Every guide to generative engine optimization explains what to do. Almost none of them say when it pays off, which is the only question anyone actually asks before committing budget to it.
That gap existed for a good reason. Answering it properly needs something nobody had bothered to build: a set of pages published at a known moment and then checked against several answer engines every single day afterwards. Without that, every claimed timeline was somebody's impression of their own client work, and impressions are exactly the wrong instrument for a system that regenerates its answers every couple of days.
The measurement now exists. Eighty-one newly published pages on a single domain, queried daily for thirty days across Google's AI Mode and ChatGPT search. Here is what it found.
The gap at the start is larger than most people expect. Within 24 hours of publishing, 36% of the test pages were already being cited somewhere in Google AI Mode. ChatGPT search had picked up 10% in the same window, more than three times slower off the line. By day seven, Google was at 56% and ChatGPT at 17%.
Then the curves change character, and this is the part that matters. Google's coverage fluctuated week to week, rising toward a peak of 59% and falling back from it repeatedly. ChatGPT's climbed steadily and did not give ground: 35% at two weeks, 42% at thirty days. Google is fast and noisy. ChatGPT is slow and cumulative.
That has a direct consequence for how you report on this work, and it is where people get burned. A project reviewed at four weeks looks like a Google success and a ChatGPT failure, and both readings are artefacts of the review date rather than facts about the work. The correct checkpoint is somewhere past six weeks, and the correct framing is two separate curves rather than an average that describes neither.
Perplexity sits outside the comparison for a structural reason. A significant share of its crawling is triggered by a user's question rather than by a schedule, so on a narrow, specific prompt it can be the fastest of the three, and on a broad one it can lag both.
The first week is not a content problem, and treating it as one wastes it. OpenAI runs three separate bots and the distinction matters: OAI-SearchBot surfaces sites in ChatGPT's search features, GPTBot crawls for model training, and both obey robots.txt, while ChatGPT-User fetches a page live because a person asked a question and is explicitly not bound by the same rules. Perplexity splits the same way, with PerplexityBot building the index under robots.txt and Perplexity-User fetching live during a question.
Both operators quote roughly the same latency on the control side, about 24 hours from a robots.txt update for their systems to adjust. That is a small number with a sharp edge, because an accidental block costs a day to undo after somebody notices. And the most common cause is not robots.txt at all. It is a bot-protection rule in a CDN that nobody remembers switching on.
Google is a different case entirely, and simpler. Its own documentation states there are no additional requirements to appear in AI Overviews or AI Mode and no special markup necessary, but that to be used as a supporting link a page must be indexed and eligible to be shown in Google Search with a snippet. The AI clock and the classic indexing clock are the same clock. There is no separate AI index to get into, and if the page is not indexed, no amount of GEO work reaches it.
So before blaming your content, check whether the crawlers are in your server logs at all, whether a firewall rule is turning them away, whether the page is indexed, and whether it renders anything useful without JavaScript. A crawler that receives an empty shell has technically fetched you and learned nothing. Those four cover most cases of "we are not being cited anywhere".
The number worth carrying out of this is not a speed, it is a limit. Over a full month, Google AI Mode peaked at 59% of the test pages and ChatGPT search reached 42%. That was on a domain with strong established authority publishing squarely inside its own subject area. Even under those conditions, roughly half of everything published was never cited anywhere.
This is the honest correction to how GEO is usually sold. The pitch implies that structuring content correctly makes it citable. The data says structuring content correctly makes it eligible, and selection is a separate, competitive step you do not control.
Plan on that basis and the economics stay sane. A page that fails to get cited is not a failed page, it is the expected outcome for about half of them, and the ones that do get cited tend to keep the citation. The practical consequence is that publishing a small number of genuinely specific answer pages beats publishing many broad ones, because the ceiling is per-page and broad pages lose to bigger entities every time.
This is the least intuitive finding in the area and it should change how you measure. Ahrefs tracked more than 43,000 keywords with at least sixteen recorded AI Overviews each and found an average persistence of 2.15 days, meaning an AI Overview has roughly a 70% chance of being different from one observation to the next. More usefully, only 54.5% of the cited URLs overlap between consecutive runs of the same query. Close to half the source list is replaced every time.
What does not change is the meaning. The same query re-run produces answers with an average cosine similarity of 0.95, and 54% of the named entities stay put. The system is confident about what it thinks and casual about who it credits, which means a citation appearing or disappearing on any given day carries far less signal than it feels like it does.
The rule follows immediately. Never measure AI visibility with a single check. Run a fixed prompt list at least ten times across several days and report the share of runs in which you appear, and freeze that list early, because changing it mid-measurement destroys the comparison and is how most in-house tracking quietly dies. A binary "are we cited" screenshot is roughly a coin flip, and I have watched more than one agency relationship turn on one taken at the wrong moment.
For most of the AI-search era the reassuring answer to "how do we get cited" was "rank well, the rest follows". In July 2025 that held up, when Ahrefs analysed 1.9 million citations from a million AI Overviews and found 76.1% of cited pages ranked in Google's top 10.