The 90-Day GEO Plan: Eleven Artifacts, In Order

Eleven chapters, eleven artifacts, and one honest problem. Chapter 11 just told you nobody has proved any of it works.

Why run a plan at all

Start with the objection. It is the right one.

The previous chapter reviewed the evidence and found it thin. No published study links a specific change to a change in AI citations under experimental control. The one controlled field study measures referral traffic. And it cannot clear its own placebo test.

So why spend ninety days?

Because that is not the only question. There are two claims in this book and only one of them is shaky.

The shaky claim: doing all of this lifts your citation share by some amount. Nobody can support that. Nor will this chapter pretend otherwise.

The solid claim is narrower and it holds. A crawler that cannot read your page cites nothing from it. A model that has your company confused with another company is answering about the other company. A passage with no self-contained answer cannot be lifted as one. None of that is disputed.

Those are mechanisms, not correlations. Chapters 3, 6 and 4 established them. Each is verifiable on your own site this week. Not next quarter.

So the plan is not a growth program. It is a defect list. Shorter, and honest to sell.

Which is a smaller promise and a much more defensible one. You are removing reasons a machine cannot use you. Whether removing them produces a citation is a separate question that nobody can answer yet.

If your agency told you otherwise, ask which study. Then wait.

The order is not arbitrary

Most plans in this category are a list of tactics in no particular sequence. This one has a dependency chain and it runs in one direction.

Readability comes first. Everything else sits downstream. All of it.

Consider what happens if you get the order wrong. You spend three weeks rewriting passages for extraction. Then you ship them inside a client-rendered shell. No crawler outside Google sees a word of it.

Or you fix rendering and structure. Then you discover the model thinks your product is a different product with the same name. Every improvement landed on somebody else's entity.

Both are real. Both are expensive. And both are avoided by testing the cheap things first.

So the sequence runs in four questions. Can they reach you? Is it you they think they are reading? Can they lift a piece of it? And is there anything about you off-site?

In that order.

Measurement wraps all four. At the start, and at the end.

The first hour

Before the plan, do this. It takes about an hour and it decides how the ninety days should be shaped.

Open a terminal. Fetch your three most commercially important pages with no JavaScript engine. Count the words in each response. Then open the same pages in a browser and count them again.

Two numbers per page: raw and rendered. That is the whole test.

Now read your robots.txt, line by line, out loud if it helps. Look for three things: a Disallow covering scripts or build output, a named AI agent you did not choose to block, and a directive somebody added years ago for a reason nobody remembers.

Then one search: your brand name, in the four major engines, once each. Read the answers rather than scoring them. Is that your company? Does it describe what you actually sell?

You now know three things: whether machines can read you, whether you are blocking them by accident, and whether they know who you are.

Most teams have never held those three facts at once. And they change what the next ninety days should contain.

If the raw counts are healthy, the robots file is clean and the answers describe you accurately, your problem is layer four. Skip ahead and read Chapters 8 and 9 again.

If any of the three fails, you have a defect. Start at day one.

Days 1 to 14: can they reach you

Two weeks. Four artifacts. Almost no engineering. This is the phase most teams skip, and it is the one where the defects actually are.

Start with the two-body diff from Chapter 10. Three URLs, no JavaScript, count the words. Ten minutes, start to finish.

About half of teams find nothing. If that is you: congratulations. Move on.

If your raw response is a tenth of your rendered page, stop here. Nothing later in this plan matters until that is fixed.

Next, the reachability runbook from Chapter 3. Seven steps. About ninety minutes.

Then read your robots.txt properly, against the working version in Artifact 3.2. Search open, training as a decision you make deliberately.

Most companies never made that decision. They inherited a file.

Now the baseline, and this is the step that makes the other eighty-eight days legible.

Run the measurement stack from Chapter 11. Four counts: crawler hits by verified agent, referral sessions, both engine reports, the off-site totals.

Write them down with the date. Then leave them alone. Do not tune them.

Last, the decomposition prompt from Chapter 2. Run it against two models. Keep only what both agree on.

That shows how the machine currently breaks your category apart. It is a diagnostic, not a target. Do not optimize toward it.

Total for the fortnight: about two days of actual work, spread over ten. Nobody needs to write code yet.

Days 15 to 45: identity, then the page

A month, and this is where the effort sits. Open with the collision audit from Chapter 6: twelve runs, one afternoon.

Why first? Because a name collision invalidates everything downstream of it.

If a model believes your brand is a different company, every passage you rewrite is about them. You would be optimizing somebody else's entity. With your budget.

So check it before you spend the month.

Then the cut-point audit from Chapter 4. Twenty pages, one afternoon. It shows where a retrieval system would slice your content. And whether the slices survive alone.

Most do not. That is the normal result. It is also fixable.

Which leads to the two-layer rewrite from Chapter 5. Six moves, then one check.

Do not do this to your whole site. Pick the templates carrying your commercial answers. Do those.

Ten pages properly beats two hundred partly. Every time.

Finish the phase with the markup decision from Chapter 7. Seven questions, one afternoon.

Note the framing. It is a decision, not a rollout. Chapter 7 found most structured data types do nothing visible. So the honest output is a shorter list than you ship today.

One warning about this month. It is the only phase that touches production. Which makes it the only one that can break something.

Ship it as you ship anything else. Staged. Reviewed. Reversible.

Days 46 to 75: the corpus around you

A month, and most of it is waiting. Not a mistake.

This phase behaves differently from the last one. You are not changing your own property. You are counting and participating in things other people own.

Start with unprompted recall, Artifact 8.5. Seven steps, two afternoons. Ten category prompts a buyer would actually type, run properly, scored against named rivals.

It answers one question. When nobody names you, do you come up?

Most companies have never asked. The answer is usually humbling. It is always useful.

Then the off-site audit from Chapter 9. Six steps, one afternoon, no subscription.

Count threads you did not start. Count videos you did not pay for. Divide by the ones you did.

Under one to one, and you are the loudest voice about yourself. Which is the exact profile Reddit's spam systems are trained to find.

Now the discipline part. It is mostly restraint, and mostly waiting.

Chapter 9 was clear about the rules. Answer where you are already named. Disclose the relationship in a sentence agreed with whoever owns legal risk. Never seed a thread.

The Endorsement Guides describe a discussion board and an employee posting promotional messages. That example is forty years older than this industry and it fits perfectly.

Why does this phase get a month for two afternoons of work?

Because the corpus moves slowly. There is nothing to do but participate honestly. Then let time pass.

Which is the least satisfying instruction in this book. It is also the accurate one. We have tried the alternatives.

Days 76 to 90: re-measure, and be honest about it

Two weeks. Re-run the four counts from Chapter 11. Same day of the week, same definitions you wrote down on day one. No improvising.

Then the part almost nobody does. It is what separates this plan from a case study.

Compare it against your control set.

Chapter 11 made the argument. Without an untouched comparison set, a rise in your numbers cannot be told apart from the platform growing underneath you. The one controlled study found treated pages grew 5.7 times. Untouched pages on the same domain grew 3.5 times.

Most of that was the tide. Your numbers have a tide in them too. Every set does.

So on day one, before any of this, mark a comparable set of pages as untouched. Did you? If not, mark them now, and accept a weaker read this cycle.

Then interpret carefully. The four counts differ in weight.

One more instruction for day 90. It costs nothing.

Write down what you expected to happen, then what happened. Keep both.

Do that for four cycles and you have something nobody in this field currently has. A private record of which interventions moved your own numbers. On your own site.

That record beats any benchmark you can buy. Nobody else has your site.

The whole thing on one page

Here is every artifact in the book. In the order to run them.

If you only have thirty days

Sometimes ninety is not available. A board asked, a quarter is ending, somebody wants a number by month end.

So here is the honest short version. Three items. We would defend the cut to anybody.

Run the two-body diff. Run the reachability runbook. Run the collision audit.

That is the whole list, and it prices out at ten minutes, ninety minutes and one afternoon.

Those three find the defects that make everything else pointless. Each has a mechanism behind it, not a correlation. And if all three come back clean, you have learned something real: the problem is not technical, and no amount of technical work will fix it.

What we would not do in thirty days is the rewrite. It is the most expensive item and it touches production. You will not see a signal inside a month even if it works.

And skip the vendor subscription until day 91. You cannot evaluate a tool against a baseline you have not taken.

After day 90

The plan ends. The work does not. The shape it settles into matters more than the first cycle did.

We run it as a quarterly rhythm with a monthly check. The monthly check is deliberately small.

The five ways this fails

All five are ones this practice has watched happen. Three are sequencing errors. Two are errors of appetite.

The first: starting at layer four. Somebody buys a visibility subscription in week one, and the dashboard becomes the project. Nine months later the raw HTML is still empty.

The second: rewriting before checking identity. A month of good work, landing on an entity the model has confused with somebody else. Chapter 6 exists because we watched it happen.

The third: no baseline. The team does everything right, then cannot say whether anything changed. Nobody wrote down day one. It is the most common failure here, and the easiest to prevent.

The fourth: doing all two hundred templates. Effort spread evenly across pages that matter and pages that do not. The commercial answers get the same attention as the careers page.

The fifth is different. It is the one nobody admits: the plan runs, the numbers barely move, and somebody quietly decides the whole category is fake.

That last one deserves an answer.

Ninety days is one cycle. The slow counts are slow by nature. The off-site corpus took years to accumulate for whoever is beating you. It will not reverse in a quarter.

What ninety days can tell you is whether the defects are gone. That is a real answer to a real question. It is just not the answer people hoped for.

What not to do in ninety days

Shorter list. Every item on it will be pitched to you.

Do not commission a schema rollout across every template. Chapter 7 traced what most types actually do. The honest output was a shorter list, not a longer one.

Do not seed Reddit threads. Chapter 9 priced that exactly. An employee posting without naming their employer is Example 8 in the Endorsement Guides. Reddit now catches roughly 25,000 spam items a day.

Do not buy a dashboard before you hold a baseline. Compare it to what?

Do not serve different content to crawlers. Google calls dynamic rendering a retired workaround. No AI vendor publishes any policy on it at all, and Chapter 10 covered what happened to a magazine that tried.

Do not chase llms.txt. Google states that Search itself does not use it, and Chapter 7 carries the quote.

Last: do not rewrite everything for "AI readability" as a distinct discipline. Layers one through three are largely good technical practice. With a new label on the invoice.

What to tell whoever is paying

This plan creates a conversation you have to manage upward. So here is the frame.

Do not promise a citation lift. It cannot be forecast. Chapter 11 explains why in detail, if anybody asks.

Promise defect removal instead. There is a specific number of reasons a machine currently cannot use your site. This plan finds them. At day 90 you will know how many are left.

Give them the cost honestly: roughly six days plus a scoped rewrite. That is small enough that the decision does not need a forecast to justify it.

Then set the day 90 expectation on day one. Two of the four counts will have moved. Two will barely have twitched, and that is the designed outcome rather than a disappointment.

Say the uncomfortable part out loud early. If the audits come back clean and nothing improves, the problem was never technical. Which is worth knowing. It is cheaper to learn this way than after a year of retainer.

Who actually does this

The plan needs three kinds of hands. It does not need three people. Often one will do.

Somebody who can read a server log and run a command line owns phase one. In most companies that is a developer for half a day, not a hire.

Somebody who owns the words does the rewrite. That is the largest block of effort, and the one most often handed to whoever is free. Which is how a month of work lands on templates nobody visits.

And somebody has to own the legal sentence before anybody posts anything off-site. Chapter 9 was specific. The Endorsement Guides expect training and monitoring, not a line invented in the moment by a support engineer.

In a small company all three can be the same person across three months. In a large one the risk inverts: four teams each own a layer, and nobody owns the order.

Which is the failure we see most. Not incompetence: sequence.

What this costs

Roughly six working days across ninety. Plus the rewrite. That is all.

The rewrite is the variable, and it is worth scoping honestly. On a modern stack with a handful of commercial templates, days. On a large legacy site, weeks. The correct move there is ten templates that carry your commercial answers, not two hundred that do not.

Everything else here is a person, an afternoon, a text editor.

Then the item to examine hardest: the recurring one. A visibility subscription is the only line here that bills every month. Whether or not it changes a decision.

That is not an argument against buying one. Take the baseline first. Then ask the runs-per-prompt question before you sign, rather than after.

The objections this chapter has to answer

"Ninety days is arbitrary."

It is. But the ninety-day re-measure is not, because the two trustworthy counts in Chapter 11 move too slowly to read at thirty. The phase boundaries are a convenience. Move them.

"This is just technical SEO with a new name."

Phases one and two, largely yes. Chapter 1 called that belief wrong overall, while granting the part that is true: no index, no answer. Which is what phase one protects. So this is not a criticism of the plan. It is a criticism of anybody selling those phases as a new discipline, at a new price.

"You told us in Chapter 11 that none of this is proven, then gave us a twelve-step plan."

Fair, and it is the tension this chapter opened with. The plan is defect removal justified by mechanism, not a growth program justified by outcome. Every step is cheap, reversible, and independently sensible even if AI search disappeared tomorrow.

"Our competitors are moving faster than this."

Some are. Chapter 9 covered what the fast version usually is. Reddit now catches roughly 25,000 spam items a day, using models trained to find exactly that pattern.

Then the disclosure, one last time. We sell this work. And the plan above is written so a competent team can run all of it without us. Which is either a poor commercial decision or the only version worth publishing.

What would change our mind

Three things, and they are the same three that would change the whole field.

A randomized study with citations as the outcome. Matched pages, one variable, a real control. The design is not hard and nobody has published it.

An engine reporting clicks from AI answers. Google gives impressions, Bing gives citations, and neither closes the loop to a visit.

A vendor publishing its prompt set and its runs per prompt. One would force the rest.

Until then, this is the most defensible sequence anyone can build from what is actually known. Better that than a forecast.

A note on how fast this ages

Every claim in this book carries a date, and that was not a stylistic choice.

Consider what moved while these twelve chapters were being written. Bing shipped citation reporting to publishers in February 2026, and Google shipped generative AI impressions in June. Microsoft published its most detailed crawler documentation in years in March. Reddit went to court over scraping and won most of a motion in July.

Four material changes in seven months, in a field that did not exist as a job title three years ago.

So treat the specifics here as perishable and the method as durable. Any figure in this book should be re-checked against its source before you put it in a deck with your name on it, and every source is printed at the end of every chapter for exactly that reason.

That is also the honest answer to the question of whether to buy a book about a moving target.

You are not buying the numbers. You are buying a way of interrogating them, which is the part that survives the next release.

If a chapter here turns out to be wrong, better you catch it with the sources than believed it because we wrote it confidently.

Twelve chapters in twelve lines

Before the plan leaves your hands, here is what the book actually found. One line each.

One: there is no list, so there is no position to hold in it.

Two: an answer is assembled in stages, and you are visible at only some of them.

Three: each engine feeds from a different pipe, and one of them is a training corpus that closed years ago.

Four: the unit of retrieval is a passage, not a page.

Five: a passage that needs the paragraph above it cannot be lifted alone.

Six: if two companies share your name, the model may be answering about the other one.

Seven: most structured data types do nothing visible, and the newest protocols do not read markup at all.

Eight: models recognize far more brands than they recall, and recall is built from accumulated mentions.

Nine: Reddit is a licensing contract and YouTube is open for watching and closed for reading.

Ten: most AI crawlers fetch your JavaScript and never execute it.

Eleven: the instrument moves more than the thing it measures.

Twelve: fix the defects in order, measure against a control, and be honest about what you cannot prove.

Read those twelve lines back and notice what is missing. There is no growth promise anywhere in them.

That was deliberate. We could not find the evidence for one. So we did not write one.

What this book did not cover

Four gaps, named deliberately. A book that claims to cover everything is selling something.

Non-English engines. Baidu, Yandex, Naver and the Chinese assistants operate under different rules, different crawlers and different law. Nothing here transfers unchecked.

Paid placement inside AI answers. OpenAI documents an ads crawler and the surface is moving quickly. By the time you read this there may be a market where this book describes none.

Agents that buy things. Everything here assumes a human reads the answer. When the agent completes the purchase, the questions change: not whether you are cited, but whether you are machine-transactable.

And the legal layer beyond disclosure. Chapter 9 covered the Endorsement Guides and one live case. Publisher licensing, training-data litigation and the copyright questions underneath all of it are a book of their own, and not one we are qualified to write.

So where does that leave the scope? Narrower than the title implies. Which is honest.

What is left

Nothing. Which is rather the point.

Twelve chapters, and it has been one argument the whole way. There is no list to rank in. There is a machine that fetches, reads, retrieves and assembles. Every chapter took one stage of it and asked three questions: what is documented, what is measured, what is being sold.

Much of what is sold turned out to rest on very little. A 40% lift measured in a simulator. A rendering study twenty months old. A citation share on an unstated denominator. Protocols that do not take the markup everyone is shipping.

And some of it held. Crawlers that cannot execute JavaScript. Entities that collide. Passages that cannot stand alone. Corpora that accumulate slowly and cannot be bought.

The difference between those two lists is the whole book. And the method for telling them apart is the only durable thing in it. Read the document. Check the date. Then ask who measured it, on what, and whether they had a control.

The engines will change again this year. They changed twice while we were writing this.

The method will not.

So run the first hour this week, before the plan, before a budget, before anybody proposes a retainer. Three fetches, one robots file, four questions asked of four engines.

An hour of your time, and you will know more about your own position than most companies in your category know about theirs.

Then decide what your ninety days should contain, with three facts in hand instead of a proposal.