How Far Ahead Do You Have to Publish to Get Cited by AI?

The Assumption This Corrects

There is a number underneath every question about when to start answer engine optimization, and almost nobody quotes it. It is the median age of the material an engine actually cites.

The working assumption across most of this field, including in our own writing on why memory rather than content is the moat in AI search, is that live retrieval is current to the minute. Publish something this morning, the reasoning goes, and an assistant can quote it this afternoon. That is true of what retrieval CAN do. It is not true of what these systems were measured doing.

A 2026 comparative study of web search against generative answer generation recorded the median age of the sources each engine actually cited, broken out by engine and by category. The answer is not one number. In consumer electronics it runs from 62.3 days on Claude to 130.4 days on Google Search. In automotive the same measurement runs from 148.0 days to 492.9. Sixteen months, in the category where the purchase is largest.

That gap is not a curiosity about how retrieval works. It is a deadline. If the material an engine reaches for is typically months or years old, then the answer served in a given quarter was decided by work that had to exist well before it. The question of when to start stops being a matter of appetite at that point. It becomes a subtraction.

The distinction matters because the two readings produce opposite plans. If retrieval is effectively instant, the correct move is to publish close to the moment demand appears, which favours a reactive calendar with a small standing team. If the material being cited is typically months old, the correct move is the reverse, because the work has to be in place long before the demand it serves. Those are not adjacent strategies. They are opposite ones, funded differently, staffed differently, on different review cycles. A company that picks the first because nobody measured the second will be reacting to a market that stopped accepting new entrants for that quarter some time ago.

There Is Not One Lead Time, There Are Eight

The single most common mistake in planning for AI visibility is to ask how long it takes in general. The measurement says the question has no general answer. It has eight answers across four engines and two categories, differing by roughly a factor of eight.

Read the corners. A consumer electronics brand aiming at Claude is working against a median of about two months. An automotive brand aiming at Google Search is working against a median of about sixteen. Those are not variations on a theme. They are different businesses with different planning horizons, and averaging them produces a number that describes neither.

The blended average is the specific error worth naming. Take the mean of those eight figures and you get roughly 180 days, a number that would send the electronics brand into a panic it does not need while giving the automotive brand a deadline it has already missed by nearly a year. A planning input that is wrong in opposite directions for two readers is not a planning input. It is noise with a decimal point.

There is a second reading of the same table that is easy to miss. The spread across engines within one category is roughly two to one. The spread across categories within one engine is closer to four to one. Your category is therefore a stronger determinant of your deadline than your choice of engine, which inverts how most teams approach the problem. The usual question is which assistant to optimise for. The measurement says the more consequential variable was settled the day you chose what business to be in, leaving the engine question as a refinement on top of a horizon you do not control.

The Second Clock, in the Engines' Own Words

Age of cited material is the first clock. The second is how long it takes an engine to notice that something exists at all, and here the engine owners publish their own numbers.

Google states that a recrawl after a page is submitted or updated takes a few days to a few weeks, and that resubmitting the same URL buys no speed. OpenAI states that a publisher-side access change reaches its systems in roughly 24 hours. These are the cheapest possible changes, a single directive or a single page, and they still cost real time.

Model training cutoffs form a third layer belonging to a different argument that we have made elsewhere. This piece treats them as out of scope, taking the retrieval path as given. The published cutoffs are a matter of record for anyone who wants them. What matters for scheduling is that the second clock is small next to the first. Recrawl is measured in weeks. Cited-source age is measured in months to years. The deadline is set by the larger of the two, which means it is set by the age of what gets cited.

It is worth being careful about how the two clocks combine, because they do not simply add. The re-fetch window governs when a page becomes eligible to be considered at all. The cited-source age describes what tends to get chosen once a large pool is already eligible. A page can be indexed inside a week while still waiting months before it is the thing an engine reaches for, which is exactly why the smaller number cannot be treated as the answer. Eligibility is a gate. Age is the queue behind it, and nobody clears a queue by arriving at the door faster.

The Publish-By Arithmetic

The DSF Publish-By Map is a subtraction with five steps, and the output is a single date. Fix the quarter you need to be present in, tied to a real commercial event rather than a budget cycle. Read the median cited-source age off the matrix for your engine and your category. Add the engine's stated re-fetch window. Subtract both from the target. Compare the result to today.

Run it twice and the shape of the problem appears. An automotive brand that wants to be present in Google Search answers in autumn 2027 subtracts 492.9 days, then a few more weeks for the recrawl, and lands in spring 2026. A consumer electronics brand aiming at Claude for the same quarter subtracts 62.3 days and lands in the summer of 2027. Same target, same arithmetic, two deadlines fourteen months apart.

Most companies do not have a single target. They have one category, which fixes a row of the matrix, plus a set of engines their buyers genuinely use, which selects several columns. The practical resolution is to run the subtraction against the LONGEST of the applicable medians rather than the mean of them.

A schedule built on the shortest deadline arrives on time for one surface while arriving late everywhere else, which is the worst of both plans: the cost of starting early without the coverage that justified it. This is ordinary critical-path discipline. The date that governs a project is the one furthest out, and treating that date as pessimism rather than as the schedule is how delivery slips.

The uncomfortable half of that example is the automotive line. A company deciding this month whether to fund the work is not deciding about autumn 2027 at all. That window closed before the question was asked. It is deciding about 2028, and the only thing still available for 2027 is whatever the shorter-horizon engines will accept. Knowing which of those two situations you are in is the entire practical value of running the subtraction. If this is the point at which the calendar stops being an abstraction, that is what Answer Engine Optimization exists to do.

Why the Spread Is Widest Where the Purchase Is Biggest

The pattern in the matrix is not random. The category with the oldest cited material is the one where the purchase is largest, the research most deliberate, and the buying group biggest. The category with the newest is the one where a specification changes every season.

The study's own finding is that answer engines return newer cited material than Google on the median, with the gap widening in automotive. That fits how those decisions get made. Gartner found that 45% of business buyers used generative AI during a recent purchase, working across an average of seven information sources. A decision assembled from seven sources over months does not reward whatever was published this week. It reaches for what has been sitting there long enough to be corroborated.

Why an established source wins that contest is a separate mechanism, and we have written it up in full under stale-source bias. For scheduling purposes it is enough to take it as given. The point here is narrower and purely practical. The heavier the purchase, the further back the deadline sits, which is precisely the opposite of how most companies sequence their marketing investment.

That inversion carries a budgeting consequence worth stating plainly. Marketing plans are usually sequenced by expected return, with the largest categories funded first because they pay back most. The matrix says those same categories carry the longest lead time, so funding them first within a calendar year still leaves them latest to arrive in an answer. Sequencing by return produces one order. Sequencing by deadline produces another. Only the second one is answerable by arithmetic, which is a strong argument for letting it go first.

The categories with the longest lead time are the ones where the deal is biggest. The deadline moves furthest away exactly where the money is.

What the Calendar Says About Starting in September

Set against the demand side, the arithmetic gets sharper rather than softer. The surfaces this material feeds are not waiting politely for anyone's planning cycle.

Google reports AI Mode past one billion monthly active users with queries more than doubling every quarter since launch. Salesforce measured agentic search as the first step in the shopping journey growing 200% year over year. Neither of those is an argument that a window is closing, which is a claim this article deliberately does not make. They are the scale against which a sixteen-month deadline should be read.

The CMO Survey found four in ten companies already running generative engine optimization, a capability absent from every previous edition of the survey. Read that next to the calendar rather than as a scare. It tells you the material competing for autumn 2027 answers in your category is being published now, which is the only reason the subtraction has any urgency at all.

There is a way to misread those figures that is worth heading off. Rapid growth in assistant usage does not shorten anybody's deadline. If anything it lengthens the queue, because more material competes for the same citation slots inside a corpus that already skews old. The growth numbers are not themselves the reason to move. The reason, where there is one, is arithmetic: a target quarter minus a measured median leaves a date, and dates do not negotiate. Everything else in this section is context for how much that date is worth.

What the Calendar Cannot Tell You

A planning tool that overstates its own precision is worse than no tool, so it is worth being exact about what this arithmetic does not establish.

No published source measures the elapsed interval from publication to first appearance in an answer set. Every latency figure here is either an engine's stated recrawl window or the measured age of material already being cited. The publish-by date is therefore a planning floor, derived from two adjacent measurements rather than observed directly, and it should be treated as the earliest sensible date rather than a guarantee.

Age is also a precondition rather than a cause. Old material that nobody corroborates does not get cited for being old. The matrix tells you when the work has to exist. It does not tell you whether the work is any good, and publishing early to satisfy a calendar while ignoring whether the material can be verified is how a page ends up in the growing pile of optimised documents that engines are learning to discount. Detected prevalence of that kind of page rose from 7.02% to 16.36% between 2024 and 2026.

Finally, material published to age still has to survive the wait. Google budgets a few weeks or more for URL changes on a medium site and advises holding redirects for at least a year. Pew found 38% of the webpages that existed in 2013 were gone by October 2023, with roughly one in five from 2021 gone within two years. An unmaintained page does not age into authority. It ages out of existence, and the deadline you met becomes a deadline you have to meet again.

FAQ — Publishing Lead Time

Next Steps — Run the Subtraction

▶ Name the quarter you need to be present in, tied to a real commercial event rather than a budget cycle.

▶ Identify which engine your buyers actually use for that decision, because the deadline moves by months between them.

▶ Read the median cited-source age for your engine and your category off the matrix, then refuse the blended average.

▶ Subtract that median and the engine's stated re-fetch window from your target date to get one publish-by date.

▶ Compare that date to today, and if it has passed, reset the target quarter rather than the ambition.

When the subtraction returns a date that has already gone, the work is to pick the next reachable quarter and start against it. Answer Engine Optimization is where that schedule gets built and met.