We publish measurements and method.
Nobody needs another ten-tips post about answer engines. What's missing is somebody sampling them on a schedule and printing the numbers with the intervals attached, including the numbers that argue against buying anything from us.
Latest notes
Each one names its sample, its date, and where it could be wrong. That is the whole format.
ChatGPT called one tool both GEO and AEO on the same day
We asked ChatGPT, Perplexity, Claude and Google AI Overviews 25 questions about tools that track brands in AI answers. None of the questions said AEO or GEO. The answers used GEO in 64 of 180 and AEO in 31, often for the same vendor. Here's where the two terms came from, why the label won't help you choose, and the split that will.
Read the noteSame AI mention rate, named first 38 times vs 15
Most AI visibility tools lead with a mention rate. In one category we track, two companies were each named in 70 of 160 answers from ChatGPT, Perplexity, Claude and Google AI Overviews. One was named first 38 times, the other 15. A mention rate can't see that. Here's what a tool should measure before you pay for it, and the questions that tell you whether it does.
Engines cited 129 sites that publish llms.txt 6,343 times. Once, it was the file
llms.txt is the most recommended fix in AI search, and the least checked. We took the 300 domains that ChatGPT, Perplexity, Claude and Google AI Overviews cited most in our samples and fetched each one's file. 129 publish one. Those 129 sites were cited 6,343 times, and one citation pointed at the file. Eight of the ten most cited sites have none. Here's what the file is for, and what to spend the hour on instead.
Our tracker missed Google's AI Overview on 96% of searches. One setting fixed it
From August 22 to September 7 our sampler recorded no Google AI Overview for 193 of 201 searches. Google had shown one almost every time. The Overviews load after the page, and our data vendor only returns them if you ask. On the same 36 questions after the fix, 3 of 487 searches came back empty. Here's what that means for any number a tracker gives you about Google.
Seven universal sites took 5% of citations in 15 categories
Ask which sites to get mentioned on and you'll get the same list: Reddit, Wikipedia, YouTube, LinkedIn. Those lists are real, and they're built from every topic people ask about. We counted the citations behind buyer questions in 15 categories instead. The universal sites barely registered. Almost everything cited belonged to one category alone.
When one AI engine named the brand, all four did 32% of the time
Most AI visibility reports give you one number, or one number per engine that looks about the same. We put the same unbranded questions to ChatGPT, Google AI Overviews, Perplexity and Claude on the same day and compared the answers question by question. The engines disagreed with each other more than any one of them disagreed with itself the next day.
Competitor websites took 16% of the AI citations we counted. The brand's own site took 0.87%.
Half the advice says publish depth on your own domain and the engines will find it. The other half says your own domain barely matters. We counted every citation in 1,205 answers to unbranded buyer questions. Vendor websites were the biggest single block of sources, so the channel plainly works. They were almost never the website of the brand doing the asking.
Software directories are 1.8% of AI citations, and the cited page is almost never yours
The standard advice is to claim your G2 and Capterra profiles and fill in every field, because review profiles are where AI answers come from. We counted every citation in 1,099 sampled answers. Directories took 1.8% of them, and the page an engine took was usually a category ranking or a competitor's alternatives list. The brand's own profile came up 8 times out of 9,198.
ChatGPT named the brand 78% of the time it searched, and 6% when it didn't
You ask ChatGPT who the best tools in your category are, your competitors come back, and you don't. The standard explanation is that your content isn't authoritative enough. In our sample the likelier explanation was that the model never looked anything up. It answered from memory, and when it did that it named the brand 6% of the time instead of 78%.
59% of the YouTube links in AI Overviews jumped to a timestamp
Every guide says to start a YouTube channel, because Google favors its own platform. YouTube was indeed the most cited domain in our sample, at 10.4% of Google AI Overview citations and 0.7% of ChatGPT's. Then we looked at the links themselves. Most of them skipped to a specific second inside the video, usually inside the first two minutes, and 11 of 533 were on the channel of the brand we were measuring.
Reddit is 3% of AI citations, and every one is a single thread
Every guide to getting recommended by ChatGPT says to get on Reddit. We counted what two engines actually cited over five weeks. Reddit sat near the top of the domain list and still carried about 3% of the links. All 183 of them pointed at a single thread, and the two engines almost never picked the same one. That changes what a brand should do there.
Thirty days can change what ChatGPT looks up, not what it remembers
You asked ChatGPT for the best tools in your category and it named five companies that aren't you. Before you publish anything, know which half of the answer a new page can reach. The model's memory was frozen months ago. The lookup half reads the web today, and only when the engine decides to search. Here's how we'd spend the first thirty days, and why we wouldn't grade new pages at the end of them.
Fix the line a roundup already wrote about you
Every guide to AI visibility tells you to get named in more roundups. Nobody tells you that the roundups already naming you are often wrong about your price, your plan and who you're for, and that the model repeats their sentence instead of your homepage. Zapier publishes its selection criteria, its testing method and a form, and the form's one firm commitment is to fix incorrect information. That's the cheapest move on the board.
Your error bars overlap and the 12 point drop is real
Your visibility went from 42% to 30% and the error bars still touch, so nothing fired. Two 95% intervals drawn from identical populations overlap more than 99% of the time, which makes the overlap check roughly a 1% test instead of the 5% you wanted. On the same 200 prompts a proper test gives p = 0.012, and comparing the prompts that actually flipped gives p = 0.00006. Here is the arithmetic, including what our own alert gate gets wrong.
Every AI visibility number you have is a turn-one number
Every answer-engine tool, ours included, runs each prompt in a fresh chat and reports who got named. Buyers do not stop at one question. In a 200,000 conversation experiment across fifteen models, delivering the same request over two or more turns cut performance by 39% and more than doubled the gap between the best and worst run. The shortlist at turn two is a different measurement, and nobody is taking it.
The known brand wins every time, until a competitor is a tenth of a star better
A June 2026 study put three models in front of ten products where only the brand name differed. The known brand won every single trial. Then the researchers gave an unknown competitor the smallest advantage they could measure, and the monopoly collapsed in one step. Once any real difference exists, brand identity explains 1.2% of what gets ranked first.
Mentions hold still. Sentiment flips 6.7 times more often.
Whether an engine names you on a category question turns out to be close to a settled fact: 77.5% of tracked cells were strictly always or never mentioned. What the answer says about you is the part that moves, flipping on nearly half the cells that had anything to report. Most dashboards lead with the stable number and set alerts on it.
Ask for the best software and 73% of the cited sources are third-party sites
An AI answer isn't a verdict on your page. It's a set of roughly eight sources, and on a software question about three quarters of them belong to reviewers and comparison sites. Co-citation, a measure defined in 1973, is the useful lens: two documents become neighbors because somebody else kept reaching for both. You don't pick who you appear next to, and that company shapes how a buyer reads you.
Zero citations doesn't mean the model didn't use your page
The link chips under an AI answer feel like a receipt. They are closer to a caption: text the model produced about its own retrieval, after the fact. The Tow Center handed eight AI search tools an exact paragraph and asked who published it, and the tools got it wrong more than 60% of the time. The gap between what an engine retrieved and what it credited is where a lot of brand visibility quietly goes missing.
You asked once and the model named you. The honest reading is 21% to 100%.
Someone types the category question into ChatGPT, sees the brand in the answer, and screenshots it for the channel. That screenshot is one Bernoulli trial. Put a 95% Wilson interval on it and the range that stays consistent with what you saw runs from 20.7% all the way to 100%. The engines are non-deterministic by design, and most of what gets reported as a change in AI visibility is a sample too small to have a finding in it.
Google's AI Mode turns one question into a dozen searches. You track one of them.
Google has written down how AI Mode works, and it is not a ranked list. One question is broken into subtopics and answered by a set of searches the engine writes itself. A Google engineering director puts the everyday case at a dozen. If you are tracking one head term, you are watching the only question in the set that you chose.
Blocking GPTBot doesn't remove you from ChatGPT. Blocking OAI-SearchBot does.
There's a robots.txt line going around that blocks GPTBot, and a lot of people added it thinking they'd made a decision about ChatGPT. They made a different one. Here's which agent actually gates which answer, straight from each engine's own docs, and how to check your site in an afternoon.
When an AI summary appears, clicks fall from 15% to 8%
Pew watched 900 people search for a month. When an AI summary showed up, people clicked a result 8% of the time instead of 15%, and clicked a source inside the summary 1% of the time. Here is what that does to a brand, and where the number could be wrong.
Three findings we keep coming back to
All three are our own live measurements, in the category we tested. None of them are market forecasts, because we can't source one we'd trust.
Listicles were 44% of what ChatGPT cited for “best X” in the category we tested. Review sites 24%, vendor pages 18%, editorial 14%. Most of those listicles take submissions, so they're the cheapest gap in the whole report to close. There's a post in that, and no secret.
No engine we track documents llms.txt as a ranking signal. It's the most-recommended item in this space. It's cheap to publish, and we publish one, but it won't get you named on its own. We'll keep saying so, including when it costs us a sale.
The same five to seven names hold across almost every phrasing. If the list reshuffled at random it wouldn't be worth winning. It doesn't. It holds, which is why it's worth getting into and worth defending once you're there.
Every post names its sample, its date, and where it could be wrong. That's the whole format.
Ask the same question tomorrow and 9 to 27% of the names change.
Perplexity moved most at 27% day over day, Claude least at 9%, in our own sampling. An article about answer engines goes stale much faster than an article about search ever did. That's why we'd sooner publish a running measurement than a take.
- Every figure carries its interval, a 95% Wilson score interval, so we can't sell you a wobble as progress
- If a post quotes a number, it quotes the day we sampled it
- When a finding stops holding, we say it stopped holding
Perplexity27%
ChatGPT18%
AI Overviews12%
Claude9%Average change in which brands get named, same question, one day apart. Proofsource sampling.
On the writing list
What we are measuring and writing up next.
One answer is not a measurement
Why a single run carries an interval, and what a Wilson interval actually rules out.
Reading a citation like a map
The sources behind an answer tell you which pages you have to be on. We'll show the working.
Presence, position, framing
Being named, being named first, and being named well are three different measurements.
Your own numbers beat our best post.
The shortlist in your category is being written right now, whether anyone reads this page or not. A free trial, 25 questions on four engines today and tomorrow, tells you if your name is in it.
Google AI OverviewsGoogle AI Mode
