Measurement & Diagnosis

When AI Citations Fall, What Have You Actually Lost?

A falling AI citation count can mean fewer answers, shifting questions or different pages cited. How to tell which, and what to investigate next.

A falling citation count in an AI visibility report can mean three different things. Fewer answers may have cited the business. The business may have gained ground on some questions and lost it on others. Or the AI system may simply have cited different or fewer pages from the same site. The headline number alone can’t tell these apart or show what to investigate next.

A 19% Drop, and What It Didn’t Say

For one client, we monitored the same 60 questions on Google across two separate seven-day periods. The number of URL citation records pointing to the client’s website fell from 278 to 224, a drop of about 19%.

On its own, that is the kind of number that gets a meeting called.

The decline was real. The export recorded fewer citations to pages from the site. But that count alone couldn’t say whether fewer answers cited the site, which questions gained or lost citations, or which pages were cited.

Count a different way and the picture changes. In the first week, 201 observed answers cited at least one page from the site. In the second, 202 did.

The difference is in what each number counts. A URL citation record is one URL from the site recorded as cited in one observed answer, so an answer that cites three of the site’s pages adds three records. Counting answers instead treats each question on each day once, however many of the site’s pages it cited. The number of answers citing the site was essentially unchanged. Among answers that cited the site, the average fell from about 1.4 pages per answer to 1.1.

The Steady Number Isn’t the Whole Story Either

It would be easy to stop there and conclude that nothing much changed. That would be the second misreading.

Across the 60 questions, 19 were cited on more observed days in the second week than the first. Eighteen were cited on fewer. Twenty-three were unchanged, but 11 of those were stable only in the sense that no citation of the site was observed for those questions in either week.

So 201 and 202 weren’t a sign of stillness. They were the net of movement in both directions.

Measure What it counts First week Second week
URL citation records Each URL from the site recorded as cited in each observed answer 278 224
Answers citing the site Each question on each day with at least one page from the site, counted once 201 202

The table shows how many citations to the site’s pages were recorded and how many observed answers cited the site. Neither count tells you which questions gained or lost citations. That takes a question-by-question comparison.

Same Presence, Different Page

The comparison above shows which questions the site was cited for on more or fewer days. To see what changed within a question, look at which pages were cited.

On a question about current conditions, the site was cited on five observed days in each week. In the first week, a general FAQ accounted for five citation records, and a page dedicated to the topic accounted for one. In the second week, no citation of the FAQ was observed, and the dedicated page accounted for all five.

The dedicated page is a more direct match for the question. That much is an observation about fit. It doesn’t tell us why Google chose it, and nothing in the data does.

What it does show is that one page from a site can drop out of the sources while another from the same site takes its place. Counted as records, that question went from six to five, a small loss. Counted as presence, it held. Looked at page by page, the source changed to the page built to answer the question.

A Better Match, Cited Less Often

On another question, the page cited became a better match, even as the site was cited less often.

In the first week, a general FAQ and other pages from the site were cited. In the second, a guide written specifically to answer the question accounted for all four observed appearances.

Over the same two weeks, the number of days on which the site was cited for that question fell from six to four.

Both are true. A page that answers a question more directly can become the page cited while the business appears less often for that question. Neither finding cancels the other. Either reading on its own tells only half the story. And neither example establishes why the engine chose a particular page, or whether the change helped the business.

A Competitor’s Gain Isn’t Automatically Your Loss

When a competitor’s citations rise in the same period, it’s natural to read that as the competitor taking your place. Check that first.

A competitor’s FAQ page appeared 11 times in the second week. In seven of those answers, our client was cited too. Those seven are co-appearances, not displacement. The other four show only that the competitor was cited where the client wasn’t observed. They don’t establish that the competitor caused a loss.

What Two Weeks Can and Can’t Show

One more detail matters before any of this is read as reassurance. In the records labelled Google, 343 observations included at least one citation in the first week, compared with 369 in the second. The number of observed answers citing the site barely changed, but the pool of observations with any recorded citation was larger in the second week. The site’s share of that pool fell from about 59% to 55%.

Those percentages describe the share of observations with at least one recorded citation, not a rate across all completed runs. The monitoring data can’t distinguish an answer that cited nothing from a run that didn’t happen, so a blank means no citation of the site was observed, not that it definitely wasn’t cited.

Two weeks on one engine is enough to show what a headline number can conceal. It isn’t enough to say where this business’s visibility is heading.

What to Investigate Instead

If your report gives you only a citation total, ask for a before-and-after breakdown using the same questions and the same AI engine. How many observed answers cited your site? Which questions gained or lost citations, and which pages were cited for the questions that matter most to your business? Ask how many runs completed in each period, too. Without that, you can compare recorded counts, but you can’t tell whether the rate changed across all runs.

If fewer observed answers cite the business, look at where it stopped appearing, and check whether the monitoring ran as fully in both periods before treating the gap as real.

If the number of answers is steady but individual questions moved, the aggregate is a net figure. Look at which questions gained and which lost, and how much each one matters to the business. A loss on a question central to the business deserves more attention than a gain on a peripheral one.

If the questions held but the cited pages changed, read the newly cited pages and ask what they actually answer. Sometimes the change is toward a better match. Sometimes it isn’t.

And if a competitor gained citations, find out first whether it appeared alongside the business or in answers where the business wasn’t cited.

“Citations down 19%” is a real observation. What the business actually lost, if anything, is a separate finding, and it is the only one that tells you what to do next.

Questions

Frequently asked questions

Why would AI citations fall if a business still appears in the same number of answers?

In this report, the citation count counts each cited page separately within an answer. One answer can contribute several citation records from the same site, so if answers cite fewer of its pages, the citation count falls even when the business is cited in about as many answers as before.

If a competitor's AI citations go up, has it replaced us?

Not necessarily. Check whether the competitor was cited in the same answers as your business. Where both appear, the competitor has joined the answer rather than displaced you. Where it appears without you, that shows a gap, but not that the competitor caused it.

Does a blank in AI monitoring data mean the business wasn't in the answer?

Not always. Some monitoring exports can't distinguish an answer that cited no sources from a run that didn't complete. A blank is better read as "no citation observed", and it's worth checking whether monitoring ran equally in the periods being compared.

The most important AI conversation about your business may be one you are not part of.

Start a conversation