Why do my AI citations keep appearing and disappearing?

Every answer is written from scratch

A search engine hands back a stored ranking. An AI engine writes its reply word by word, with some randomness built in, so the same question can return a different list of brands each time. SparkToro and Gumshoe.ai measured this in January 2026. 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI Overviews 2,961 times. The odds that ChatGPT or Google's AI gave the same list of brands twice were less than 1 in 100. The odds of the same list in the same order were less than 1 in 1,000.

Strong brands still showed up reliably. When volunteers asked ChatGPT about cancer care hospitals, City of Hope appeared in 69 of 71 answers, though it was the first pick in only 25. The list shuffles, but a brand the engine trusts keeps coming back. That repeat appearance is what you want to track.

Your buyers don't word their questions the same way either. In a follow-up, the researchers collected 142 prompts written by people for one shared need: picking headphones for a family member who travels. The prompts scored just 0.081 on the semantic similarity measure the team used, so most pairs were only loosely related in wording. A check that tracks one exact phrasing is following a question few real buyers type.

The sources underneath change

Engines with live search fetch web pages for each answer, and the pages they pick keep changing. SISTRIX tracked 1.5 million weekly snapshots across 82,619 prompts from December 2025 to April 2026. About 74 percent of the domains ChatGPT Search cited in a given week hadn't been cited the week before. Google AI Mode came in at 56 percent and Google AI Overviews at about 5 percent. Counted page by page instead of site by site, ChatGPT's weekly turnover rose to 85 percent.

Share of cited domains that are new each week

• ChatGPT Search 74%
• Google AI Mode 56%
• Google AI Overviews 5%

You can see it inside a single day, too. SE Ranking ran 10,000 keywords through Google AI Mode three times on June 20, 2025. On average, only 9.2 percent of the cited URLs overlapped across the three runs, and about one keyword in five had no URL in common at all.

Google's AI Overviews are steadier, with a catch. Ahrefs found in November 2025 that their wording changes about every 2.15 days on average, and that 45.5 percent of the cited URLs differ from one check to the next. That fits with SISTRIX's 5 percent because SISTRIX counted sites, not pages. The individual pages change often. The sites behind them mostly don't.

The engines change too

AI companies release new models and adjust how their search features pick sources, often with little notice. When an engine switches models, who gets named can shift overnight for every brand at once. In your numbers that looks like a sudden drop, but your site had nothing to do with it.

Before you change anything, go through these five questions:

• Did it happen once, or does it keep happening? Ask the same question several times. If you show up in some answers and not others, you're near the edge of the list, which is normal. If you've dropped out of every run, look closer.
• Is it one week or several? One bad week on ChatGPT, where most cited domains change weekly anyway, says very little. A decline that lasts three or four weeks is a trend.
• Is it one engine or all of them? A drop on one engine usually points to a change at that engine. A drop on every engine at once points to something on your side.
• Can AI crawlers still reach your site? This is the most common cause of a real, sudden drop. A CDN bot-protection setting, a firewall rule or a robots.txt edit can block AI search crawlers with no visible error on your site. Our guide to allowing AI crawlers shows what to check.
• Has one competitor taken your place? If the same rival now shows up where you used to, week after week, that's a real shift in your category, not noise.

Treat a citation as a rate, not a spot you win. "We appeared in 7 of 10 answers this week" tells you something. "We were cited this morning" doesn't. The SparkToro researchers came to the same view: visibility percentage across many repeated runs is a reasonable metric, but a brand's position in any one list is not.

In practice that means asking each question more than once per engine, so one odd answer can't swing the result. Track each engine on its own, because they move at very different speeds. Then judge the trend over several weeks instead of reacting to the latest number. Measuring AI citations covers how to set that up. If you aren't showing up at all yet, start with how long it takes to show up in AI search, or check whether AI search mentions your brand by hand first.

I ran the Free Snapshot twice and got different results. Which one is right?

Both are real, and neither tells the whole story. Each run is one answer to one question from one engine, and AI answers vary from run to run. Getting two different results usually means your brand sits near the edge of the list for that question. What counts is how often you appear across many runs, which is what repeated weekly checks measure.

How many checks do I need before a result means something?

More than one, spread over time. One answer can't separate a real change from normal variation. Ask the same question several times per engine, then watch the rate over several weeks. Web Cited's paid plans ask each question 3 times per engine every week and show a 4-week average next to each reading for exactly this reason.

Do some AI engines change more than others?

Yes, by a wide margin. In SISTRIX's May 2026 study, about 74 percent of the domains ChatGPT Search cited in a given week hadn't been cited the week before. For Google AI Mode it was 56 percent, and for Google AI Overviews about 5 percent. A drop on ChatGPT is far more likely to be noise than a drop on AI Overviews.

Should I rewrite my page when a citation disappears?

Not after one check. First make sure the drop holds across repeated runs and more than one week, and that AI crawlers can still reach your site. If you rewrite a page because of normal variation, you may be changing something that was working.

The Free Snapshot asks Claude one buyer question about your category and shows whether you were named. It's a quick first look, and as this page explains, it's one sample. We built Citation Monitor to be the repeated version, and the SparkToro findings above are why it works the way it does.

It never reports where you landed in a single answer's list, because that spot changes on the next run. Instead it asks each of your buyer questions 3 times on each of six AI engines every week, then reports the share of those answers that named or cited you. On the Watch plan's 10 questions that is 180 answers a week, and on Command's 25 it is 450, when every engine responds. Every share comes with a 95 percent confidence range, and the trend chart adds a 4-week average, so one noisy week doesn't read as a trend. Because real buyers word the same need so differently, it's worth spending a few of those question slots on different phrasings of your most important one.

Try the Free Snapshot   Start Monitoring

By the Web Cited Editorial Research Team. Last updated 9 October 2026. Sources: SparkToro and Gumshoe.ai (January 2026), SISTRIX (May 2026), SE Ranking (August 2025), Ahrefs (November 2025).