Buyers Are Picking Vendors They've Never Heard Of. Thank ChatGPT.
POINT Key points
- 1 33% bought from a vendor they'd never heard of before
- 2 The shortlist now gets written where sales can't see it
HexScope Lens
Decoding the future through the lens of AI.
Which Of Your Press Hits Does The AI Actually Read?
POINT Key points
- 1 Half to 60% of the third-party mentions that work carry no link
- 2 Coverage lists need one more column: body text or footer
Your Press Release Gets Its First AI Citation In About Eight Hours
POINT Key points
- 1 Over 8,000 wire releases tracked for one week in May 2026
- 2 Well-structured content pulled 2.6x the average citations
Rewriting One Page Did Nothing. Six Aligned Pages Nearly Doubled AI Mentions
POINT Key points
- 1 One page got 35.9%. Six aligned pages got 67.2%
- 2 Align what your existing surfaces say before adding new ones
Same Search Result, 2.2x The Clicks If The AI Named You
POINT Key points
- 1 Cited brands pulled 2.2x the clicks on identical queries
- 2 Rank trackers have no column for "did the model name us"
41.4% Start Shopping With AI Now. Brand Sites Just Lost First Place.
POINT Key points
- 1 41.4% of 4,040 US and UK shoppers now start with AI
- 2 The AI share barely moved; going direct dropped 21 points
Claude Thinks On A Tiny Workbench. Your Brand Has To Get On It
POINT Key points
- 1 Claude holds a few dozen concepts at once, under 10% of its activity
- 2 Being known and getting named turn out to be separate measurements
The 'Which AI Did You Ask?' Problem Starts Below The Top Three
POINT Key points
- 1 Asked four AI assistants for the best social media management tools, 60 times
- 2 Three names came back in all 45 usable answers, then the models split
What 137,000 Sites Tell Us About Where AI Actually Reads
POINT Key points
- 1 "Just drop in an llms.txt file" is the standard AI advice
- 2 Server logs show AI reading the pages you already publish
What 25,000 AI Citations Say About Which Page Wins In Your Industry
POINT Key points
- 1 Top-cited page type flips by industry: 61% lists, 55% homepages
- 2 Comparison pages get cited 45% above average on 4.1% of supply
Let's Look At Where The AI Actually Files Your Company
POINT Key points
- 1 Brand-and-attribute memory maps to one cell on a grid inside the model
- 2 Nudge that cell and the answer changes, so you can act on it
Want AI To Quote You? Bring A Number It Can't Make Up
POINT Key points
- 1 Original research tops the citability poll at 82%, video sits at 2%
- 2 The budget case for making data instead of more blog posts
ChatGPT Names You, Nobody Clicks, And Visits Still Go Up 2.5x
POINT Key points
- 1 AI referrals stay under 1%, the lift hides in branded search
- 2 A ChatGPT mention drives 2.5x the visits of a competitor
Ask The AI About You 100,000 Times And One Number Holds Still
POINT Key points
- 1 Ask again and a named brand reappears more than 9 times in 10
- 2 Sentiment flips 45.5% of runs, so track appearance, not praise
Whose Job Is AI Search? 81% Of Merged Teams Saw More Traffic
POINT Key points
- 1 'AI search needs its own brand-new specialist team' is the received wisdom
- 2 But 81% of merged SEO teams reported gains, versus 36% when split
One Tiny Circuit Decides Which Source The AI Believes
POINT Key points
- 1 The jump to wrong info is a few middle-layer attention heads
- 2 The circuit is narrow, so how AI describes you is trackable
What 150,000 AI Citations Say About Who The Models Trust
POINT Key points
- 1 'You can't control word of mouth' is the usual shrug
- 2 Brands talked about on Reddit pull 3.9x the AI citations
Your Backlinks Barely Register With The AI. Your Reputation Does.
POINT Key points
- 1 Brand mentions correlated 0.664; backlinks stalled at 0.10 to 0.218
- 2 The lever is digital PR, not one more link-building sprint
Turns Out AI Picks Brands By Their Facts, Not Your Rank
POINT Key points
- 1 Fact data like Wikidata correlated up to 0.71; search rank explained just 5.8%
- 2 The real lever sits in the brand info other sites hold about you
ChatGPT And Perplexity Don't Read The Same Web. So Stop Optimizing Them The Same Way
POINT Key points
- 1 ChatGPT leans on Wikipedia, Perplexity leans on reddit
- 2 About 86% of AI citations come from sources you control
Want ChatGPT To Name You? Add A Quote And Watch Citations Jump 40%
POINT Key points
- 1 Quotes, stats, and sources lifted AI citation odds by 30-40%
- 2 Keyword stuffing went the other way — down about 8%
Can't Win At Search, So You're Doomed In AI? The Long Tail Says No
POINT Key points
- 1 Most AI citations come from pages ranked #11 and below
- 2 Rank-11 content still gets quoted, if the substance is there
Your #1 Ranking Means Less Now. Getting Cited By AI Means More
POINT Key points
- 1 68% of US Google searches now end without a single click
- 2 Being named inside the AI's answer beats sitting at #1
The Sale Pitch That Works On People Quietly Loses You The AI
POINT Key points
- 1 Real-review wording lifted one model's recommendations 334%
- 2 'Only 3 left' and 'members only' backfire and drop you off the list
When The AI Explains Why It Picked You, It Might Be Making That Up
POINT Key points
- 1 Slip a bias into the input and the model follows it, then explains its answer without ever mentioning the bias
- 2 Treat the AI's stated reason as a hypothesis; track the appearance rate you can actually observe
AI Plays Favorites With Famous Brands. A 0.1-Star Edge Undoes It.
POINT Key points
- 1 Spec-for-spec, the AI hands the famous brand 100% of the recommendations
- 2 What flips it isn't hype — it's a sliver more rating and outside backing
Reviews Take Your AI Citation Rate From 1% To 75%
POINT Key points
- 1 Just having a review profile lifts AI citations from 1% to 53.5%
- 2 Collect reviews and reply, and it hits 75.3%, so visibility is buildable
Why ChatGPT Keeps Changing Its Mind Depends On What You Asked
POINT Key points
- 1 Ask a purchase question twice and only 4 of 10 brands show up both times — a 40% match
- 2 To track your AI visibility, fix the intent you measure with and read trends, not single runs
Your Best Buyers Now Come From AI. Can It Even Read Your Site?
POINT Key points
- 1 AI-referred shoppers now convert 42% better than regular ones
- 2 But a third of product pages can't be read by AI
Turns Out AI Traffic Beats Paid Social, Especially On Complex Products
POINT Key points
- 1 The viral '9x from ChatGPT' is one B2B case, not a rule
- 2 Across 973 stores, AI referrals beat paid social but trail search
Shoppers Would Let AI Buy — If They Could Hit The Brakes
POINT Key points
- 1 A third expect AI to run part of their shopping within a year
- 2 What unlocks it isn't a smarter AI, it's a working stop button
You Figured AI Citations Were Out Of Your Hands. 86% Aren't.
POINT Key points
- 1 6.8 million citations counted, and 86% came from sources you control
- 2 Before blaming Reddit, the fix is your own site and listings
Turns Out ChatGPT Mostly Cites Pages From The Last Year
POINT Key points
- 1 About 65% of the pages AI cites went up in the last year
- 2 So 'publish once' is out, and freshness is the new lever
Is ChatGPT Ranking Your Brand, Or Just The Order You Listed?
POINT Key points
- 1 Reshuffle the same candidates and the AI ranking moves about 30%
- 2 So shuffle and re-ask before you trust an "AI put us first"
Is Your ChatGPT Visibility Number Still True Five Weeks Later?
POINT Key points
- 1 Brand citations per answer fell from 4.95 to 2.96 in five weeks
- 2 The total came back, but the mix of cited sources didn't
The "Whose ChatGPT Did You Even Check?" Problem
POINT Key points
- 1 ChatGPT's factual recall went from 67.9% in 2025 to 82.8%
- 2 So a blank-account spot check measures nobody in particular
Eight Products On The Shelf, And The Agent Sees Two Or Three
POINT Key points
- 1 One Fitbit went from 45% to 77% between two model versions
- 2 Rewriting the product copy raised the share of picks across most models
Say The Same Fact Lots Of Ways And ChatGPT Can Recall It
POINT Key points
- 1 Taught one fixed phrasing, the model couldn't answer questions about it
- 2 Backwards, from the description to the name, accuracy sat near zero
Turns Out The Model Plans Before It Answers, Just Not Like You
POINT Key points
- 1 Made-up problems it can't have memorized still get solved
- 2 Mistakes plus fixes beat clean answers as training data
What Happens When You Build An AI Copy Of 1,052 Real People
POINT Key points
- 1 The AI stand-ins matched the real answers about 85% of the time
- 2 Good for drafting hypotheses, not for measuring your real brand image
ChatGPT Cited Your Brand. Now Go Read What It Said.
POINT Key points
- 1 Journalists graded 3,000 AI answers: 45% had a significant problem
- 2 A bad attribution dents the brand that got quoted, not the model
Ranking #1 And Getting No Clicks: When The AI Just Answers For You
POINT Key points
- 1 On pages with an AI summary, only 8% click an external link vs 15% without — roughly half
- 2 The metric shifts from 'what rank am I' to 'does my brand show up inside the AI's answer'
Why Won't ChatGPT Just Say "I Don't Know"? Blame The Grading
POINT Key points
- 1 Marking "I don't know" as zero taught the models to bluff
- 2 A confident tone tells you nothing about whether the answer is right
Does ChatGPT Pick The Best Brand, Or The Best-Known One?
POINT Key points
- 1 Rich-country prompts got luxury brands 88 to 100 percent of the time
- 2 Local and newer names start behind, whatever the product is worth
Why Bigger AI Gets Smarter Turns Out To Be A Packing Problem
POINT Key points
- 1 Double the model's width and it misses roughly half as often
- 2 Gives you a footing for how far scaling keeps paying off
Somebody Watched An LLM Grow Up, And The Parts Split Into Specialists
POINT Key points
- 1 The score curve slides down smoothly while roles get assigned underneath
- 2 Judging an AI on its final score alone gets harder to defend
Berkeley Did The Math On Why 'Here's An Example' Prompts Work
POINT Key points
- 1 Shown two or three examples, the model builds the best rule on the spot
- 2 It all evaporates when the chat closes, so template the examples
You Think AI Gets Smarter Slowly. It Might Just Lurch.
POINT Key points
- 1 "Grokking": a long memorization plateau, then a sudden jump
- 2 A one-time verdict on any AI ages badly within months
The Trick To Getting Cited By AI Is Boring: Add Numbers And Sources
POINT Key points
- 1 Tested GEO tweaks lifted a page’s citation rate in AI answers by up to 40%
- 2 Keyword stuffing did nothing; what worked was numbers, sources, and third-party voices
The AI Quotes Reddit, Not Your Website
POINT Key points
- 1 Reddit tops the AI's source list -- shown across a 30-million-source analysis
- 2 What earns citations isn't your own site, it's the third-party places where others talk about you
The AI Read 38,000 Of Your Pages And Sent You One Visitor
POINT Key points
- 1 Google returns a visitor every 5.4 pages. Anthropic, every 38,065.
- 2 Brand-name searches held up while general search fell over 20%
ChatGPT Agrees With You 49% More Than People Do
POINT Key points
- 1 Eleven current models backed users 49% more often than humans
- 2 One exchange left people surer they'd been right
Six In Ten Have Now Used Generative AI. Trusted Brands Matter More.
POINT Key points
- 1 Generative AI use jumped from 40% to 61% in a single year
- 2 AI-made news scores -18 on trust, so brand names do the vouching
There's A Golden Gate Bridge Knob Inside Claude, And Someone Turned It
POINT Key points
- 1 Anthropic pulled about 34 million features out of Claude 3 Sonnet
- 2 Read AI answers for learned bias, not just for the wording
A 70B Model Beat A 280B One. So Much For Bigger.
POINT Key points
- 1 Chinchilla's rule was double the model, double the training data
- 2 What to ask a vendor instead of which model they run
Let's Take "The AI Is Just Predicting" To A Brain Scanner
POINT Key points
- 1 Models better at guessing the next word matched brain activity better
- 2 Fluent isn't sourced, so read AI answers beside primary data
Does ChatGPT Grade Your Copy Like A Customer Would?
POINT Key points
- 1 GPT-4 picks its own writing 89% of the time
- 2 Human raters pick it 36%, so the scorecard splits in two
When The AI Sounds Certain, That's Your Cue To Check
POINT Key points
- 1 Steering how sure the model sounds cut confident hallucinations 30% on short answers
- 2 That tone is a real internal signal, but not a correctness meter
How AI Decides Which Brands It Remembers
POINT Key points
- 1 Exact AI recommendation lists almost never repeat, but appearance rate stays stable
- 2 Being remembered comes from a consistent category story across your sources
When To Trust The Decimals On Your AI Dashboard
POINT Key points
- 1 Fine decimals prove nothing when the measurement can't tell them apart
- 2 Trust a number that shows its source data and reproduction conditions
What Turns An AI Persona Into An 85% Match
POINT Key points
- 1 85% of a real person's answers, reproduced from a two-hour interview
- 2 Thin attribute personas drift to an average; final calls still need human data
What Hundreds Of Real ChatGPT Questions Reveal About 'Best Tool' Prompts
POINT Key points
- 1 Real questions to ChatGPT run 71% longer than estimated prompts
- 2 Best-type prompts alone measure a conversation users rarely have
Is The AI Actually Inconsistent, Or Just Badly Measured?
POINT Key points
- 1 Individual AI answers wobble, but brand appearance rates stay stable
- 2 Much of the measured flakiness comes from the eval method itself
2,961 AI Answers Later, Rank Is Noise And Rate Is Signal
POINT Key points
- 1 Across 2,961 AI queries, the exact list matched under 1%, order near 0.1%
- 2 The fix is tracking appearance rate and standardizing how many times you measure
Before You Hand Consumer Research To AI, Split It Three Ways
POINT Key points
- 1 AI respondents, AI question writers and real prompt logs solve different problems
- 2 Pick which stage to use them on, not which one to adopt
'Prompt Search Volume' Looks Precise. That's The Problem.
POINT Key points
- 1 The panels behind the number skew to desktop Chrome and miss mobile
- 2 Safe to read as direction, never as an absolute you can budget against
1,052 People, Cloned By AI, Matched 85% Of Their Own Answers
POINT Key points
- 1 Two-hour interviews, not demographic labels, drove the 85% match
- 2 Best for front-loading hypotheses and survey questions, not replacing real research
Your AI Tool Is Guessing The Prompts. Use It Anyway?
POINT Key points
- 1 Estimated prompts run 8.8 words; real ones run 15.1 and stay personal
- 2 But the brand rankings barely move, so read them as comparison, not truth
Your AI Rank Keeps Moving. ‘Appearance Rate’ Still Measures Your Brand.
POINT Key points
- 1 Across 2,961 AI queries, individual rankings almost never reproduced
- 2 What works in practice is the appearance rate from repeated measurement
Your Next Buyer Could Be An Agent. Stock Two Shelves.
POINT Key points
- 1 Forecasts run from 8.8% of ecommerce transactions to half of them by 2029
- 2 Recommendations swing by model, so log ChatGPT, Claude and Gemini separately
ChatGPT Took A Fifth Of Search. The Answer Isn’t To Abandon SEO.
POINT Key points
- 1 ChatGPT-habit households cut search volume 20%+ within 20 weeks
- 2 Informational queries fall — defend core SEO, migrate the rest to GEO
Letting AI Draft Your Survey Is Fine. Shipping It Raw Isn't.
POINT Key points
- 1 AI drafts survey questions fast, but ships them verbose and double-barreled
- 2 A three-step flow — draft, unify terms, human review — keeps the data clean
ChatGPT Names A Different Brand Every Time. Measure Anyway.
POINT Key points
- 1 The recommendations really do shuffle, but part of that shuffle is your measuring method
- 2 What holds steady is appearance rate across repeats, not the single-shot ranking
Only 27% Go On Gut Now. Design Proof Into Your AI-Era Content.
POINT Key points
- 1 68% doubt the content is authentic, 61% doubt it's even reliable
- 2 Track whether the verification path gets clicked, not just how it reads
Does One Wrong Word Really Break ChatGPT's Answer?
POINT Key points
- 1 Reword the prompt and accuracy moves 3-10%, not the meltdown you'd expect
- 2 What pays off is clearer instructions and how you grade answers, not magic phrasing
Your Customers Trust AI. They Just Don't Trust You Using It.
POINT Key points
- 1 60% trust AI's answers, but 50% avoid brands that flaunt it
- 2 The fix isn't whether you use AI, but where you let it show
ChatGPT Traffic 13x'd In A Year. I'd Still Wait On The Budget.
POINT Key points
- 1 GenAI retail traffic 13x'd in a year, but referral share is still 0.21%
- 2 Set up two-axis measurement and small tests before you move budget
The Harder The Product, The More ChatGPT Traffic Pays Off
POINT Key points
- 1 AI traffic is about 1% of visits but pulls roughly 10% of sales
- 2 So roll out GEO on high-involvement categories first, not all at once
ChatGPT Traffic Out-Converts Paid Social. Read Two Numbers, Not One.
POINT Key points
- 1 LLM referrals out-convert paid social and grew 40% in one quarter
- 2 Share is still near 1%, so track conversion and referral share together
A Small Model Just Crashed The Big-Model Party
POINT Key points
- 1 A bolt-on layer lets small models reach the big-model band
- 2 So 'wait for next year's AI' is the actual losing move
You Fed The AI Your Data, So Why Won't It Answer? A Study Breaks It Down
POINT Key points
- 1 The same fact stored in only one phrasing is hard for the model to pull back out
- 2 Reverse-direction questions and a hard capacity ceiling are the blind spots in your setup
When You Tell The AI To "Think," Is It Really Thinking?
POINT Key points
- 1 In controlled math tests the model ran a hidden reasoning step before it answered
- 2 Models trained on mistakes-and-fixes reasoned more accurately than ones fed only correct answers
Two AIs Can Tie On The Test And Still Be Different Machines
POINT Key points
- 1 How a model stores word position alone changed how well it understood
- 2 So pick the AI that fails quietly on your task, not the top score
Does The AI Actually Understand, Or Just Sound Like It?
POINT Key points
- 1 Peek inside the model and you find it rebuilds a sentence's structure, not a flat word list
- 2 So compare candidates by how they fail, not by how fluent they sound
The Highest-Scoring AI Isn’t Always The Smartest One
POINT Key points
- 1 A top benchmark score can hide whether the model understood the problem or just pattern-matched a familiar shape
- 2 Before you trust a leaderboard, read the test design — that’s what tells you if a score will hold up on your work
Nice to meet you —
I'm Hex-ko.
I edit this publication, HexScope Lens. We cover the heart of AI marketing, grounded in research papers, in a relaxed way.
- Who is it for?
- Marketers, SEO leads, and executives — people with an eye for seeing through vague hand-waving about AI, but who aren't engineers.
- Tone
- Written with the warmth of a friend's opening line when they tell you about a cool study at a party. No hype, no talking down, no reporting-speak.
- Who runs it
- The owned media of HexScope, an AI brand-tracking SaaS.