Your AI Citations Disappeared in February. They Came Back Different.
Between mid-January and early March, the average number of citations ChatGPT included in answers to brand questions dropped from 4.95 to 2.96. That's a 41 percent decline in five weeks, pulled from over 170 million AI answers and 500 million citation records by Profound, reported in Built In a week ago.
If your brand gets cited on questions like “Is HubSpot worth it” or “Salesforce vs Pipedrive,” there's a good chance you saw something weird in February. Visibility trackers got noisy, referral traffic in GA4 looked off, and for a few weeks the dashboards stopped making sense.
By late March, citations had recovered to about 4.5 per answer, roughly 90 percent of the December baseline. Category queries (broader stuff like “best CRM for startups”) dropped 16 percent and came back to almost full recovery. The standard reading of all this is that ChatGPT rolled out ads, pushed a model update, the channel wobbled for a few weeks, and now we're back to normal.
I went through the Profound numbers a couple of times and I think that reading misses the most important part.
What Profound found
Profound tracked which sources got cited, not just how many citations went out. When the volume came back in March, the names attached to those citations were not the same names that had been there in December. A brand cited in December that got knocked out in February was not necessarily back in March, and the slots that opened up got filled by sources that had mostly not been getting cited before the drop, especially third-party reviewers, comparison content, and category-defining publishers.
So citation volume recovered while the composition of who gets cited shifted underneath it. Most teams' AI visibility tracking only asks whether they got cited at all, which means March looked clean and the alerts stopped firing, even though underneath the recovering volume number there was a quiet reshuffle going on that nobody was seeing.
This is consistent with what SE Ranking published in November 2025, finding that domains with strong presence on Quora, Reddit, and review sites like G2 and Trustpilot are about three to four times more likely to get cited by ChatGPT than domains without that footprint. The data is from last fall, so it's been kicking around for months. The thing that's worth re-examining in light of February is the threshold. Brands that scraped by in December with thin third-party citation surfaces appear to be struggling more in March, which suggests the bar for what counts as “enough” third-party signal has moved up.
The variance was already there
The other thing the February story buries is how much movement happens in AI citations on a normal day. AirOps published research late last year showing that only 30 percent of brands stay visible across consecutive answers to the same prompt. You run a query three times, and seven out of ten brands that showed up the first time are gone by the third. That's the baseline, with no model updates or ads launches involved.
Washington State University ran a similar experiment across more than 700 hypotheses, asking ChatGPT identical questions ten times each. The model gave consistent answers 73 percent of the time, and that was on the updated GPT-5 mini, so it's not a problem you can write off to an older version of the model.
This inconsistency was already in the system long before February. The February drop pulled enough brands out of the answer at once that the underlying noise became hard to miss, and a lot of teams that had been comfortably not measuring this carefully had to start paying attention.
You can't audit this quarterly
Most marketing teams are running AI visibility audits the way they run quarterly SEO audits, and that schedule does not match how AI citations behave. A regular SEO audit asks where you rank for a set of keywords right now and what it would take to move up, and that answer holds steady over a week or a month so you can plan against it.
AI citations move too fast for a quarterly cycle. You need to be sampling continuously, looking at what percentage of runs you appear in, and watching whether that percentage is moving.
A few things I've seen working for teams that started building for this earlier in the year.
Run your priority prompts across ChatGPT, Perplexity, Gemini, and Claude on a regular schedule. Daily for the prompts that matter most, weekly for the rest. Log which brands appear, which sources get cited, and at what frequency.
Track the whole citation mix, not just your domain. For your top 20 or 30 prompts, log every citation in every run. When competitor citations start climbing or new third-party sources start dominating, you have a leading indicator that the citation surface is moving, and some lead time to react before it shows up in referral traffic.
Set alerts at the prompt level for week-over-week changes above whatever threshold you decide matters. Teams I've talked to running this catch real shifts about five to seven days before GA4 sees them, which is usually enough time to investigate and do something about it.
This is certainly a chore, and most of it looks more like a quality monitoring program than a search campaign, which is probably one reason it's been slow to catch on. Most marketing teams are built around campaigns, and the systems we're now optimizing against do not behave like search engines in a way that rewards that structure.
What to do this quarter
If you've been on the fence about investing in AI visibility monitoring, the February data should settle it. The variance is here to stay, the composition of who gets cited will keep moving as model updates land, and the teams that started building for this in 2025 already have a year of data the rest of us don't.
A few things to do. Get a continuous monitoring setup in place this quarter, whether you build it yourself or buy it. Pick your prompt set, pick your platforms, pick a frequency that matches the stakes. Audit your third-party citation surface and figure out where reviewers, analysts, comparison publishers, and Reddit talk about your category, who they mention, and where you're missing. Stop handing leadership a single visibility number; a distribution of appearance rate, citation share, and source-mix stability tells you something you can act on, where a single number mostly tells you whether to feel good or bad.
The brands that came through February-March in the best shape were not the ones leaning hardest on schema markup or an aggressive AEO playbook, although a lot of them had those too. They were the ones who had spent enough time building a deep third-party citation surface that no single model update could push them entirely out of the answer. That kind of position takes a while to build, and the teams that started in 2025 are already further along than most of us.
The February drop is over for now, but whatever caused it is still in the system, and the next move like it is going to land harder on teams that treated this one as a one-off than on teams that used it as a reason to set up better tracking.
Jarred Smith is the author of Explainable: Why AI Recommends Some Brands & Ignores Others, an Amazon bestseller on AEO, GEO, and SEO. He's a marketing leader with nearly 20 years of experience across healthcare, public media, retail, and environmental services. Find him at jarredsmith.com.