division nobody performed. Organic lost roughly 5 million sessions a month. AI referrals, after 163% growth, run about 200,000 a month. Even counting all 200,000 as incremental — the implied base was around 76,000, so net gain is closer to 120,000 — the recovery is 4% of what was lost. Apply the 1.5x key-event rate generously and call it 300,000 organic-equivalent sessions, and you reach 6%. For every 100 sessions lost, AI gives back the effect of about six. "Less traffic but better quality" is not reassurance at that ratio. It is a 94% hole described as a channel opportunity.

Second, 1.5x is survivorship bias, not channel superiority. The person who clicks out of an AI answer has already read the summary and wants to verify one more thing. That is a filtered, late-funnel cohort. Comparing them to all organic traffic — most of which is informational — is comparing finalists to the entry list. The fair denominator would be transactional-intent organic traffic, and that figure is not in the report.

Third, the key-event definition inflates itself. Brainlabs states plainly that key events include making a purchase, signing up to a newsletter, "or just scrolling to the bottom of a web page." AI referrals typically land on a single content page, and readers who finish it scroll to the footer — which in most GA4 configurations fires a key event immediately. Organic ecommerce traffic lands on category pages, filter pages and carts, where scroll-to-bottom rates are naturally lower. The 1.5x may be measuring page-type differences rather than purchase intent.

Fourth, there is no control group. Fourteen months contained more than AI Overviews: multiple core and spam updates, SERP layout changes, consent-mode shifts and GA4 data modelling adjustments all move session counts. Attributing the full −10.5% to AI Overviews means picking a month and calling it a hinge, not running a structural break test. And eight of the 54 clients did not decline at all, which the report leaves unexplained.

Reproducing This Without a Subscription

Brainlabs used tools you already have: GA4, Search Console, server logs. The only difference is that someone sat down and configured them once.

☐ Create an AI Referral channel group. GA4 Admin → Data display → Channel groups. Match source against chatgpt|perplexity|copilot|gemini|claude|you\.com. It applies going forward only, so build it now.
☐ Redefine your key events. Strip scroll and engaged-session metrics from the list. Keep four real commercial actions: enquiry form submitted, checkout completed, phone call placed, messaging channel joined. Remove the inflated numerator and the number becomes usable in a negotiation.
☐ Isolate a transactional baseline. In Search Console, filter queries containing price, best, versus, where to buy and cost, and export those clicks. That is the correct denominator for AI referrals.
☐ Monitor crawler logs. Run grep -E "GPTBot|ChatGPT-User|PerplexityBot|ClaudeBot" over your access log weekly and tally hits by URL. The pages crawled most often are your body of work as far as the models are concerned.
☐ Sample citation rate by hand. Write 20 prompts you most want to be recommended in and run them monthly against ChatGPT and Perplexity. Forty queries, about an hour — substantially what the monitoring subscriptions sell.
☐ Fill in structured data. FAQPage, Product, Offer and LocalBusiness first. One-off cost, continuing benefit.

To turn this into a monthly report a client can actually read, or to assess how far your sector has been absorbed by AI answers, get in touch:

FAQ

My traffic dropped. How do I tell whether AI Overviews caused it?

Look for the intersection of two signals. In Search Console, split queries into informational and transactional: if impressions hold steady while clicks fall, you are still ranking and simply not being clicked — that is the classic AI Overviews fingerprint. If impressions fall too, it is a ranking problem and AI is not the culprit. Then cross-check against Google's confirmed update timeline. If the dates do not line up, do not blame AI.

How do I separate AI traffic inside GA4?

It will not do it by default. ChatGPT usually lands in referral and sometimes in direct. Create a custom channel group in Admin and match on source domain. Two caveats: channel groups never backfill, so your baseline starts the day you build it; and a meaningful share of AI traffic gets logged as direct because users open a separate tab. There is no clean fix for that today — the honest move is to state the margin of error.

Does "AI traffic converts 1.5x better" hold for a small business?

Almost certainly not. Twenty-nine of the 54 sampled companies employ 1,000+ people. These are firms with an established brand entity, and models recommend them because the training corpus is full of third-party writing about them. A local company with no encyclopedia entry, no press coverage and no forum threads is not in the same order of magnitude for citation probability. Solve "is anyone writing about you" before you optimise conversion rate.

So should I buy an AI visibility tool or not?

Use one threshold: whether you exceed roughly 100,000 organic sessions a month. Below that, the sample a tool measures is too small — monthly movement is mostly noise, not trend, and a manual 20-prompt check is enough. Above it, and with someone on staff who will change content based on the report, the subscription starts paying. A dashboard nobody opens is the most common waste in this category.

My Take

The consensus reading is: traffic will fall, but the people AI sends are worth more, so add GEO to your KPIs now. I think the opposite. By mid-2027, "AI traffic converts better" will be reclassified as a measurement artefact rather than a channel advantage.

All three supports are loosening. Key-event definitions are being tightened across the industry, and scrolling to a footer will not count as conversion forever. As AI assistants take on transactional queries, the people who click through shift from "the decided few" back to "the still-comparing many," and 1.5x drifts toward 1.0 on its own. And OpenAI is already testing ads and shopping placements — once position inside an AI answer can be bought, the high-intent traffic that currently arrives via organic citation gets bought out from under you first.

The less flattering corollary: the GEO tooling business model assumes visibility is a continuously optimisable variable. If citation is mostly determined by brand entity strength, and brand entity strength is determined by PR, media and word of mouth, then these tools are selling a dashboard for a public relations problem — measurable, but not movable.

The implication for ScriptWalker is concrete. Do not compete for "AI visibility monitoring," a market already crowded with funded companies. Compete for the gap they leave behind. A tool tells the client they appear in 12% of relevant prompts, and then stops. The billable work is the next step: rebuilding the pages that do get cited into pages that close, converting FAQs and specifications into structured data, and repairing GA4 key events from inflated to credible. That is a one-off project plus a monthly retainer, it sits well inside our technical range, and it does not require out-engineering a company with a billion-dollar valuation.

Sources

Primary

Third party