From SEO to GEO: A 90-Day Migration Roadmap for Existing Search Teams
You do not migrate from SEO to GEO by switching disciplines. You migrate by adding three new layers — a prompt set, a visibility baseline, and a retrofit programme — on top of a search operation that keeps running. Ninety days is enough to install all three, provided you spend the first month measuring instead of writing.
Most teams get this backwards. They commission a batch of new "AI-optimised" articles, and three months later have no idea whether any of it worked, because they never recorded where they started. The migration failed at week one, not week twelve.
This roadmap is for a team that already runs SEO — content calendar, technical backlog, rankings dashboard — and needs to fold generative engine optimisation into it without stopping the machine. It covers what to do each fortnight, who owns it, what to stop doing, and how to tell at day 90 whether the shift is working.
Key takeaways
- Spend days 1–30 on measurement, not content. Without a baseline, every later change is unattributable, and you will end the quarter arguing about anecdotes.
- Retrofit before you commission. Existing pages that already have authority convert into citations faster than new pages written from scratch.
- The migration is mostly organisational. The hard parts are reporting, budget defence and who owns off-site mentions — not writing.
- Do not cut SEO to fund GEO. Generative engines retrieve from search indexes; degrading crawlability degrades your AI visibility with it.
- Day 90 is not a finish line. It is the point at which AI visibility becomes a standing metric with an owner, a cadence and a budget line.
The 90-day plan at a glance
| Phase | Weeks | Primary goal | Main output | Natural owner |
|---|---|---|---|---|
| Baseline | 1–4 | Know where you actually stand | Prompt set, share-of-model baseline, retrievability audit | SEO lead |
| Retrofit | 5–9 | Convert existing authority into citations | Rewritten priority pages, entity fixes, off-site plan | Content + PR |
| Institutionalise | 10–13 | Make it survive the quarter | New dashboard, reallocated budget, standing cadence | Head of marketing |
Sequencing matters more than the calendar. If your baseline slips to week six, push everything else back — retrofitting pages while still blind means you can never separate the effect of your edits from ordinary model drift.
Days 1–30: build a baseline you can defend
The first month produces no published content. Its deliverable is a defensible answer to one question: how often, and in what terms, do generative engines currently mention us?
Weeks 1–2: assemble the prompt set
Your keyword list is not a prompt list. Keywords are fragments typed into a box; prompts are full questions typed to an assistant, and they carry context a keyword never does — budget, company size, the alternative the buyer is already using.
Build the first set from four sources you already own:
- Sales call recordings and demo notes. The exact phrasing prospects use when they describe the problem. This is the highest-value source and almost nobody mines it.
- Support tickets and onboarding questions. These map to retention and expansion prompts, not just acquisition.
- Your existing keyword research, rewritten as questions with context attached.
- Competitor comparison pages — the "X vs Y" and "best X for Y" shapes buyers ask assistants directly.
Sort the result into three intent buckets, because they behave differently and you will want to report them separately:
| Intent bucket | Example shape | What a mention here signals |
|---|---|---|
| Discovery | "How do I track whether AI tools recommend my brand?" | You are considered category-relevant |
| Purchase | "Best AI visibility platform for a mid-market B2B team" | You are in the shortlist |
| Alternative | "Alternatives to [incumbent] for share-of-model tracking" | You are a credible displacement option |
A hundred prompts is a workable starting set for a single product line. Precision matters more than volume: ten prompts that mirror how buyers actually speak beat two hundred synthesised from a keyword export. The set will keep growing after day 90, but freeze it at the end of week two so the baseline means something.
Week 3: record the baseline
Run every prompt across the engines that matter to your market — typically ChatGPT, Gemini and Perplexity — and record, per prompt: whether you were mentioned, which competitors were mentioned, in what order, and with what framing.
Two properties of generated answers make this harder than rank tracking, and both need to be designed around:
Answers are probabilistic. Ask the same question twice and you may get different brands. A single run is an anecdote. You need repeated runs per prompt to get a stable rate, which is the practical reason manual spot-checking does not scale past a handful of queries.
Sentiment is part of the result. Being named as "a lower-cost option with a thinner feature set" is not the same result as being named as the category standard, even though both count as a mention. Sentiment and credibility shape how a mention converts, so capture the framing, not just the appearance.
The output of week three is a single number per engine and intent bucket — your Share of Model — plus a competitor set you will now track against. Automating this collection is what Geomyze does; settle on the method now rather than rebuilding it in month three.
Week 4: audit what is actually retrievable
Now look at your own estate through the eyes of a retrieval system rather than a crawler. You are checking whether your best claims can survive being pulled out of context, because that is what happens to them.
Work through your top pages against five questions:
- Does each section answer its own heading in the first two or three sentences? Retrieval takes fragments. A section that builds to a conclusion in paragraph six loses to one that leads with it.
- Are comparisons in tables? Prose comparisons parse badly; structured ones get lifted intact.
- Do headings describe content or perform cleverness? "How pricing is calculated" is retrievable. "The nitty gritty" is not.
- Can a fragment stand alone? Every "as we discussed above" is a sentence that becomes meaningless once extracted.
- Is your factual data on your own domain, current, and consistent with what third parties say about you? Contradiction between your site and review platforms suppresses confident recommendation.
Score each page and stop there. Do not start fixing — you are building the week-five work queue.
Days 31–60: retrofit before you commission
The instinct at day 31 is to brief a batch of new articles. Resist it for another month. Pages that already hold authority and traffic get retrieved sooner than new URLs, so the fastest path to a citation runs through content you have already published.
Weeks 5–6: fix the answer shape of priority pages
Take the twenty to thirty pages closest to your purchase-intent prompts and rebuild them for extraction: lead every section with its answer, convert comparison prose to tables, replace performative headings with descriptive ones, and add an FAQ block whose H3s match real prompt phrasing.
Prioritise with a simple grid rather than a scoring spreadsheet:
| Page has authority | Page is retrievable | Action | Order |
|---|---|---|---|
| Yes | No | Restructure — fastest available win | First |
| Yes | Yes | Leave alone; add to monitoring | Fourth |
| No | No | Consolidate or retire | Third |
| No | Yes | Build authority off-site, not on-page | Second |
The top-left cell is where a migration earns its early results. These are pages search engines already trust that generative engines currently cannot use. The mechanics of writing for extraction apply the same way whether the page is new or eight years old.
Weeks 7–8: close the entity gaps
Generative engines recommend brands they can describe confidently. Confidence comes from consistent, corroborated facts about who you are, who you serve and what you cost — and gaps in that description show up as vague or hedged mentions.
Fix, in this order: your own canonical facts (pricing, positioning, integrations, ideal customer) stated plainly in text rather than only in graphics; then structured data and organisation markup; then the third-party surfaces where your category gets discussed — review platforms, comparison directories, industry wikis.
Any factual inconsistency between these surfaces is worth more attention than a new blog post. Several distinct factors feed a model's willingness to name a brand, and unresolved contradictions suppress most of them at once.
Week 9: hand the off-site layer to someone
This is the week the migration usually stalls, because the work belongs to nobody. In SEO, you influenced off-site signals through link building, which lived with the SEO team. In GEO, the currency is corroborated mentions across the sources engines retrieve from — and that is closer to PR, community and partnerships than to outreach for links.
Name an owner. Give them the competitor set from week three and one objective: get the brand accurately described in the places assistants read when answering your purchase-intent prompts. Without a named owner, this line item quietly disappears and the migration becomes an on-page project with a ceiling.
Days 61–90: change the reporting, then the roadmap
The last month is where the migration either becomes permanent or evaporates. The technical work is done; what remains is organisational.
Weeks 10–11: rebuild the dashboard
Your current reporting will show the migration as a failure. AI-referred sessions are small and badly attributed, while classic organic clicks are flat or falling for reasons unrelated to anything you did. Report only on sessions and you will be defunding GEO by week twelve.
Add three things to the existing dashboard rather than replacing it: share of model per intent bucket, mention sentiment, and citation rate on the retrofitted pages. Keep rankings and organic sessions in place next to them — the two disciplines share a foundation, and understanding where GEO and SEO actually diverge is what stops the comparison becoming a turf war.
Set expectations on timing at the same time. Retrieval-driven wins can appear within days of a structural fix. Changes to what a model "knows" about your brand move on the timescale of training data and third-party coverage — months, not weeks. Reporting both against one deadline guarantees a false negative.
Week 12: reallocate, do not reassign
Fund GEO from the parts of the SEO budget that were already declining in value — thin content refreshes aimed at informational queries that AI answers now absorb, and volume link acquisition. Protect the technical foundation absolutely. Crawlability, site speed, index coverage and structured data serve both disciplines, because generative engines retrieve from the same indexes search engines build.
Week 13: set the cadence and review
Fix the operating rhythm before the quarter closes: monthly share-of-model review, quarterly prompt-set refresh, and a standing retrofit slot in the content calendar. Then run the week-three measurement again with the same prompt set and compare.
What to protect during the migration
Four things should not change while you are migrating, because each one degrades AI visibility if you touch it:
- Technical SEO health. A page that cannot be crawled cannot be retrieved.
- Pages that rank and convert. Restructuring for extraction should not mean rewriting a page that is already doing its job.
- Publishing consistency. Freezing the calendar for three months while you "figure out AI" removes the recency signal that both systems reward.
- Existing internal linking. Entity relationships expressed through internal links help engines understand what you are, not just what a page says.
Common migration failures and how they show up
| Symptom at day 90 | Underlying cause | Fix |
|---|---|---|
| "We can't tell if it worked" | No baseline recorded in month one | Restart measurement; treat the last 90 days as unattributed |
| Mentions rose, pipeline flat | Prompt set is discovery-heavy | Rebuild the set around purchase and alternative intent |
| Results swing wildly month to month | Too few runs per prompt | Increase sample size before drawing conclusions |
| Organic traffic fell and nobody can explain it | SEO defunded to pay for GEO | Restore technical and refresh budget immediately |
| Pages restructured, still not cited | Retrieval is not the bottleneck — brand knowledge is | Shift spend from on-page to third-party corroboration |
| Programme dropped after one quarter | No owner, no standing metric | Assign owner and put share of model in the monthly pack |
Frequently asked questions
How long does an SEO-to-GEO migration actually take?
Ninety days installs the capability: a prompt set, a baseline, a retrofit queue and a reporting line. Meaningful movement in how often engines recommend you takes longer, because part of it depends on third-party coverage accumulating. Treat day 90 as the point where the programme becomes measurable and permanent, not the point where results are final.
Do we need to stop doing SEO to start doing GEO?
No, and doing so actively harms results. Generative engines ground their answers in retrieved documents from search indexes. Weakening crawlability, page quality or index coverage weakens the retrieval layer that AI visibility depends on. The migration reallocates effort within search; it does not exit it.
Can we run this migration without a dedicated tool?
You can run the retrofit and entity work manually. The measurement layer is where manual work breaks down, because generated answers vary between runs — establishing a stable rate means repeating prompts across engines on a schedule and logging framing as well as mentions. That is a data collection problem, not a judgement problem, and it is worth automating early.
Who should own GEO — the SEO team or someone else?
Split it. On-page retrofitting and technical work belong with SEO. Off-site corroboration belongs with PR, community or partnerships, because the currency is accurate mentions rather than links. Measurement and reporting need a single owner senior enough to defend the budget line at the quarterly review.
What is the first thing to do if we only have two weeks, not ninety days?
Build the prompt set and record the baseline. Everything else in this roadmap can be deferred; the baseline cannot, because its value only exists retroactively. Two weeks of measurement now makes the next quarter's work attributable.
How do we know which competitors to track?
Do not use your internal competitive set. Use the brands that actually appear alongside you in week-three answers — generative engines frequently surface a category peer group that differs from the one your sales team names, and those are the brands you are genuinely competing with for the mention.
Start where the migration actually starts
Every part of this roadmap after week three depends on the baseline you record in week three. Without it, the retrofit work is unmeasurable, the budget conversation is unwinnable, and the programme dies at the first quarterly review.
Geomyze runs that measurement layer for you: it generates intent-based prompts across discovery, purchase and alternative shapes, runs them repeatedly across ChatGPT, Gemini and Perplexity, and reports your share of model against the competitor set that actually appears in the answers — with the framing captured, not just the mention count.
Run your first AI visibility report free → — one full report, no card required. It takes minutes, and it gives you the day-one number the rest of the ninety days is measured against.
