Welcome to the 2nd edition of the AI SEO Newsletter.
Every week, we curate the latest AI SEO news, studies, tactics, and tools that can shape your strategy.
Our Latest Insights

How Google's AI Overviews pick their sources
We looked at 857 AI Overview citations in B2B SaaS. Here's what we found:
-
Only 33.2% of cited pages rank in Google's top 10
-
YouTube is cited 5x more often than Reddit
-
Being on-topic matters more than site authority
Search & AI Updates
- Google tests AI-written summaries in the citation cards next to AI Overviews
Brodie Clark on X
Instead of quoting the source page, some citation cards now show a summary Google writes from what it thinks matters most for the query. The snippet beside your link may no longer be your own words. Brodie Clark spotted the test, which shows for some queries but not others.
- Sundar Pichai says Search is becoming part of a single agentic layer across Google
Sources with Alex Heath
In an interview with Alex Heath, Google's CEO said Google plans to bring agentic capabilities into Search “in a pretty deep way”, and expects AI agents to reach websites both by browsing pages and through APIs.
- Google: robots.txt controls crawling, not how AI uses your content
Google Search Central
In a Search Central video, Gary Illyes said robots.txt only governs which URLs crawlers can request. To keep content out of Gemini training and grounding, you need to block the Google-Extended token. A standard for usage preferences is still only an IETF draft.
- OpenAI explains how ChatGPT picks which ads to show
OpenAI
An orchestrator model follows the conversation across messages and nominates relevant ads, then a separate auction ranks, prices and selects them. The language model decides which ads enter the auction, but not which one wins.
- Google moves links higher in AI Mode and AI Overviews and widens news carousels
Google
Google says use of Preferred Sources has grown more than 7x since the start of the year. It's also testing a slot at the top of answers for one or two official pages, plus clickable creator and publisher names on video results. Search Console still shows no click data for AI results.
- ChatGPT's sudden referral traffic drop looks like a tagging bug, not lost visits
Profound
After GPT-6 launched, the share of ChatGPT answers with a UTM-tagged link fell from about 98% on Oct. 8 to 66% on Oct. 14, so GA4 counted those visits as Direct. One analyst reports a fix rolling out from Oct. 15; OpenAI hasn't commented publicly. Annotate Oct. 9 in your reports and don't compare the weeks either side of it.
- Cloudflare signals helped OpenAI find hundreds of millions of new pages for ChatGPT search
Cloudflare
The pilot covered about 1% of Cloudflare's network, sending OpenAI's search crawler signals about which pages exist and when they change, so it could find useful pages that few other sites link to. Check that robots.txt allows OAI-SearchBot, which OpenAI controls separately from GPTBot.
- DeepSeek opens its own search engine to developers and says it filters "GEO manipulation"
DeepSeek
The Search and Fetch APIs cost $0.75 per 1,000 requests and draw on an index of hundreds of billions of pages. A site: query shows which of your pages DeepSeek has indexed, with an authority tier from 0 to 4. It doesn't say what counts as GEO manipulation.
- Gemini 4 Argon comes to AI Mode for Google AI Pro and Ultra subscribers
Google
Google's new frontier model is built for deep, multi-step work in areas like coding, finance and legal. Pro and Ultra subscribers can use it in AI Mode and the Gemini app, and developers can use it through the API in AI Studio.
- Bing Webmaster Tools launches Topics Portfolio (Beta) in its AI Performance report
Bing Webmaster Blog
It groups the grounding queries that cite your site into topics, showing each topic’s size, how many sites are cited, and Top Opportunity Queries where competitors appear but you don’t. It only appears for sites with enough data, and Microsoft says the numbers are approximate and meant for planning.
Worth Reading
- ChatGPT sent 0.62% as many sessions as Google organic across 100 sites
Marie Haynes, Search Engine Journal
Marie Haynes analyzed 12 months of GA4 data (145M+ sessions). ChatGPT visitors converted at 13.8% vs 3.02% for Google organic, but Google still drove 96.5% of conversions. Her conclusion: prepare for agentic search rather than chasing LLM citations.
- Why citations aren't the whole AI visibility game
Graphite
Graphite argues that what a model already knows from training shapes both what it searches for and what it picks. Of about 957,000 ChatGPT fan-out queries from non-branded comparison prompts, 59.6% already contained a brand name, and brands the model didn't know reached only 11% median visibility even when mentioned on 15%+ of cited pages.
- Does AI drive more direct traffic? Kevin Indig tested it and couldn't find it
Kevin Indig
Across 12 mostly B2B sites, Direct traffic rose 140.9% from January to September while engaged Direct sessions moved -1.0%. In 632 tests across four brands, AI visibility didn't reliably predict Direct or branded search, so he recommends triangulating AI answer presence, self-reported discovery and revenue.
- Listicles make up 66.6% of the pages AI engines cite
OtterlyAI
OtterlyAI logged 617,985 citations from seven AI engines answering buyer questions in five US industries. Publishers who disclose affiliate commissions took 36.6% of citations versus 11.7% for editorial publishers, and Claude cited listicles 78.2% of the time.
- How a site of made-up statistics ended up all over AI answers
Metehan Yesilyurt
Metehan Yesilyurt generated about 50,000 unchecked statistics pages for roughly $50. Google buried the site, but it drew 72,000 ChatGPT-User requests in 30 days and links from 330+ sites, and its invented figures turned up in an academic paper. LLMs rerank for relevance, not truth.
- AI Mode queries are leaking into Search Console: what 155 sites show
GSC Wizard
Jan-Willem Bobbink classified 16 months of queries across 155 properties and found AI Mode-style reply queries grew about 21x, showing up on 94 of the sites. Those queries click through at just 2.9%, against 35.6% for all queries.
- How do we measure SEO if the SERP is generative?
iPullRank
Mike King argues that as Google moves toward model-generated results pages, there's no single stable SERP left to rank on. Treat each capture as one sample: take several a day, report visibility as a rate with a confidence interval, and anchor rank tracking to Search Console and referral data.
- Eight of 10 AI search tools cited a made-up concept within a week of it being posted
arXiv (Fudan University)
Fudan University researchers planted articles about a fictional concept on Medium, YouTube and other sites AI tools cite often. ChatGPT, Perplexity, Grok and Google AI all cited them within seven days; Grok linked to the planted posts in 30 of 32 answers.
- OpenAI’s API Only Partly Matched ChatGPT’s Citations, Paper Finds
arXiv
Researchers traced how ChatGPT rewrites queries, retrieves pages and picks citations from GPT-5.2 to GPT-5.5. Google Search and OpenAI's Responses API, even combined, recovered only about a quarter of the URLs ChatGPT cited, but matched much better at the site level. Track visibility by site rather than exact URL, and rerun prompts, since results vary run to run.
- Poisoned Google SERPs are skewing rank tracking data
Nozzle
Nozzle analyzed 27 million Google results pages across 52 countries. Since September 13, 8.4% came back "poisoned", with most of a keyword's usual top-10 sites missing, against 0.17% before, peaking at 28.6% on September 29. A tracked site missed the top 10 on 60.2% of poisoned pages, but 68.9% recovered within three days, so check before reacting to a sudden drop.
Tools Worth Testing
-
Query Fan-Out Extractor
Our free tool that shows the hidden searches an AI model runs before it answers, across Google AI Mode, ChatGPT and Perplexity. Those sub-queries are the subtopics your content needs to cover.
Try It → -
Prompt Tracking Frameworks skill
A free text file from Orit Mutznik that you attach to ChatGPT, Claude, Gemini, Copilot or Perplexity. It asks about your brand, buyers and competitors, recommends one of three prompt tracking frameworks, and writes 25 to 50 prompts to track. Worth a try before you pay a tracker to monitor prompts picked by gut feel.
Try It → -
Jev by TypeSafe AI
A new kind of AI model that returns decisions instead of text: choices, scores and probabilities your scripts can use directly. Worth testing for jobs like classifying keyword intent or mapping redirects in bulk. In early access, and you'll need a little code to use it.
Try It →
Have any questions or suggestions? Email us at hello@position.digital.
See you next week.
Cheers,
The Position Digital team