AI Traffic Stays Small as AI Answers Reach BuyersThe Visibility Report #40 | Week of October 20-26, 2026 AI assistants still send very little identifiable referral traffic compared with Google. Marie Haynes found that ChatGPT sessions equaled 0.62% of Google organic volume across 100 sites and more than 145 million sessions. The same week, OpenAI explained how ChatGPT ads use a separate language model to read multi-turn intent, retrieve candidates, and run an auction. Those findings put the measurement problem in the right order. Keep spending proportional to what you can prove. Also watch the answer, recommendation, and ad surfaces that can influence a buyer before a trackable visit appears. Marie Haynes found that ChatGPT sent 0.62% of Google's organic trafficHaynes analyzed 12 months of GA4 data from 100 sites with more than 145 million sessions. In that sample, ChatGPT sessions were 0.62% of Google organic sessions. All measured LLM referrals combined reached 0.73%. Among the 55 sites that tracked conversions, ChatGPT visitors converted at 13.8%, compared with 3.02% for Google organic visitors. The higher rate did not overcome the volume difference. Google organic produced 96.5% of the conversions across the channels in her analysis and 27 times as many conversions as all LLMs combined. This is an aggregated, self-selected set of sites, not a market census. GA4 also misses untagged referrals and buyers who see an AI answer before returning through search, direct traffic, or another channel. The findings still make one budget question unavoidable: how much are you spending to measure a channel that sends less traffic than several smaller search engines? SparkToro reviewed Haynes' work beside two other recent studies and argued against reorganizing marketing budgets around AI traffic. Mike King at iPullRank accepts the low-referral premise but frames AI answers as a branding channel. Both positions fit one operating rule: measure identifiable sessions and conversions first, then label mention share, recommendation context, and branded demand as directional influence rather than attributed revenue. Sources: Marie Haynes | SparkToro | iPullRank OpenAI separated ChatGPT's ad auction from answer generationOpenAI says ChatGPT uses a dedicated language-model orchestrator for ads. It reads the conversation, forms several possible interpretations of the user's intent, retrieves candidate ads, applies eligibility and relevance checks, and sends the remaining candidates to an auction. A final check can block the winning ad if it does not fit the conversation. The ad service runs separately from the model that writes the answer. OpenAI says the answer model does not see the ads or wait for them, and a late ad can miss the response even after it wins. That separation is OpenAI's description of its architecture, not an independent audit of the boundary. OpenAI also reported a $1 billion annualized revenue run rate in less than 200 days and more than 100,000 advertisers since launch. It did not publish advertiser performance benchmarks. Bids matter after the system narrows the inventory for relevance, so advertisers need to map creative and landing pages to the full buying conversation rather than a keyword alone. Keep organic mentions and sponsored delivery in separate reports. Source: Brooke Osmundson in Search Engine Journal Sundar Pichai said Google agents will use web pages and protocolsGoogle CEO Sundar Pichai told the Sources podcast that agents will reach sites in two ways: by browsing pages much as people do and through protocols built for agents. He expects protocols to scale better, but he also expects them to take longer to develop. Google already documents Google-Agent as the user agent for hosted browsing and backs the Universal Commerce Protocol for shopping tasks. This is a product direction, not a deadline for replacing websites with APIs. Brands need both paths to work: accessible pages that an agent can read now, and structured endpoints or commerce protocols when the business case is clear. Test the unauthenticated path through important tasks. Check whether an agent can find prices, availability, policies, and contact steps without executing a fragile interface. Then track protocol work separately so a future integration project does not become an excuse to neglect the site customers use today. Source: Matt G. Southern in Search Engine Journal Google tested AI-written descriptions inside AI Overview citation cardsSearch Engine Roundtable documented a Google test in which citation cards used an AI-generated description instead of text copied directly from the cited page. Brodie Clark first shared the example, and Barry Schwartz reproduced the behavior on some queries but not others. The limited observation does not establish a broad rollout. It does show another place where a publisher's words can be summarized before the click. Check how high-value pages are described in citation cards, save the query and date, and compare the summary with the source. If the description changes the meaning, fix ambiguous source language before assuming Google invented the problem. Sources: Search Engine Roundtable | source post on X Foundation found that Google's AI cited YouTube far more often than ChatGPTFoundation's B2B Citation Index ran the same 5,000 bottom-of-funnel software prompts across 80 categories for 13 weekly runs from June 29 through September 21. The US tests, run with search enabled, produced 182,631 answers and 2,227,822 citations. Of those citations, 83,835 pointed to YouTube. Google AI Mode and AI Overviews produced 98.2% of the YouTube citations in the study. Google's two AI surfaces cited YouTube in 36% of answers, compared with 2.1% for ChatGPT. Among the 150 most-cited videos, view and subscriber counts had almost no relationship with citation frequency. More than half of the 17,780 cited videos appeared in only one answer. The sample covers bottom-of-funnel B2B software prompts, not all search behavior. It was also powered by Writesonic, a vendor relationship Foundation discloses. Treat 36% and 2.1% as engine-specific findings inside that panel. If a buyer needs a demonstration, comparison, or review, test a clear video against the same prompts over several weeks before expanding production. Source: Foundation's YouTube citation study Semrush found AI Overviews on 88% of its commercial query panelSemrush tracked a fixed panel of 400 US desktop queries across banking, ecommerce, and software from September 2025 through August 2026. The panel covered five commercial modifiers: alternatives, brand comparisons, best-for queries, price or cost, and reviews. In August 2026, 88% of the panel triggered an AI Overview, up from 45.75% in September 2025. Review queries rose to 90%. Brand-comparison queries exceeded 86%, while alternatives exceeded 82%. Across all five modifier groups, coverage landed between 82% and 96%. This is a 400-query vendor study, not proof that 88% of every commercial search shows an AI Overview. It does move the monitoring priority closer to revenue. Track the AI answer on pricing, reviews, alternatives, and head-to-head comparisons. Record which brands appear, how they are described, and which sources support the answer. Source: Semrush commercial-modifier study From Practitioners
From the Tool Blogs
What to watch
What to do next
AI referrals are small enough to demand spending discipline. AI answers are close enough to buying decisions that ignoring them would also be a guess. Measure the traffic you have, sample the answers buyers see, and keep directional influence out of the attributed-revenue column. The Visibility Report | Will Scott |
Join 500+ digital marketers getting weekly AI visibility tools, tactics, and search updates. Free • 5-min read • No spam.
The Visibility ReportIssue #43 | Week of November 9 - November 15, 2026 Google Search Is Querying More Than Your Query Google's newest Search model can split one complicated question into several related searches, then build an answer from what it finds. That makes the pages behind the answer harder to predict, but it does not make the work mysterious. Subscribe to The Visibility Report for a source-checked briefing every Tuesday. Google expands query fan-out with Gemini 4 Hydrogen Google...
The Visibility ReportIssue #42 | Week of November 2 - November 8, 2026 ChatGPT Webmaster Tools May Be Coming OpenAI appears to be testing something search marketers have wanted for years: a first-party view of how businesses appear in ChatGPT. Think of it as ChatGPT Webmaster Tools, with one important caveat. This is an unconfirmed, unfinished feature seen by a third party, not an OpenAI product announcement. Subscribe to The Visibility Report for a source-checked briefing every Tuesday....
The Visibility ReportIssue #41 | Week of October 26 - November 1, 2026 Publishers Get a Clearer Choice About AI Training Publishers have been stuck with a blunt choice: let AI companies use their content or risk disappearing from AI answers. New controls from Google, OpenAI, and Microsoft make that choice a little less blunt. Microsoft draws the clearest line between model training and grounding, which is the use of current web content in generated answers. A publisher can say no to training...