Last Week in AI Search: November 16
What this answers
- How many items did the AI-search wire publish this week?
- 19 items between 2026-11-09 and 2026-11-15, across the lanes Engines, Research, AEO. Each is dated and sourced on the wire.
In this issue: Too little context and AI guesses. Too much and it ignores you. Four ways to tilt the odds · Plus: OpenAI's Business Hub and 79% of sites now get AI…
The Story
AI results are a roll of the dice. 4 ways to tilt the odds your way
AI is a gamble. When Profound had the same made-up person ask ChatGPT, Gemini or Claude the same question twice, only about 2 in 5 of the recommended brands matched.
That’s a problem if you want to be the brand it names. It’s a bigger one if you use AI to do your job, because the answer you got today may not be the one you get tomorrow.
You can’t take the dice away. You can tilt the odds, and most of that comes down to context: what the AI knows before it answers.
Too little context, too much context
Give it too little and you’re rolling the dice. Ask for “a launch plan” and the AI fills every blank with a guess: your audience, your budget, your tone. New guesses each time, new plan each time.
Give it too much and it starts ignoring you. A 2023 study led by Stanford researchers found models use what’s at the start and end of a long input best, and miss what’s buried in the middle. In 2025, Chroma tested 18 models. In one comparison, it asked the same questions two ways: with only the few paragraphs that mattered, or with those paragraphs buried in a long chat history. Every model in that comparison did significantly better with the short version.
The goal isn’t more context. It’s the right context.
Rule or standard?
Before you hand AI a job, sort each part of it. A rule has one right answer you can check: the speed limit is 65. A standard takes judgment: drive at a safe speed.
Most jobs are both. In a campaign recap, every number has to match the report. That’s a rule. Which result leads for your boss is a standard.
Things go wrong when you mix them up. Give AI the context to reason out a standard, and it’s better equipped to judge it. Ask it to judge a rule without enough context, and it’s bound to get something wrong, usually with total confidence.
Four ways to tilt the odds
- Write the context down once. Keep a plain text file, and attach it or set your AI tool to load it before every job. It should say who the audience is and what good looks like, with one good example and one bad one. Markdown (.md) files are the usual kind, and they’re your best friend. You stop retyping, and it does less guessing.
- Keep it short. Put what matters most at the top, and cut anything the AI doesn’t need for this job. A longer file isn’t a safer one.
- Put the rules in a script, not a prompt. A script is a few lines of code that does the same thing every time, like checking every figure against its source. You can have the AI write it for you. Test it on work you know is right and work you know is wrong, then run it before anything goes out. Skills, the saved instructions an AI pulls up for a task, still have a job. Skills guide the work. Scripts check the rules. A script doesn’t have moods.
- Give standards context, then a person. Let the AI make the judgment call with the file in front of it. Then someone who knows the audience reads it before it goes out.
The engines are doing it too
Every time someone asks an AI search engine “what’s the best…?”, it’s answering a standard, and your pages are part of the context it reasons from. So make your rules easy to read. Keep the same facts and claims everywhere, and publish your news on your own site first. In October, researchers got 8 of 10 AI search tools to cite a topic they made up. Nothing contradicted it. Don’t leave room for someone else’s version of you.
Give AI context for the judgment calls. Build checks for the rules.
Profound’s persona test is a vendor study. Chroma sells a database for AI apps, and its tests were deliberately simple tasks. The Stanford-led study tested 2023 models; Chroma’s 2025 tests found the same weakness in newer ones. The planted-topic study is an arXiv preprint.
The Follow-Up
OpenAI’s Business Hub showed one business 690 mentions and 24 clicks
Last Tuesday, OpenAI opened Business Hub, a beta portal where a company proves it owns its ChatGPT profile, then fixes its details, adds products with prices and sees how often ChatGPT mentions it. Two dashboards shared on X give an early look. One showed 690 mentions and 24 clicks over 28 days, the other 28 mentions and 1 click. Nobody outside OpenAI has said what counts as a “mention,” and the clicks tile warns: “Clicks don’t confirm website visits or conversions.”
Claim it if you can verify it. It’s a direct way to correct what ChatGPT has to work with. Read the counts month to month, not against anyone else’s. And don’t buy ads because the dashboard is there. OpenAI says businesses can sponsor their Business Agent as a landing spot for ChatGPT ads, and Seer Interactive found that people who searched a brand by name were shown a different company’s sponsored card in 5 of 6 industries. In last week’s one outside study, a visit from a ChatGPT ad brought stores 89 cents in sales, against $2.45 for a Google ad. (Beta; the two dashboards are anecdotes; Seer’s sessions weren’t a random sample. OpenAI’s page refused our fetch, so its wording comes from Search Engine Roundtable and PPC Land.)
The Numbers
- 79% of 28,500 sites with steady Google traffic got at least one visit from an AI tool in September, up from 39.7% in January 2025. The typical site with 1,000+ monthly search visits got 2 to 5 AI visits for every 100 from search. (SE Ranking study; a higher ratio can also mean falling Google traffic) Read →
- ~21% of one site’s Google clicks came from AI Overviews and AI Mode, by Cyrus Shepard’s estimate after he switched them off for a month and back on. (Single-site test) Read →
- 80%+ of the 1 million UK keywords Sistrix tracks now trigger a Google AI Overview. Retailers rose fastest: Go Outdoors went from 21.5% of its keywords on August 31 to 87.4% on November 9. (Sistrix data) Read →
The Rest
- Google started its November core update on November 10 and made a new model, Gemini 4 Hydrogen, available across Search two days later. If your Google traffic or AI citations move this month, note both dates before you guess at a cause.
- Yahoo became the first paying customer of the Internet Licensing Collective, a group of more than 3,000 publishers that licenses their content to AI companies together. Publishers are paid by how their content is used, including whether they’re credited. Terms weren’t disclosed.
- Bing now lets people turn AI answers off with a browser extension or by adding “-ai” to a search. Microsoft gave no figures on how many want to.
- Corrections — SparkToro’s open-web traffic study reports visits to the top million sites fell 12.2% from their 2022 peak. Our item first called traffic flat. All corrections →
All 19 items from November 9–15, dated and sourced: the week on the wire · State of Search · Corrections · Archive. Seen something the record should hold? Reply to this email; tips are credited by name.
The full wire
20 items published November 9–15. Every item dated, sourced and linked. An item can carry more than one lane, so lane totals exceed the count.
Google Search · 7
- Nov 9 — Zyppy opt-out test: AI Overviews, AI Mode drove ~21% of clicks — Zyppy SEO (Cyrus Shepard)
- Nov 10 — Google tests AI Overview local pack without action buttons — Search Engine Roundtable
- Nov 10 — Google brings Gemini 4 Carbon to AI Mode for AI Ultra subscribers — Google Keyword Blog (company announcement)
- Nov 11 — Google November 2026 core update: rollout may take 2 weeks — Google Search Status Dashboard (company announcement)
- Nov 11 — Sistrix: AI Overviews on over 80% of UK keywords — Sistrix Blog (vendor study)
- Nov 12 — Google Search adds Gemini 4 Hydrogen for all users worldwide — Google Keyword Blog (company announcement)
- Nov 13 — Google AI Overviews image generation reappears for some searchers — Search Engine Roundtable (single source)
Research · 6
- Nov 10 — SparkToro and Similarweb: open-web traffic flat since 2021 as web users grew 24% — SparkToro (vendor study)
- Nov 11 — Peec AI survey: 24% of companies say they can prove ROI from AI search — Peec AI (vendor study)
- Nov 12 — Seer study: ChatGPT ads reached 12 of 19 users by summer — Seer Interactive (vendor study)
- Nov 13 — Profound Black Friday AI search study: AT&T, Wayfair gain — Profound research blog (vendor study)
- Nov 13 — SE Ranking: 79% of sites got AI traffic in September 2026 — SE Ranking (SE Visible) (vendor study)
- Nov 13 — Scrunch: 91% of AI-driven visits carry no AI referral — Sitecore (Scrunch) (vendor study)
Engines · 4
- Nov 9 — Bing AI Mode now shows search and shopping ads at the top of results — Search Engine Roundtable (single source)
- Nov 10 — ChatGPT Ads now show during image generation — Search Engine Roundtable
- Nov 11 — ChatGPT Business Hub: OpenAI adds mention and click counts — Search Engine Roundtable
- Nov 11 — Bing adds a toggle and -ai operator to turn off AI answers — Bing Blogs (company announcement)
Tools & Vendors · 2
- Nov 12 — Profound adds Prompt Analysis to Prompt Volumes — Profound (vendor study)
- Nov 15 — TechCrunch robots.txt drops ChatGPT-User, Google-Extended blocks — Crawler access panel (first-party data)
Law & Money · 1
- Nov 12 — Yahoo Scout signs as first Internet Licensing Collective licensee — Digiday Media
Why it matters: The one place on the wire where the week is read rather than reported — compiled from the published record, framed by the editor, checkable line by line.
Glossary: SEOAI OverviewsAI Mode
Via AEO Wire ↗
Posted to the wire November 16, 2026. Edited by Joe Balewski.