AI Citation Stability: How Often Do AI Answers Change?

AI citation stability is low from one day to the next: in Ahrefs, October 2026, at least half of citations had disappeared by the next day on every tool tested, and about 75% of the dropped ones came back within 20 days. A single check tells you almost nothing. A repeated, fixed sample does.

Key takeaways

• Never judge AI visibility from one run. Count how many of the last seven days a page or brand appeared, per prompt.
• Track brand mentions and citations as separate metrics. Mentions survive far more often than links.
• Report a rate with its sample size and margin of error. Call a change real only when it sits outside that margin and holds for a week or more.
• Split prompts by type (your own facts, reviews and comparisons, "best X"). Each group has a different normal level of churn.
• Do not credit a page edit for one comeback. Most dropped citations return on their own.

What did Ahrefs find about how stable AI citations are?

Ahrefs followed about 37,500 citations across five AI tools for 20 days, asking the same questions daily and tracking every website cited on day one. The study was published on 29 October 2026 by Mateusz Makosiewicz and Xibeijia Guan. Here is the core table, reproduced from Ahrefs, October 2026.

Gemini and Copilot look worse partly because their answers often cited nobody: 64% of Gemini answers and 53% of Copilot answers had no citations. Skip those empty answers and all five tools looked alike, with 50 to 60% of citations disappearing by the next answer that included citations.

"Don't judge AI visibility from one check." Ahrefs, October 2026

Three other findings matter for measurement:

• Most losses reverse. Roughly 75% of citations that disappeared at least once returned during the study, and about half returned within three days.
• Age predicts return. Of citations that lasted one day, 34% (AI Overviews), 41% (AI Mode) and 35% (ChatGPT) never came back. For those that lasted 8 to 14 days, the figures were 4%, 2% and 0.7%.
• Brand names outlast links. On every tool, about 50 to 70% of cited websites appeared on just one day, against about one in five brand names.

Is this a one-off result, or do other studies agree?

Other published datasets point the same way, with different methods. That is useful: it means the pattern is not an artefact of one prompt set.

Read together: the exact list is close to random, the leading brands are not. In Detailed's data, the brand a platform named most often for a prompt appeared on 82% of days on ChatGPT and 89% on Google AI Overviews. So the right unit of measurement is a rate over many runs, not a snapshot. Our post on why AI engines disagree about your brand covers the cross-engine side of the same problem.

How do I measure my brand's AI visibility when answers keep changing?

Treat each tracked prompt as an experiment you repeat. This five-step method uses only a fixed prompt set, a schedule and a spreadsheet.

1. Freeze the prompt set. Write 20 to 50 unbranded buyer prompts, version them and do not edit them mid-test. If you need ideas, try the free Noma AI Prompt Finder, which suggests questions buyers ask about your category. Our guide to which prompts to track explains the selection.
2. Tag prompts by type. Ahrefs found the brand's own site was cited on 97 to 100% of days for factual questions about its own pricing on AI Overviews, AI Mode and ChatGPT, but only 5% (AI Overviews and AI Mode) for reviews and 0 to 5% for comparisons. That came from 22 questions about one brand, so treat it as direction, not a rule. Still, it shows why one blended number hides a strong group and a weak one.
3. Run every prompt daily on each engine. Record per run: brand named (yes or no), position, URL cited (yes or no), and whether the answer cited anyone at all. The last field stops you counting an empty Gemini answer as a loss.
4. Normalise brand names. Ahrefs grouped variants such as "Adobe Acrobat Online" and "Adobe Acrobat", and Detailed's manual check of 55 brands found 30 to 36% of apparent losses were the same brand under a different name. Keep an alias list.
5. Aggregate over a window. Report mention rate and citation rate over 7 and 28 days, per engine and per prompt type, never from one day.

How many runs do I need before a change counts?

The runs you sample set how wide your uncertainty is. For a yes-or-no outcome such as "brand named", the 95% margin of error is roughly 1.96 times the square root of p(1-p)/n, where p is the observed rate and n the number of runs. This is a standard approximation, and it assumes runs are independent, which repeated runs of one prompt only roughly are. The figures below are our own arithmetic for a brand named in 60% of runs.

A move from 40% to 55% mention rate is a 15-point change. Using the usual two-proportion sample size formula at 80% power, you need about 175 runs in each period to detect it. With 30 runs a period, only swings of roughly 35 points or more are distinguishable from noise. That is why a 5-point wobble on a small prompt set should never reach a slide.

A decision table for separating noise from change

One more rule from the same study: do not credit an edit for a single comeback. Because most lost citations return, a page you changed on Monday and saw cited on Thursday may simply have returned on its own. Look for a sustained rise in how often you are cited.

How should I report volatility to stakeholders?

Stakeholders read a falling number as failure. Pre-empt that by reporting stability as part of the metric. Our guide on reporting AEO performance to a CFO covers the finance angle; the reporting layout below fits any audience.

• Lead with the rate and the denominator. "Named in 58% of 300 runs (28 days, 4 engines), plus or minus 6 points."
• Split core and tail. Detailed defined core brands as those named on at least 80% of days. Report how many prompts you sit in the core for, and how many in the tail.
• Show mentions and citations side by side. Ahrefs suggests citations may be a weak KPI and mentions a better candidate because they are steadier.
• Show prompt types separately. Facts about your brand, reviews, comparisons and "best X" each get a line. SISTRIX's May 2026 study found brand domains anchored while co-cited sites rotated at 70% a week.
• Annotate known events. Page launches, model releases, robots.txt changes. See what a new model release changes for visibility.
• Name the trend test. "We call a change real when it leaves the margin and holds for 7 days."

Share-of-voice figures need the same discipline; how AI share of voice is calculated shows the formula, and the AI Visibility Measurement Standard covers measuring AI visibility with fixed prompts and repeated runs.

What should I check this week?

• Pick your 10 most important prompts and check each by hand on three engines, three times on different days. Note how often you are named and how often linked.
• Confirm AI crawlers can reach your pages. OpenAI's crawler documentation says OAI-SearchBot surfaces websites in ChatGPT's search features. The free Noma AI Crawler Checker shows whether AI crawlers can reach your site.
• Look at which of your pages rank in Google for your target questions. Ahrefs found that ranking pages were about seven times more likely to keep citations for a week on AI Overviews and AI Mode.
• Update stale pricing and comparison pages. Ahrefs reported two cases where freshness changed citations, though it labels these observations outside the study.

Where does Noma fit?

Noma by Nurdd is an AEO (answer engine optimisation) and AI visibility platform that tracks your brand across 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude. That gives you one place to repeat the same buyer prompts across engines, which is what the method above needs.

See it on your own brand: book a demo or start the 3-day free trial. A card is taken at signup so billing can start on day 3; cancel before then and nothing is charged.

What are the limits of this evidence?

• Ahrefs says its figures may underestimate returns, because citations that dropped late in the study had less time to come back, and it tracked only the first disappearance.
• The brand-fact, review and comparison split rests on 22 questions about one brand.
• The ranking analysis covered only 230 questions, and Ahrefs describes the results as associations, not causes.
• Studies differ in prompts, countries and definitions (page versus domain, daily versus weekly), so compare patterns, not exact percentages.
• The margin-of-error and sample-size figures above are our arithmetic, not published findings, and rely on independence assumptions.

How we researched this

We read the Ahrefs study in full, then reviewed the SISTRIX and Detailed studies, a Search Engine Journal report on SparkToro's research, Profound's earlier AI search volatility post and OpenAI's crawler documentation. We did not run our own experiment. Statistics, dates and product facts were checked on 2026-10-30.

Next step

Freeze a prompt set this week, run it repeatedly, and report rates with their margins. To track your brand across 13 AI engines, book a demo or start the free trial.