Ahrefs: Up to 83% of Links in AI Answers Vanish Within a Day. Here's How to Measure Visibility

A familiar scene: yesterday ChatGPT linked to a client's site, today the link is gone. You want to open an audit right away, rewrite the copy and call the SEO contractor. Don't. Most likely, the problem isn't on the site but in the way you measure.

A Disappearing Link Is Normal Behavior

For 20 days, Ahrefs tracked roughly 37 500 citations across five AI services every day. A citation here means a link to the source site beneath an answer. The result: by the next day, between 50 and 83% of the links disappear. About three quarters of the vanished links do come back within 20 days.

The reason is that ChatGPT builds the answer from scratch for every question. When it searches the web, it rewrites the user's question into one or more search queries, and the set of pages changes. Ahrefs found that ChatGPT repeated exactly the same searches as the day before in only 26% of cases. Add account memory, region, language and device, and two people will get different results. Still, by Ahrefs' estimate, changed queries explain only part of the disappearances.

To rule out a problem on your side, check three things: site availability, recent edits, and pages blocked from search crawlers. If all of that is clean, you're looking at noise.

Count a Share, Not a Fact

The metric is simple. A question is asked n times, the site is cited in k answers, and visibility equals k / n. Peec AI calculates it the same way, but for brand mentions. The problem is the small sample: 3 answers out of 10 give 30%, but the confidence interval (the range in which the true share lies with 95% certainty) is 11–60%. The author of the piece calculates it with the Wilson formula: with a small n and a share near 0 or 100%, the ordinary formula spills outside the 0–100% range. In Python, it takes a few lines with a square root.

The more answers, the narrower the interval. With the same 30% share, it looks like this:

• 10 answers: 11–60%
• 20 answers: 15–52%
• 40 answers: 18–45%

Hence the recommendation: 10 repeats of the question per measurement, one measurement per week, and a window of the four latest measurements, which is 40 answers. Track two shares: links to the site and brand mentions. In Ahrefs' data, for Google's tools and ChatGPT the link disappeared while the mention stayed in 12–16% of cases; in 9–12% it was the other way around.

Keep the measurement conditions unchanged: one account with no saved memory, one region, language and device, and one wording of the question. Otherwise there is nothing to compare.

When Is It a Failure, and When Do You Wait?

The thresholds are as follows:

• Failure signal: two consecutive measurements. Take the latest 20 answers as a single share. If the upper bound of its interval is below the lower bound from the four earlier measurements (40 answers), look for the cause on the site.
• Loss: three weeks without a return.

Example: the share used to be 30% (18–45%). Two consecutive measurements gave 0 out of 20, that is 0–16%, which is below 18%, and that is a signal. But 2 out of 20 gives 10% with an interval of 3–30%, which overlaps the earlier one, so there is no reason to panic.

Why wait exactly 20 days? It all depends on the age of the citation. A link that had lasted 8 days or more came back in 96–99% of cases. New citations are fragile: about 70% disappeared the next day, and only 34–45% of them came back. So a first appearance after a single measurement is not yet a success.

What to Show the Client

Show the share over the last four measurements with its interval, plus the month-by-month trend. For example: the site appears in 30% of answers (18–45%); a month ago it was 20% (11–35%). The intervals overlap, so the growth may still be random, and it's better to say so up front. On the other hand, the answer to "why weren't we there last week?" is ready: one week accounts for a quarter of that figure.

Show the brand's position in the list separately. SparkToro found that City of Hope hospital appeared in 69 of 71 ChatGPT answers, or 97%, but was named first in only 25 of them.

Limitations

The Wilson interval assumes independent repeats, and repeats from one account under the same conditions are unlikely to be independent. The true margin of error may be larger. For Gemini and Copilot, Ahrefs itself calls the 20-day comparison unreliable: the citations were present on all 20 days for only 4% of Gemini's questions and 27% of Copilot's.

The practical takeaway: a week without a link is a reason to wait for the next measurement, not to rebuild the site. Set up a spreadsheet with k and n for each question, record the conditions, and make decisions only from the intervals.