Ahrefs Tracked 37,500 AI Citations, and Daily Snapshots Look Mostly Useless

Ahrefs Tracked 37,500 AI Citations, and Daily Snapshots Look Mostly Useless
In Ahrefs' 20-day test, 50% to 83% of AI citations vanished overnight, but about 75% of the lost ones came back.

Ahrefs tracked roughly 37,500 AI citations across ChatGPT, Google AI Mode, AI Overviews, Gemini and Copilot for 20 days and found that 50% to 83% of day-one citations were gone the next day. About 75% of those lost citations came back within 20 days. A single-day AI visibility check is mostly noise; the useful metric is days cited out of the last seven.

That's the headline from Ahrefs' new citation stability study, written up by Mateusz Makosiewicz with data work from Xibeijia Guan. The setup was simple: ask five AI tools the same questions every day for 20 days, then follow every site cited on day one. AI Mode alone accounted for 1,996 questions and 18,009 cited websites. ChatGPT had 3,173 questions. Perplexity isn't in the data, which matters if that's where your audience lives.

I think this study matters less for what it says about AI search and more for what it says about the dashboards a lot of teams started paying for this year. If a weekly report has a line that says "we lost 12 AI citations," there's a decent chance most of them are already back.

Half of everything is gone by breakfast

The next-day loss rates by tool: AI Overviews 50%, AI Mode 58%, ChatGPT 59%, Gemini 81%, Copilot 83%. The Gemini and Copilot numbers look apocalyptic until you see the other finding. Gemini returned answers with zero citations 64% of the time, and Copilot 53%. So a big chunk of their "loss" is just the tool deciding not to cite anyone that day. Strip out the empty answers and all five tools land in the same 50% to 60% churn range.

The part that surprised me was ChatGPT. Ahrefs found it reused the same fan-out searches only 26% of the time, which you'd expect to explain the churn. But even when the searches were identical, the cited sites changed 81% of the time. So the model isn't only asking different questions each day. It picks different sources off the same results. If you've read the earlier Ahrefs breakdown of ChatGPT's query fan-out, this is the next layer down, and it's messier than the fan-out story alone suggested.

Glen Allsopp's AI volatility research at Detailed, which Ahrefs checked its patterns against, points the same direction. The most-cited brands and domains changed 11% to 14% day to day. Everyone else churned at 61% to 78%.

Think of AI citations less like rankings and more like a restaurant's daily specials board. The regulars are always there. Everything else rotates, and getting upset that the soup changed on Tuesday doesn't tell you much about the kitchen.

Most of it comes back (if it lasted long enough)

Only 20% to 27% of day-one citations disappeared and never returned during the 20 days. About half of the lost ones came back within three days. On AI Overviews and ChatGPT, roughly 40% returned after a single day off.

The more useful finding is what Ahrefs calls a Lindy effect: the longer a citation has lasted, the more likely it is to come back after it drops. Citations that lasted only one day never returned 34% to 41% of the time, depending on the tool. For citations that held for 8 to 14 days, the never-returned rate fell to between 0.7% and 4%. And next-day return rates jumped from 14% to 25% for one-day citations to 69% to 87% for citations that had held for 8+ days.

Which flips how you should read both wins and losses:

A new citation is not a win yet. A long-held citation that disappears is probably not a loss yet.

In Brand Radar's tracked questions, about 70% of brand-new citations were gone the next day, and only 34% to 45% of those came back. So the screenshot someone posts in Slack the morning you first show up in ChatGPT is, statistically, a coin flip at best.

Rankings still do the heavy lifting on Google's AI surfaces

This is the part I'd forward to anyone who's decided SEO is dead and GEO is the new job. Pages that ranked in Google for the same question were about seven times more likely to stay cited for a week or more on AI Overviews and AI Mode. The split: 29% vs. 4% on AI Overviews, 28% vs. 4% on AI Mode. ChatGPT showed a weaker gap, 11% vs. 4%.

Organic traffic tracked stability too. On AI Mode, the median cited page that lasted one day had 1 monthly organic visit; pages cited 15+ days had a median of 124. Domain Rating barely moved (typically 64 to 78 regardless of how long a citation lasted), so DR seems close to useless as a predictor here. That lines up with Ahrefs' earlier study of 1.9 million AI Overview citations, which also tied citations back to ranking positions.

And in fairness to the GEO crowd, there's a real counterpoint in the same data. Around 75% of AI Mode and ChatGPT citations came from pages that didn't rank for that question at all. So ranking isn't required to get cited. It just seems to be what keeps you cited. Getting in and staying in look like two different problems, and most AI visibility tools only measure the first one.

It also connects to something we covered earlier today: HubSpot lost 80% of its blog traffic and kept growing revenue. If AI citation persistence follows organic traffic, sites that shrug off traffic losses may quietly be shrugging off their AI staying power too. I'm not sure HubSpot cares. Smaller sites probably should.

Brand mentions are the steadier signal

Most citations don't come with a brand name attached. On AI Mode, 84% of cited sites weren't named in the answer text. On ChatGPT, 54%. And when a citation disappeared, the brand name sometimes stuck around anyway: 40% of the time on Gemini, 13% on ChatGPT, 12% on AI Overviews.

There's also a data hygiene problem. In a manual check of 55 brands, Allsopp found that 30% to 36% of apparent losses were the same brand showing up under a different name. So if a tracker only matches "Acme Analytics" exactly, it may be logging "Acme" or "AcmeHQ" as a loss.

Brand-named questions are the most stable of all. When the question names a brand, that brand's own site was cited all 20 days 28% to 33% of the time on Google's tools and ChatGPT, compared with 4% to 7% for any cited site. But in Ahrefs' smaller 22-question test, that mostly held for factual prompts. On review prompts, the brand's own site was cited 5% of the time on AI Overviews and AI Mode. On comparison prompts, 0% to 5%. Google's AI apparently doesn't want to hear your opinion of yourself, which, honestly, is fair.

"Best X" queries are the casino floor

On ChatGPT and AI Mode, 64% to 76% of cited pages changed day to day regardless of question type, but "best X" queries were among the least stable and brand-named ones the most stable. Allsopp's industry cut is wild: health questions kept the same first source 56% to 79% of the time, fashion questions just 13% to 23%. More near-identical listicles fighting for a handful of slots seems like the obvious explanation, though Ahrefs frames that as a likely cause rather than a proven one.

BrightEdge has published its own citation volatility analysis with a similar split between frequently and sporadically cited domains, though it measures week to week rather than day to day. Different methodology, same rough shape. I'd trust the shape more than any single number from either vendor, since both sell tools that track this stuff.

A 7-day persistence score you can set up this week

Here's what I'd change in AI visibility reporting, starting with the next report:

  • Replace the daily count with a persistence rate. For each tracked prompt, record days cited out of the last 7. Report the share of prompts at 5/7 or better. That's the number to trend, not raw citations on a given day.
  • Set a 14-day rule before touching a page. Ahrefs' own advice is to wait two to three weeks before reworking a page over a missing citation. If a citation held 8+ days and then vanished, the odds it returns the next day are 69% to 87%. Don't burn a content sprint on it.
  • Filter empty answers on Gemini and Copilot. Before logging a loss, check whether the answer cited anyone at all. In one tracked report, 68% of lost citations on those two tools happened in answers with no sources.
  • Track brand mentions as a separate column, and add every name variant you can think of (product names, abbreviations, the old company name).
  • Cross-reference with Google rankings. For prompts where you're cited but don't rank, expect churn. If you want those citations to hold on AI Overviews or AI Mode, the ranking gap is probably the first fix.

One caveat that keeps nagging at me: this is one month of data, and the AI tools changed underneath it. In one tracked report, the share of ChatGPT losses caused by answers that cited nobody jumped to 58% between August 19 and September 7, against 11% across the full window. Any baseline you build will drift, so I'd recalibrate it quarterly rather than treat it as settled. If you want a reminder of how strange these systems can get, our piece on AI search tools citing a fake concept within a week is a decent palate cleanser.

Where the 7-day number probably settles

My guess: by mid-2027, at least half of the AI visibility tools on the market will default to a rolling persistence metric instead of daily citation counts, because clients will notice their "losses" keep coming back. Until then, the tools will keep reporting the specials board as if it were the menu.

I don't think anyone needs to stop tracking AI citations. It's more that the people who look at a week of data before reacting will probably spend a lot less time chasing ghosts than the ones who check every morning.