You’ve checked how your brand shows up in AI, right? You asked ChatGPT a few questions about your category, counted how often your own name came back, and wrote the number down.
I did the same thing. And I filed that number away the way you’d file a height measurement: you take it once, and it’s good for a while.
But it isn’t good for a while. Your visibility inside an AI answer moves in weeks when you changed nothing at all, because the platform changed something instead.
So let’s look at a report that watched that movement across 170 million real answers, and work out what a marketer is supposed to do about it.
Sixteen Weeks Of Watching 170 Million AI Answers
The report is The 2026 State of AI Search, published by AirOps, a company that builds AI visibility measurement tools. (Worth flagging up front: they sell into this market, so read the framing with that discount applied.)
They tracked answers from the major engines — ChatGPT, Perplexity, Gemini — for about 3,000 brands, continuously.
The scale is the reason to care. From December 2025 through March 2026, sixteen weeks, they logged 170 million AI answers, and the citation records inside them came to more than 500 million.
A “citation” here is just the page the AI leaned on while writing its answer — the record of which source it used to talk about which brand. For you, that’s the count of how often the AI picked your brand up at all.
The Week The Ads Landed
What makes this report unusually vivid is that it was already running when something happened.
The something: on February 9, 2026, OpenAI started showing ads inside ChatGPT. More than 600 companies bought placements, and the answer screen started carrying advertising.
So what happened on brand queries — the “what’s a good X?” kind of question? The average number of brand citations per answer fell from 4.95 to 2.96. Five weeks, a 41% drop.
Nobody’s marketing changed. The exposure just left.
It’s the seismologist who happened to be staring at the needle when the real quake arrived: lucky, and a little horrifying. And from where you sit, it means your visibility moved for reasons that have nothing to do with whether your work was any good.
The Total Came Back. The Line-Up Didn’t.
But it didn’t stay halved, which is the part I keep chewing on.
By late March the citation count had climbed back to around 4.5 — roughly 90% of where December sat. Look only at the total and you’d close the tab thinking “fine, a temporary shock, we’re back.”
Here’s the catch. The total mostly came back, but AirOps points out that the mix of source types cited on brand queries changed permanently. Same headcount, different faces.
Reading the sum and relaxing, when the contents got swapped out underneath, is like looking at a team that replaced its entire roster and calling it last year’s team because the points tally matches. Your number recovered. Where the AI goes to talk about you did not.
So Where Does The AI Even Get You From?
Which raises the obvious question. Who are those faces? And the answer stings a bit, because most of them sit somewhere you can’t edit.
In the AirOps measurements, about 48% of the citations in AI answers came from community platforms like Reddit and YouTube. Narrow it to brand mentions and it gets starker: 85% started from a third-party page rather than the brand’s own domain.
“Third-party page” means everything that isn’t your official site — review write-ups, comparison pages, forum threads, other people’s turf.
So when the AI talks about you, most of its evidence lives outside your control. You can polish the official site forever; what the model is mostly reading is what the world wrote about you. When that moves, you move.
Conclusion: Treat It Like A Vital Sign, And Look Past Your Own Domain
Pull it together and the point is this: AI brand exposure isn’t a static number you measure once and shelve. Every time a platform adds ads or ships a model update, it can swing hard in a matter of weeks, with your own work held constant.
So the first thing I’d change is the cadence. Measure it the way you’d track a vital sign — same check, fixed interval. A once-a-year snapshot would have missed a 41% move across five weeks entirely, and it’s only the month-over-month line that tells you whether a change came from your campaign or from a shift in the platform underneath you.
The second thing is where you point your attention. With roughly half of citations coming from community platforms and 85% of mentions starting on third-party pages, the reviews, comparison posts and threads about you are doing most of the talking. Your own domain is a smaller lever than it feels like.
Maybe I’m reading too much into one report from a vendor with a stake in the answer. But the direction seems hard to argue with: tidying your own site and calling it done is the part worth retiring. Start by finding out who out there is currently telling the AI who you are.
Sources
- AirOps, “The 2026 State of AI Search”, airops.com (continuous measurement of AI answers from ChatGPT, Perplexity and Gemini for roughly 3,000 brands. December 2025 to March 2026, sixteen weeks: 170 million AI answers and more than 500 million citation records. After OpenAI began showing ads inside ChatGPT on February 9, 2026, with more than 600 advertisers, average brand citations per answer on brand queries fell from 4.95 to 2.96 over five weeks, a 41% decline; by late March the count had recovered to about 4.5, roughly 90% of the December baseline, but the mix of source types cited on brand queries changed permanently. About 48% of citations came from community platforms such as Reddit and YouTube, and 85% of brand mentions originated on third-party pages rather than the brand’s own domain. Note: AirOps sells AI visibility measurement tools.)
