Getting Cited by AI Is Faster. Staying Cited Is Harder.

For much of my career in search, one of the hardest parts of the job was waiting. You could substantially improve a page, clean up its structure, strengthen the evidence behind it, and know you’d made it better, then wait weeks or months to find out what Google thought of the change. The timing belonged to the search engine, not to you.

That was never a fixed rule. Google’s own guidance puts crawling anywhere from a few days to a few weeks, and says plainly that requesting a crawl doesn’t guarantee fast inclusion, or inclusion at all. Some changes showed up quickly. Plenty didn’t. But the operating assumption businesses learned was patience: make the page better, then wait to see if the market noticed.

AI-assisted search has introduced a different problem. The feedback can be much faster. So can the loss.

Quick to Arrive, Easy to Lose

• Traditional search trained businesses to expect slow feedback: publish or revise, then wait to see whether a ranking moved.
• Search-enabled AI can surface current evidence on a timetable very different from the ranking cycles businesses learned to expect.
• The same speed cuts both ways. Citations can also disappear, get replaced, or vary across systems and prompts surprisingly quickly.
• Getting cited and staying cited are different problems, and mistaking the first for the second is the real trap here.
• None of this is a case for publishing constantly. It’s a case for evidence that keeps deserving its place, not just earning it once.

AI Answers Can Run on a Different Clock

A search-enabled AI system operates differently from a traditional ranked results page. It can retrieve current web information while forming an answer, so evidence can potentially begin influencing an AI response without first establishing the kind of durable organic ranking businesses historically watched for. OpenAI’s own documentation describes ChatGPT retrieving current information from the web at the time a question is asked, separate from whatever the underlying model learned during training. That distinction matters more than it sounds like it should.

There are really three different things happening, and they’re worth keeping separate:

• What the model learned in training. Editing a webpage today doesn’t mean that information has become part of a model’s trained knowledge. That happens on a much slower, much less visible cycle, if it happens at all.
• What a system can retrieve right now. A search-enabled AI system can pull current information from the web while forming an answer, which means newly published or revised evidence can potentially influence that answer well before it would ever become part of a future training run.
• What actually gets cited. In OpenAI’s web-search implementation, not every source consulted necessarily appears as an inline citation. More generally, which sources surface can vary by prompt, by system, and by time.

The mistake I want to head off early: none of this means a page edited today gets “learned” by ChatGPT today. It means a search-enabled system may be able to retrieve and use that edit quickly, without the edit ever touching the model’s actual trained knowledge at all.

I’ve seen the speed side of this firsthand. New content we’ve published has started getting cited within a day of going live, faster than anything I saw in years of waiting on traditional rankings to move. That’s an observation from our own work, not a measured study, but it’s consistent with what the retrieval mechanism above would predict.

The New Temptation

Faster feedback creates a temptation that businesses trained on the old system don’t have good instincts for. A citation shows up, and it’s tempting to read that as a verdict: we won.

A single citation is an observation, not a verdict. It doesn’t necessarily mean stable authority, durable inclusion, or a result that looks the same tomorrow, on a different platform, or for a slightly different version of the same question. The faster AI visibility moves, the easier it becomes to confuse movement with progress.

Getting Cited and Staying Cited Are Different Problems

This is really the core of what’s changed, and it’s worth two separate terms rather than one blurry one.

Citation velocity is how quickly new or revised evidence is first observed appearing in tracked AI-generated answers.

Citation persistence is how consistently that evidence continues appearing across repeated observations once it has surfaced.

Presence tells you whether a citation appeared. Velocity tells you how quickly. Persistence tells you whether it keeps earning its place. None of this replaces the five diagnostic dimensions this series already uses to measure AI understanding, recognition, understanding, inclusion, citation, and recommendation. It adds a temporal question to one of them specifically: citation isn’t just present or absent. It moves, and the shape of that movement is itself information.

Persistence also depends on what you’re measuring. A specific page can disappear while another page from the same source takes its place. That’s different from the source itself dropping out of the citation set, and different again from the company disappearing from the answer altogether. URL-level churn, source-level persistence, and entity-level inclusion are not the same thing.

Traditional search encouraged a positional question: where do we rank? AI citation asks a different one: how often is our evidence selected for the questions that matter, and does that selection persist? That’s not a small shift in emphasis. It’s a different unit of visibility.

Search Had Inertia. AI Citations Can Have Churn.

Traditional search was still relatively easy to think about positionally: a page occupied a place in a ranked results environment, and once that position became reasonably stable, businesses learned to treat it as something with a degree of inertia.

AI citations are different. The same question can produce a different source set across systems, across time, and even across repeated executions. That instability isn’t only something visibility tools are reporting. A 2026 peer-reviewed comparison of Google’s traditional search against five generative search systems from Google, OpenAI, and Perplexity, published in Findings of the Association for Computational Linguistics, found substantial differences in source stability between them, and showed that generative-search outputs can vary across both time and repeated executions of the same query. The authors argue explicitly that stability is a dimension traditional search evaluation was never built to capture.

I’ve also watched this happen at page level. At one large, high-volume publisher we track, pages can gain significant citation visibility and then lose much of it within a week or two. High publishing volume may be contributing to internal competition between pages, but I can’t establish that from the outside. What I can observe is the result: citation gains that initially look significant don’t necessarily persist.

It’s also not an isolated impression. One well-documented example is worth walking through, including the part where even outside experts don’t fully agree on why it happened. Semrush tracked more than 230,000 prompts weekly for 13 weeks across ChatGPT Search, Google AI Mode, and Perplexity, and found Reddit falling from close to 60% of ChatGPT responses to around 10% in September 2025, with a comparable Wikipedia decline, while the other systems behaved differently. The shift was concentrated in ChatGPT rather than appearing uniformly across the platforms Semrush tracked.

The cause was disputed, which is itself instructive: a source can lose significant citation share without the publisher having a clear explanation for why. This is vendor-tracked data, not peer-reviewed research, and it covers one platform’s behavior over one window. But between that, what I’ve watched happen to individual publishers, and what the ACL research found more systematically, this is enough to know that the instability is real, even when its cause isn’t obvious.

Why a Single Snapshot Becomes Less Useful

If citation behavior can move like that, checking once tells you very little. A representative set of questions, tracked consistently over time, across more than one system, is what actually reveals a pattern rather than a moment. A single prompt on a single day is a snapshot. The faster the environment moves, the less useful a single screenshot becomes.

What Measurement Looks Like When Citations Move

The point isn’t to build another scorecard. A handful of diagnostic questions are more useful:

• Are we being cited at all, for the specific buyer questions that matter?
• How quickly does new or revised evidence begin surfacing once it’s published?
• Does that citation persist, or does it show up once and vanish?
• When it disappears, what’s replacing it, and why might that source be winning the comparison?
• Does the same pattern hold across systems, or is this a single-platform story?
• Is the underlying evidence changing, or is it retrieval and citation behavior shifting around evidence that hasn’t moved at all?

Those questions diagnose. A single number doesn’t.

The Evidence Layer Has to Stay Active

None of this is an argument for publishing constantly. The evidence layer this series has already described, credentials, case studies, third-party corroboration, has to do more than exist once. It has to remain accurate, consistent, and worth selecting, continuously, not just at the moment it was first published.

Outdated evidence can become less useful. Stronger or clearer competing evidence may begin surfacing instead. A page that contradicts something published elsewhere on the same site can weaken the whole picture’s attribution at any point, not just at launch.

Fast-moving citations don’t create a requirement for endless content. They create a requirement for evidence that continues to deserve selection.

The Harder Test

Search taught businesses to wait for rankings to move. AI may require something different: learning to tell whether a citation is merely appearing, or whether the evidence behind it has become strong enough to keep earning its place.

Getting into the answer can happen quickly. Staying useful enough to remain there is the harder test.