How Long Does GEO Take to Work?

Most agency answers to this question are a single number, and a single number is wrong the way an average is wrong: it describes something nobody actually experiences. What produces movement in AI answers is not one process on one schedule. It is three, running in parallel, separated by an order of magnitude.

The awkward part becomes obvious on a second read. An owner sold "AI visibility" who expects a result in six weeks may have bought work that lives on the third row. Correcting that mismatch is what this page is for.

Why one number cannot be honest

An assistant answering a question about your business is not consulting one store of knowledge. Some of what it says comes from what the model absorbed during training. Some comes from documents it fetched seconds ago. Some comes from a structured record of your business that neither of those created on its own.

Those inputs update on completely different schedules. A page you publish tonight can enter a live retrieval index this month. A model's baseline knowledge was fixed before it shipped. The consensus about what your company is forms slowly, across records you do not own. Asking how long GEO takes without saying which of the three you mean is like asking how long a renovation takes without saying whether you mean the paint or the foundation.

The research supports the shape of the intervention without settling the timing question. Aggarwal and colleagues, the Princeton and Georgia Tech team that named the field, showed in their KDD 2024 paper that content-side changes such as adding statistics, quotations, and citations to authoritative sources can lift visibility in generated answers by up to 40 percent. That is well-evidenced about what moves. It says nothing about when, and no equivalent peer-reviewed timing study exists. Anyone quoting a precise week number is quoting an internal observation, not a finding.

Clock one: the source layer moves in weeks

The fastest lever is the one under your own hands, because retrieval happens at question time. An assistant searching mid-answer reads a current index, not a memory. Fix a thin page today and the corrected version can be the one that gets read, once the index catches up.

Google states the prerequisite plainly in its own AI features documentation: to be eligible to appear in generative AI features on Google Search, a page must be indexed and eligible to be shown in Google Search with a snippet. That single sentence dissolves a great deal of confusion. There is no separate AI pipeline you can enter directly. Ordinary indexing is the gate, and the gate has a queue.

How long the queue runs is partly yours to influence. Google's crawl budget documentation describes a crawl capacity limit that responds to your server: if the site responds quickly the limit goes up, and if it slows down or returns server errors the limit goes down and Google crawls less. The same page is blunt about what crawling guarantees, noting that not every page that is crawled will necessarily be indexed, and that after crawling each page must be evaluated, consolidated, and assessed to determine its suitability for the index. Crawl, index, eligibility, selection: four gates, each with its own delay, before an answer engine can quote you.

That is where the four-to-twelve-week band in the Quick Answer comes from. A clean site often sees new pages processed inside a week or two. A neglected one with slow responses and no internal linking takes far longer, and some pages never clear the fourth gate. If you are starting from scratch, our explanation of what GEO actually is covers the mechanics underneath.

Clock two: entity consolidation moves in months

The second mechanism is slower because you are not the only author. Entity consolidation is how the systems answering questions come to agree on a stable understanding of who your business is: one name, one location, one category, one set of claims, corroborated in enough independent places that the answer is safe to repeat.

You can publish your own version of that record in an afternoon. You cannot publish anyone else's. Directory records, association listings, supplier pages, local news, review platforms, and the incidental mentions that accumulate around an active business all update on their own schedule, and the systems reading them refresh at their own pace on top of that. Entity work resists compression not because it is hard, but because it is distributed across parties who do not answer to you.

Two things follow. Businesses that consolidate fastest usually had a scattered but real footprint already, since the job is reconciliation rather than creation. Businesses that take longest are new ones with no record at all, because agreement cannot form around nothing. That asymmetry is the largest source of variance between two companies given identical work.

A qualifier worth stating outright, since this is where expectations get abused: how quickly any of it lands depends on how contested your category is and how much public record already exists about you. Two businesses can receive identical work and see very different curves. That is not a hedge, it is the mechanism.

Clock three: model training moves on release cycles

The third mechanism is the one no agency can sell you, and the one most often quietly implied. An assistant answering from memory draws on a model whose knowledge was fixed at a training cutoff. Nothing published after that date exists in that memory until a later model is trained and released.

These dates are published. Anthropic lists a training data cutoff of January 2026 for its current Claude models, with the older Haiku 4.5 sitting at July 2025. The implication for an owner is direct: a company that opened in March 2026 is absent from those models' unprompted knowledge regardless of budget, effort, or website quality. The only route in is retrieval.

Reasoning from that division is the discipline. Spend effort where the mechanism is documented and responsive. Treat the training layer as weather: real, consequential, not something you schedule around.

Why AI answers move on a different clock than Google rankings

Owners often notice something odd early on: an assistant describes the business accurately before the Google ranking for that topic has moved. Or the reverse, rankings improving while ChatGPT repeats something outdated. Both are normal, and the explanation is structural.

A Google ranking is a competitive position. To move up you have to displace somebody, and everybody above you is working too. A citation inside a generated answer is a selection problem rather than a displacement problem. Google's own documentation describes query fan-out as a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results, which turns one user question into many small retrieval problems. A page that answers one narrow sub-question cleanly can win that slot without ranking first for anything.

It cuts the other way too. Because selection happens per answer rather than per position, it is less stable. We break this down further in our comparison of GEO and SEO as separate disciplines, but the timeline consequence is simple: AI citations appear earlier and vanish more easily, while rankings arrive later and hold better.

That middle stretch is the normal experience of the process, not evidence it has failed. It is also what a failing engagement feels like, which is why measurement matters more than mood.

The surface you are measuring moves on its own

Here is the complication almost no timeline article admits: part of what you observe over your first six months has nothing to do with your work. Generated answers are recomputed on every request rather than cached, and the platforms change how often and how widely they cite.

The clearest published example comes from seoClarity, whose chief architect Mitul Gandhi tracked average citations per prompt in ChatGPT across five markets including Canada. Between February and April 2026 citation volumes fell by 86 to 94 percent, with sharp inflection points on March 8 and April 19. In May they rebounded toward pre-March levels. A business measuring its own progress that spring would have watched its citation count collapse and recover while its website did nothing unusual.

Two consequences. Never judge an engagement on a single spot check: one query on one day is a sample of one, drawn from a distribution that shifts weekly. And build the measurement before the expectation. Our guide to tracking AI citations properly covers the instrumentation. The discipline underneath is Vector 11, Measure: watch citation presence as a trend, with platform volatility treated as a known variable.

If you would rather see where you stand today than argue about how long it should take, our no-charge AI visibility audit reports what the assistants currently say about your business, correct or otherwise.

A twenty-minute self-check that tells you which clock you are on

Before accepting a timeline from anyone, including us, run this. Twenty minutes, and it tells you more than a proposal will.

Score it honestly. Failures at one and five are fast fixes with fast payoffs. A failure at three or four is a real project, and anyone promising six weeks on it is selling the fast clock to fund the slow one.

What cannot be promised, stated plainly

A timelines page that refuses to name its limits is just a longer sales pitch. Here is the boundary.

Nobody outside these companies can promise a date for a specific citation in a specific assistant. The selection logic is undocumented at that level, platforms revise it without announcement, and the seoClarity data above shows how much movement one revision produces. Nobody can promise inclusion in a future model's training data, because no vendor publishes those criteria. Nobody can compress clock two below the pace at which third parties publish.

What can be committed to is the work and the evidence of it: which pages ship, which structured data is deployed, which records get reconciled, and what the measurement shows month over month. That is an engineering commitment rather than a forecast dressed up as one.

One consequence for budgeting: because the clocks run simultaneously, the spend profile that works is steady rather than front-loaded. Paying heavily for three months and stopping captures only the first clock, which is also the least durable. The arithmetic is in our breakdown of what SEO should actually cost, and the same logic governs GEO retainers.

Formative Digital's Results Guarantee applies here in its exact form: if your existing domain shows no measurable organic search results after 12 months of work with Formative Digital, we work for free until you see them. It is scoped to existing domains, because new domains run on a longer clock for the reasons above.

Written by Matt Griffin, founder of Formative Digital, Brantford, Ontario. Published 2026-07-20.