AEO monitoring
Monitoring AI citations without a dashboard habit
Engine-by-engine citation watch as a conversation, not a tab you forget.
Key takeaways
- Losses are the citations nobody notices. Nothing on your site changes when a third-party answer changes its sources.
- Verify before you react. A dramatic swing is often the feed, and an hour of checking beats a month of sprinting at a phantom.
- Track position and concentration, not just the count. Ten citations from one prompt is a fragile ten.
- Read direction against your own history. Raw counts rise everywhere while these surfaces are still growing.
- Route losses to a cause and gains to a protect list, so the output is a queue rather than weather.
Nothing changed on your site. A page ChatGPT used to name as a source stopped appearing there, because the answer found somebody else's page more useful, and the only trace of it sat in a report nobody ran that week.
Losses are why this needs a routine rather than a dashboard. A dashboard is a tab you forget to open. The traffic tells you eventually. By then the answer has had a quarter to settle into its new sources, and the page that replaced you has had a quarter to look like the obvious choice.
So the teams who do this well run a recurring conversation instead. A citation is an AI answer that names one of your pages as a source, out of the questions you track. The conversation covers which pages earn them, where that changed, and whether a swing deserves attention, engine by engine.
This matters because citation data misbehaves in ways rank data never did. It swings hard, it differs by engine, and its loudest readings are frequently artifacts. A routine with the right checks inside it turns a jumpy metric into a workable one.
What the weekly check reads
Three cuts of the same events, plus a movement layer on top.
That layer is new and lost citations. Pages that started earning citations this month, and pages that quietly stopped. The losses are the ones nobody notices, because nothing on your site changed when the answer changed its mind.
The ai-citation-monitoring skill, a packaged analysis you run by name in your assistant, does all of it in one ask on whatever schedule you set. The three cuts are below.
- Which of your pages get cited, how often, and where they sit in the list.
- Which prompts cite you most, and which cite somebody else instead.
- How each engine behaves, because ChatGPT, Perplexity, Gemini and AI Mode cite differently and a blended line hides the trades.
Ask it yourself
Which of our pages do AI answers cite, what's new or lost since last month, and how does it split by engine?
Did the number move, or the data?
Ask that first, before anybody builds a story. A citation share drops to zero for a day on one engine, the room splits between panic and disbelief, and neither is productive.
It is verification. Did the answers that day genuinely cite other sources, or did the data feed hiccup?
In the case that taught us this, the zero was real. The engine's answers had shifted their sources for a day, then shifted back. The team that checked first spent an hour on it. In the other timeline, a content sprint launches against a phantom and costs a month.
That check is part of it, so dramatic swings get an is-the-data-sound pass before they get a narrative. The answer arrives with its as-of date, the date that data was last complete, so nobody diagnoses a lag as a loss.
See also: how the as-of date keeps a lag from reading as a loss →
Position matters as much as count
A citation at the top of a source list and one seventh down are different outcomes. So track position beside frequency.
A page whose citation count holds steady while its position slides is losing ground gradually, and that only shows up when position is part of what you read.
Concentration works the same way. Ten citations across ten prompts is broader footing than ten from one prompt that could be rephrased out of existence next month, and the drilldown views show both shapes. That is often the difference between a durable page and a lucky one.
Position also catches quiet substitution. Your own count can hold steady while the pages around you in the source list turn over completely, and when the company an answer keeps changes, its framing tends to change next. Read the whole source list, not just your own row.
The workflow that does this: ChatGPT citations by cited page →
On-page work shows up in the citations
Here is one sequence worth knowing. A site refreshed metadata and internal links on a set of program pages, and over the following weeks those descriptions started appearing inside ChatGPT's citations of the same pages.
Small, concrete, and visible only because somebody was watching the citation history the week it happened.
That is an observed sequence. Here is the honest way to report it. The routine shows you what changed and when, and the correlation between your work and the citations is a candidate to test and never a closed case.
What it removes is the blindness. Not the uncertainty. Just the change you never noticed because nobody was looking that week.
See also: why two things moving together is a candidate to test, not a finding →
Repetition turns anecdotes into a trend
A one-time citation check is a screenshot with a date on it. Repeated weekly it becomes a trend line with a denominator, and the denominator makes movement mean something.
The prompts replayed each week come from your prompt basket, generated from your own first-party demand. So the measurement covers the questions you actually compete for.
Trends inherit the market caveat too. These surfaces are still growing, so raw counts rise with adoption everywhere. Direction against your own history and position against competitors are the two readings that survive that tide.
Set it as a scheduled task and it arrives without anybody asking, which is the difference between a habit and an intention.
See also: where the prompts come from, and why hand-picked lists lie →
Losses route to diagnosis, gains to protection
The output is a short worklist, and it always routes the same way.
A lost citation gets a two-step check. Did the crawlers stop reaching the page, which your server logs answer directly, or did the answer find a better source, which is a content question with the rival's cited URL attached?
A gained citation goes on the protect list, joining the linking and refresh priorities so the win compounds instead of eroding quietly.
Either way the citation history stops being weather and becomes a queue. For the standings picture, who leads the category and by how much, the fuller answer engine optimization (AEO) audit covers competitors and coverage gaps. This is what you check in between, which is why audit day holds no surprises.
See also: checking whether AI crawlers still reach the page →
Frequently asked
- How often should this run?
- Weekly is the usual answer, because that is roughly the rhythm at which answers re-source themselves and slow enough that you are not reading noise. The point is that it runs on a schedule and arrives without anyone remembering to open it.
- Our citations went to zero on one engine overnight. What now?
- Check whether the data is sound before you do anything else. Look at whether the answers that day really did cite other sources, and look at the as-of date on the report. Real zeros happen and so do feed hiccups, and the two need completely different responses.
- Can we prove our content work caused a citation gain?
- Not from monitoring alone. What you get is a sequence: you changed something, and the citation history moved afterwards. That is a candidate worth testing with a proper controlled design, and treating it as proof is how teams end up defending a story the next quarter demolishes.