Search Signal #12
62% of B2B buyers picked a vendor named in the AI's first answer. Plus: ChatGPT now reports your clicks, a core update, and Gemini 4 Hydrogen.

Vikas Kalwani
Founder, SAASY LINKS
01. OpenAI shipped a Search Console for ChatGPT, and a mention only counts if your site is cited
OpenAI opened Businesses in ChatGPT and its Business Hub in beta on 10 November. Claimed businesses see mentions, profile views, website clicks and five discovery topics, on 7-day and 28-day views. Search Engine Journal's own property logged 353,356 mentions and 2,663 website clicks from 14 October to 10 November. The catch: for online businesses a mention counts only when ChatGPT cites your website, not your brand name. Eligibility at launch is US ecommerce and online service businesses.
Why it matters: Every AI visibility vendor charging for ChatGPT share of voice has been selling a sampled estimate, and from this week that estimate is checkable against OpenAI's own counter.
Action: Claim each client domain in the Business Hub this week with a DNS TXT record, and do not report clicks as sessions: OpenAI's own tooltip says clicks do not confirm website visits or conversions.
02. 62% of B2B software buyers who chose a vendor chose one named in the first AI answer
Kevin Indig published a Wynter survey of 200 software buyers on 9 November. Of the 126 who picked a vendor, 78 (62%) picked one named in the first AI answer, against 43 (34%) who picked one that surfaced only after follow-ups. First-answer vendors converted to a demo or signup 41.0% of the time against 34.9%. Claude was the most used tool at 71%, ahead of ChatGPT at 62.5% and Gemini at 38.5%.
Why it matters: Self-reported recall from 200 people is weak evidence, but it is the only B2B-specific number published this week, and it says the first answer is the shortlist.
Action: Run your five core category prompts in Claude, not only ChatGPT, record whether the client is named in the first answer before any follow-up, and date the screenshot.
03. A core update, a Search Console AI logging bug and a model swap all landed inside four days
Google released the November 2026 core update on 10 November at 10:06 PST, global and across all languages, with a stated rollout of up to two weeks, which puts completion near 24 November. Separately, Google's data anomalies page confirms a logging error that cut impressions in the Generative AI performance report from 9 to 11 November, affecting data logging only. Gemini 4 Hydrogen then became the AI Mode default on 12 November.
Why it matters: Three of the four numbers a November report rests on moved for unrelated reasons, so any single-cause story about a November delta is wrong by construction.
Action: Annotate 10 and 12 November in every client dashboard, exclude 9 to 11 November from AI impression baselines, and follow Google's own guidance by waiting a week after the rollout completes before comparing Search Console periods.
04. Gemini 4 Hydrogen became the AI Mode default worldwide and now answers the hardest AI Overviews questions
Robby Stein, VP of Product for Search, said on 12 November that Gemini 4 Hydrogen is the AI Mode default for all users globally, free accounts included, and is also being used for the hardest questions in AI Overviews. Google says Hydrogen's reasoning supercharges its query fan-out technique and connects users to a broader range of links. It published no link, click or traffic figures to support that. Gemini 4 Carbon reached AI Ultra subscribers on 10 November.
Why it matters: Query fan-out decides which of your pages get pulled into an answer, so a default-model change is the most likely cause of citation-set churn this month.
Action: Re-run your tracked AI Mode prompts after 12 November and diff the cited URLs against your October set. If the pages changed and your content did not, the model moved, not your site.
05. Position-one clicks fell from about 7 per 100 impressions to about 5, across 151 billion impressions
iPullRank analysed 166 brands and 272 verified Search Console properties on 12 November, covering March 2025 to mid-July 2026 and roughly 151 billion impressions. Unbranded position one fell from about 7 clicks per 100 impressions in 2025 to about 5 in 2026. Position three more than halved, from about 3.5 to under 2. Positions four and below barely moved. Moving from four to three is worth about 19 clicks per 1,000 impressions; eight to seven is worth about one.
Why it matters: The study never measures AI Overviews and does not claim to, so quote the decline and not the cause. The flat tail is the real finding: rank gains below position four buy almost nothing.
Action: Reprice the ranking roadmap. Pull effort off keywords sitting at 8 to 15 and put it behind pages already at 4 or 5, where one position is worth about 19 clicks per 1,000 impressions.
06. ChatGPT cited 88% of the pages it fetched by web search, and 83% of brand answers cite the brand's own site
Ahrefs published a composite of its Brand Radar datasets on 10 November. Of the pages ChatGPT retrieved through web search, 88% were cited, against 12% of news pages and under 2% from Reddit, YouTube and academic feeds. Cited titles matched the query at 0.60 similarity against 0.48 for uncited. For brand questions, 83% of answers cited the brand's own site and 70% cited it first. A separate test on 1,885 already-cited pages found no meaningful citation change from schema.
Why it matters: This is vendor research from a company selling citation tracking, and it still carries the week's most useful mechanic: retrieval is the bottleneck, and your own site is the default source for questions about you.
Action: Audit the page a branded question should land on, usually the product or comparison page, for a title that matches the question wording, and stop selling schema as an AI citation lever on the strength of this null result.
07. One site opted out of Google's AI features for 28 days and estimated they were sending 21% of its clicks
Cyrus Shepard of Zyppy SEO excluded his own site from Google's AI features in Search Console on 1 September, opted back in on 29 September, then compared matched queries and positions across the two 28-day periods. AI features were about 10% of impressions. He estimates AI Overview and AI Mode click-through at about 0.9%, with a range of 0.7% to 1.1%, against the site's roughly 0.4% sitewide average, and about 21% of total search clicks.
Why it matters: A single site the author owns is the weakest possible sample, and the arithmetic is still the useful part: a 0.9% click rate across a tenth of impressions is not a rounding error.
Action: Do not run the opt-out test on a client site. Instead compare the Generative AI performance report against total Search clicks for the same period, and tell the client what share of their impression base AI features now hold.
08. A branded prompt returned a competitor's sponsored card in five of six verticals tested
Seer Interactive reviewed 161 recorded AI search sessions from March to September 2026 and published on 12 November. In five of six verticals, a branded prompt returned a sponsored card for a different company. ChatGPT in-answer ad exposure on free accounts rose from 2 of 22 users in March to 12 of 19 in the July to September window. Seven of the twelve participants shown an ad said there were none. Every Plus and Pro session was ad free.
Why it matters: Competitors can now buy the moment someone types your client's name into ChatGPT, and the users in this sample did not register the result as advertising.
Action: Type each client's brand name into a logged-out free ChatGPT account this week, screenshot whatever sponsored card appears, and put it in front of whoever owns their paid budget.
09. A federal judge set a 25 November deadline that forces the retrieval question into summary judgment
On 12 November, Judge Sidney Stein ordered that if the New York Times and the other news plaintiffs in the OpenAI copyright MDL want to pursue their grounding and output theories separately from their training claims, they must move by 25 November, with opposition by 11 December and a reply by 18 December. Word limits are 7,500 on the opening and opposition briefs and 2,500 on any reply. Nothing has been decided on the merits.
Why it matters: Grounding is the retrieval layer, the part of AI search that touches your clients' live pages, and a US court could rule on it in the first quarter of 2027.
Action: No action on the case itself. Do start logging, per client, which engines fetch their pages live and under which user agents, because that log is what any future access or licensing conversation gets argued from.
10. A court threw out Google's scraping complaint against SerpApi, and Brave began encrypting its result links
On 10 November a judge dismissed Google's amended DMCA complaint against SerpApi, giving Google until 30 November to amend again or see the case dismissed with prejudice, with discovery stayed. On 11 November Brave said it will encrypt result links and decrypt one URL per request, after scrapers pivoted to Brave and drove spikes of several thousand requests a second. SerpApi's goto incident reopened on 12 November and was still open on 14 November.
Why it matters: The legal case for buying SERP data got better in the same week the technical supply got worse, which means rank data keeps degrading whatever the courts decide.
Action: Check your rank provider's status page before pulling November data, and reconcile any reported decline below position nine against Search Console average position for the same queries.
11. Yahoo became the first paying licensee of a blanket AI licence covering more than 3,000 publishers
The Internet Licensing Collective, formerly the RSL Collective, launched on 12 November with more than 3,000 media brands and content owners, and named Yahoo's Scout as its first paying licensee. One licence covers training, retrieval-augmented generation, grounding, agentic applications and AI-generated overviews. Publishers join free, keep the right to do direct deals, and are paid on contribution rather than catalogue size. No rates were disclosed. The licence is expressed through Really Simple Licensing, a machine-readable standard.
Why it matters: Collective licensing is the first route by which a site too small for a bilateral OpenAI deal could be paid for retrieval, and it runs through a file you control.
Action: Read the Really Simple Licensing spec before a client asks about it. It is the first licence-terms format that sits alongside robots.txt, and whether to publish one is a client decision, not an SEO one.
12. Agent traffic to SaaS sites grew about fifteenfold since August, and Meta's Muse is now 59% of it
HUMAN Security's October agentic traffic report, published 9 November, puts total agent requests at 2.6 times September and roughly eight times August across its protected properties. Meta's Muse is 59% of agent requests, 3.7 times ChatGPT Agent's volume. ChatGPT Agent fell to 15.9% from 28.5%, and Perplexity's Comet to 3.6% from 41% in August. SaaS is only 1.24% of agent requests, but nearly four times September and about fifteen times its August level.
Why it matters: This is vendor data from one network, so treat the shares as directional, but the SaaS growth rate and the finding that 85% of agent requests land on product and search pages point the same way.
Action: Grep last month's access logs for the Muse user agent, then check whether your WAF or bot-management rules challenge it, since most edge configurations were written against a 2025 agent list.
Also worth knowing
- Traffic to the open web keeps shrinking. SparkToro and Similarweb put visits to the top million sites at 1.970 trillion in Q3 2026, down 12.2% from the Q1 2022 peak, with visits per web user down about 29%.
- Content rewrites may not move an unfamiliar brand. An audit of 1,176 market leaders posted to arXiv on 5 November found top-quintile brands cited 1.4 to 2.8 times more often than average, that web retrieval did not mitigate it, and that fact-preserving rewrites produced negligible ranking gains.
- Semrush stopped tracking Microsoft Copilot on 9 November because Copilot now requires sign-in, which puts a composition break in any share-of-voice chart spanning that date.
- Google warned that moving to a CDN can get Googlebot blocked. Gary Illyes and Chery Prommawin explained that a cold cache sends every uncached URL back to origin, 100,000 of them in Google's example, and that teams often read the spike as an attack and switch on protection that blocks verified crawlers.
- John Mueller said a high share of crawled or discovered currently not indexed pages is a quality signal, not a crawl budget problem, and that making a site smaller does not by itself raise its quality.
- The EU's proposed Digital Fairness Act would force AI agents to disclose paid placement. The Commission presented the 72-page regulation on 11 November, and Article 26(4)(b) requires disclosure plus user confirmation whenever an agent's provider is paid for proposing a contract.
SAASY LINKS
We build the visibility this newsletter reports on
Editorial links, AI citations, and organic pipeline for B2B SaaS. Senior-led, 90-day sprints, no long-term contracts.
Book a Call