Diagnosing an AI Citation Drop: One Brand, Five Months, Prompt by Prompt
The headline number bounced around and told me almost nothing. Going prompt by prompt, platform by platform, told me exactly what to fix.
We don't stop at the number. We go prompt by prompt, platform by platform, until we know exactly which question moved, on exactly which engine, and why.
Last time, I looked at every client in our monthly AI citation tracking folded together, nearly 10,000 citations across ChatGPT, Gemini, and Perplexity. This time I'm doing the opposite: zooming into one brand, five monthly check-ins, and watching what actually happened to it.
I'm keeping the brand and its competitors unnamed here. It's a financial data and underwriting AI company, the kind of buyer questions it tracks are things like "bank statement analysis software India" and "how to reduce NPA in NBFC." Everything else is the real numbers.
The Headline Number Bounced Around and Told Me Almost Nothing
Share of voice, the percentage of tracked questions where the brand got mentioned by name, swung between 17.6% and 27.1% across five monthly runs with no clean upward or downward trend.
Share of voice is the number everyone asks for first. Here is what it did over five runs:
| Run | Share of voice |
|---|---|
| April 2026 | 25.4% |
| May 2026 | 19.4% |
| June 2026 | 22.2% |
| July 2026 | 17.6% |
| August 2026 | 27.1% |
Read that as a single line and it looks like noise. Down, up, down, up. If I had only checked in twice, say May and July, I would have told this client their visibility was fading. If I had checked in April and August instead, I would have called it a steady win. Both stories use real numbers and both would have been wrong, because neither captures what the run-to-run swing actually was.
The Number That Actually Moved in One Direction
The share of all citations pointing to the brand's own site climbed from 4.9% to 12.3% over the same five months, more than doubling, almost in a straight line, while the total volume of citations handed out actually fell slightly.
While I had the raw citation data open, I checked something I hadn't looked at before: not just whether the brand got mentioned by name, but what share of all the citation links in that month's responses pointed to the brand's own website, as opposed to a competitor, a review site, or anything else.
| Run | Share of all citations pointing to the brand's own site |
|---|---|
| April 2026 | 4.9% |
| May 2026 | 5.1% |
| June 2026 | 7.0% |
| July 2026 | 10.7% |
| August 2026 | 12.3% |
That is a climb from under 5% to over 12%, more than doubling, and it moved in a straight line the entire time while the mention-rate number above was bouncing all over the place. The total number of citation links being handed out each month actually fell slightly over the same period, from 485 in April to 431 in August, so this was not a case of the brand simply getting swept up in more overall traffic. It was capturing a growing share of a shrinking pool.
If I had to pick one metric to trust for this brand, it would be this one, not the mention-rate headline. Getting named by an AI engine in passing is a weaker signal than actually being one of the pages it pulls from.
The August Peak Came From One Engine, Not Three
The August share-of-voice peak was driven almost entirely by Gemini, which jumped to 40%. ChatGPT actually got worse in the same month, dropping to its second-lowest point of the whole period.
Here is where this connects back to what I found across the whole client base last time: ChatGPT, Gemini, and Perplexity mostly don't agree with each other on what to cite. This brand's own numbers are a clean example of exactly that.
| Run | ChatGPT | Gemini | Perplexity |
|---|---|---|---|
| April 2026 | 23.8% | 28.6% | 23.8% |
| May 2026 | 19.0% | 19.0% | 20.0% |
| June 2026 | 9.5% | 23.8% | 33.3% |
| July 2026 | 16.0% | 25.0% | 10.5% |
| August 2026 | 14.3% | 40.0% | 25.0% |
The overall August number, 27.1%, its best run yet, reads like a broad win. It wasn't. Gemini alone jumped to 40%, its highest point across all five runs. Perplexity climbed back to 25%. ChatGPT actually got worse, dropping to 14.3%, its second-lowest run of the whole period. One engine drove the entire headline improvement while another was moving in the opposite direction in the same month.
A report that only shows the combined number would have told this brand "AEO is working, keep going." The truer sentence is "AEO is working on Gemini specifically, and something is actively getting worse on ChatGPT," which is a different conversation and a different fix.
We Didn't Stop at the ChatGPT Number. We Went Prompt by Prompt.
A number like "ChatGPT fell from 23.8% to 14.3%" is where most visibility reports stop. We don't stop there, because that number alone doesn't tell you what to fix. So we pulled all 21 tracked prompts for ChatGPT specifically, compared the April run against the August run question by question, and looked at exactly which sources ChatGPT cited on each one.
The first thing we found: the prompt list itself had changed. Four of the 21 questions tracked in April were swapped out by August and replaced with four new questions about a related but different topic, tax filing tools for accountants. Two of the four dropped questions were ones the brand had occasionally been cited on. That swap alone accounts for a real chunk of the raw percentage drop, and it has nothing to do with ChatGPT turning against the brand. It's a denominator change.
On those four new questions, we checked all three engines, not just ChatGPT. The brand was mentioned on one of the four, by Gemini and Perplexity, not by ChatGPT. On the other three, none of the three engines cited the brand at all. That's not a ChatGPT problem, that's a genuine content gap on a topic the brand hasn't built anything for yet, and it shows up the same way across every engine we track.
That left one real, comparable data point: the single non-branded question that existed in the tracked set in both April and August, "best bank statement analysis software in India." On that one, ChatGPT's answer went yes, yes, no, no, yes across the five runs. We pulled the actual source list ChatGPT cited each time. The competitor sites showing up in the "no" runs were the same ones already appearing in the "yes" runs, they weren't new arrivals that pushed the brand out. The brand fell out of an already-crowded citation list for two runs, then fell back in. Nothing about the competitive set itself changed.
Two separate, specific things, not one vague trend. A genuine content gap on one new topic area that affects all three engines equally. And ordinary month-to-month volatility on a single contested question within an already-stable competitive field. Neither is fixed by "improving ChatGPT visibility" in the abstract. One needs new content built for a topic that doesn't exist yet. The other needs nothing at all, since the brand already wins that question more often than not.
The Competitive Field Didn't Shrink or Grow, It Just Stayed Crowded
One more number worth having, since it puts the above in context: the number of distinct competitor domains cited stayed remarkably stable across every full run, between 40 and 54 each time. This brand isn't gaining ground because rivals dropped out, and it isn't losing ground because new ones showed up. It's the same crowded field every month, the same one we found sitting underneath that single ChatGPT prompt above. Whatever moved, moved within that fixed competitive set, not because the set itself changed.
What This Actually Is
This is the point of tracking AI citations every month, and it's the reason this data exists at all. Not to produce one visibility score and either celebrate or panic over it. When a number moves, we don't stop at the number. We go prompt by prompt, and platform by platform, until we know exactly which question moved, on exactly which engine, and why.
Sometimes that means finding a whole topic area that needs content built from scratch. Sometimes it means finding out nothing really changed and the brand just needs to sit through a noisy run or two. Those are different problems with different fixes, and a single aggregate score can't tell them apart. Going prompt by prompt, platform by platform, is the only way we've found to, and it's how we run every client engagement Magnent takes on.
Frequently Asked Questions
Why does an AI visibility share-of-voice number swing so much month to month? Mention rate is a noisy metric because it depends on which specific questions get asked and how each AI engine happens to answer them that month. In this five-month case study, one brand's share of voice swung between 17.6% and 27.1% with no clean trend, while a steadier metric underneath it, the share of citations pointing to the brand's own site, climbed almost in a straight line over the same period.
What is the difference between AI mention rate and AI citation share? Mention rate measures whether an AI engine's written answer names the brand anywhere in the text. Citation share measures whether the brand's own website appears among the sources the AI actually pulled from. A brand can be mentioned without its site being cited, or cited among several sources without being the headline recommendation, so the two numbers can move in different directions in the same month.
Why would one AI engine's citation rate for a brand drop while others improve? Each engine draws from a different, mostly non-overlapping set of sources, so a drop on one engine does not necessarily mean the brand lost ground everywhere. In this case, investigating a ChatGPT-specific drop prompt by prompt found two distinct causes: part of the tracked question set had changed to include a new topic the brand had no content for yet, a gap that showed up on every engine, not just ChatGPT, and the rest was ordinary volatility on a single contested question within an already-stable set of competitors.
How do you diagnose why a brand's AI citation rate fell? By checking every tracked prompt individually rather than trusting the aggregate score. That means comparing which questions were asked in each period, checking whether the prompt set itself changed, and pulling the actual source URLs each engine cited on every question where the brand's presence changed, run by run and platform by platform.
Does a lower AI citation score always mean a brand needs more content? No. A falling score can come from several different causes, a genuine content gap on a new topic, ordinary month-to-month volatility on a contested question, or a change in the tracked prompt set itself, and each has a different fix. Only checking the underlying prompts and source citations can tell which one is actually happening.