The week in GEO: 16 to 22 October 2026
Sixteen stories in our news feed this week. These are the ones that change how a GEO agency works, with what each means in practice.
ChatGPT citations cannot be reproduced by API or Google results
A preprint from researchers at the National University of Singapore and Xi'an Jiaotong University found that OpenAI's Responses API and Google's organic results together surfaced 26.2% of the pages cited by GPT-5.5 Instant and 48.7% of the sites behind those citations. Google results alone accounted for 6.4% of cited pages and 22.8% of cited sites.
Results varied between runs: 44.5% of the distinct pages returned for a given GPT-5.5 Instant question appeared in only one of its 20 runs. ChatGPT rewrote queries rather than searching the original wording, and for Chinese questions it searched in English alone in 69.6% of visible GPT-5.5 Instant runs. Every model tested predates GPT-6. (summary, more)
What it means for agencies
Ask any visibility tool whether its figures come from the API or the ChatGPT app, and treat API-only numbers as a lower bound. Track at page and site level, and repeat each prompt enough times to see the spread. A single run tells a client very little.
AI Overviews cover 82% to 96% of commercial searches
A year-long study of 400 keywords in banking, ecommerce and software found AI Overviews on 82% to 96% of every commercial modifier tracked, up from 30% to 63% in September 2025. Searches containing "reviews" rose from 30.14% to just under 90%. "Best X for Y" queries ended at 96.12%, and "brand vs brand" grew least, up 23.18 points to 82.05%. Weighting by search volume puts coverage at 97.08% in August 2026.
Google has also begun removing the AI Overviews label from some desktop results while keeping it on mobile, which one commentator attributed to a possible bug. (summary, more)
What it means for agencies
Commercial-intent queries now almost always carry an AI answer, so the organic first position is no longer the first impression. Audit what the overview says for each client's money keywords, including which review sites and videos it draws on. Where the label is missing on desktop, screenshots and label-dependent tools will undercount appearances.
A false AI answer was corrected within hours by publishing and indexing
Feature Upvote founder Steve McLeod found a Google AI Overview in July 2026 stating his company had no REST API, a week after launching one. The answer cited old blog posts over newer help documentation. He published a blog post, an X post and a newsletter on 24 July, then found Google had not indexed his announcement post at all. Within an hour of submitting his pages on 28 July, the overview answered "Yes" to whether the company has an API.
The answer flipped depending on phrasing for several days before settling. (summary)
What it means for agencies
Add a check for false brand facts in AI answers to every client audit, and check indexing status before assuming a fix failed. Publish a dedicated page stating the fact plainly, correct the outdated pages being cited, then submit for indexing. Test several phrasings before telling a client the answer has changed.
China publishes a voluntary GEO standard with measurement rules
China's advertising industry has published a five-part voluntary standard for generative engine optimization, six months after a CCTV broadcast showed an invented smart wristband being recommended by at least two AI models after roughly two dozen seeded articles. The series was issued by the China Association of Commercial Advertising and is marked T/, meaning voluntary rather than law.
The measurement rules are the most portable. Because generative engines rarely return the same answer twice, the standard sets a floor of 30 runs at the finest level a vendor reports on, and warns that scores built on five runs can swing 20 points with no real change. It requires visibility to be calculated only on category questions, never branded ones, and bans promises such as "100% screen dominance". (summary, more)
What it means for agencies
Adopt the 30-run floor in your own reporting and ask suppliers how many runs sit behind each figure. Separate branded from category prompts in visibility scores. Drop any pitch language promising control of model output; it now has a published industry standard against it.
ChatGPT referrals stay under 1% of Google organic traffic
An analysis by Marie Haynes covering 100 sites, 12 months of GA4 data and more than 145 million sessions found ChatGPT referral traffic amounts to 0.62% of the sessions Google organic search sends to websites. Adding other large language models raises the figure to 0.73% of Google organic traffic.
ChatGPT visitors converted at 13.8% across the 55 sites tracking conversions, against 3.02% for Google organic visitors, but Google organic still produced 27 times as many conversions as all LLMs combined. Haynes said ChatGPT referral traffic stopped growing exponentially in August 2025 and declined afterwards. (summary)
What it means for agencies
Keep GEO spend proportionate to this scale when clients ask what it returns. Report AI referrals as a share of Google organic, not of total traffic, and pair session counts with conversion rates. The higher conversion rate is real but small in absolute terms, so do not present it as replacing search.
ChatGPT Ads adds automation and measurement partners
OpenAI has added visual ad formats, a redesigned conversions page, weekly performance emails and measurement partners including Northbeam, DV Rockerbox and INCRMNTAL, plus Kochava, Tenjin and Airbridge. Creative enhancements will switch on automatically for new and existing self-serve campaigns from 30 October 2026, rewriting headlines and descriptions from ad copy and landing pages, resizing images and translating text, with opt-outs available in Ads Manager.
Advertisers on ChatGPT climbed from around 20,000 at the end of August to roughly 60,000 in early October, with clicks costing about $5 against $30 to $50 on comparable Google ads. OpenAI says the ad system runs on separate models from the answers. (summary, more, more)
What it means for agencies
Check the creative enhancement setting on every client account before 30 October and decide deliberately rather than by default. Landing page copy may now become ad copy, so review pages for claims you would not run as ads. Use the new measurement partners to test whether paid placement shifts citations before scaling budget.
Google puts ads in AI Mode and source preferences beside citations
Google has confirmed ads will appear in AI Mode results for users with Personal Intelligence switched on, reversing its earlier ad-free position. Its documentation says ads will not be based directly on Gmail or Google Photos, though summaries and inferences from connected apps may personalise results and ads.
Google is also testing a Source preferences button inside the AI Mode citations sidebar, letting users add or remove preferred sources without leaving the answer. Preferred sources have been supported since May. (summary, more)
What it means for agencies
Paid and earned placements now sit in the same AI answer, so show clients prompt by prompt where they are cited and where a paid slot would add anything. Being on a user's preferred sources list is now one click from the answer, which strengthens the case for work that earns repeat selection rather than one-off mentions.
Also this week
Cloudflare and OpenAI say a pilot across roughly 1% of Cloudflare's network helped OpenAI's crawler find hundreds of millions of pages it had not indexed, with crawler access the lever for site owners. (summary) A commentary argues AI assistants let people describe needs in more detail, so content should cover the conditions underneath a category rather than the category alone. (summary)
All the week's stories are on the news page. To compare GEO agencies, see the directory.