GEO Learning Path
Five stages, from how a language model actually reads a page to measuring whether it cites yours. Primary sources only, one essential resource per stage, and something to build at the end of each.
- Stages
- 12
- Resources
- 75
- Reading time
- 61h
- Cost
- Free
All links verified August 2026
Jump to a stage
Explain, without hand-waving, how a language model turns your page into something it can retrieve and quote.
Most GEO advice is people guessing at a black box. You do not have to guess. Tokenisation, embeddings and retrieval are all documented, and once you understand them, a good deal of the advice in circulation is visibly wrong on its face. This stage sits first because everything after it depends on it.
Break a retrieval system on purpose
Ask an assistant a question you know the answer to and watch which sources it picks. Then ask the same question three more ways. Log what changed. You are looking for the point where retrieval fails or contradicts itself, because that boundary is where all the real optimisation work lives.
Get reliable, repeatable output from a model, which is the difference between measuring AI visibility and collecting anecdotes.
Every serious AI visibility measurement is a prompting problem underneath. If your prompts drift between runs, your data is noise and your reporting is fiction. This stage is short and it is not optional.
Write a prompt that returns the same shape every time
Build one prompt that asks an assistant about a brand and returns a fixed structure: mentioned or not, position, sentiment, sources cited. Run it ten times. If the shape varies, the prompt is not finished. This artefact becomes the instrument you measure with in stage seven.
Configure crawler access from primary documentation, and tell the difference between a bot that trains and a bot that cites.
This stage has the highest ratio of confident advice to correct advice anywhere in the field. Almost every crawler list in circulation is copied from another crawler list, and the errors propagate. Read the providers directly. It takes an afternoon and it puts you ahead of most people writing about this.
Audit a robots.txt against the primary docs
Take any site, yours or a competitor's, and check every AI user agent in its robots.txt against the provider documentation above. Count how many are deprecated, how many search crawlers are missing, and how many lines address agents that ignore robots.txt anyway. Most files fail on at least two of the three.
Judge any GEO claim by its evidence rather than by how confidently it was posted.
There is one peer-reviewed study at the centre of this field and a great deal of vendor content orbiting it. Read the study, then read the systems documentation that predates the AI era and survived it. Together they give you a filter for everything else you will encounter.
Take a claim apart
Find a GEO statistic being shared confidently, the kind shaped like schema produces 2.5x more citations. Trace it to its origin. Note who funded it, what the sample was, and whether the number appears anywhere outside the company selling the tool. Do this twice and you will never repeat one of these numbers again without checking.
Write and structure a page so a single passage can be lifted out, understood alone, and attributed back to you.
A retrieved passage arrives with no page around it. Everything in this stage follows from that one fact, and most of it is ordinary editorial craft rather than anything new. The AI era did not invent these rules, it raised the penalty for ignoring them.
Rewrite one page to survive being fragmented
Take a page you own. Cover everything above each section and read that section alone. Wherever it stops making sense, fix it: name the subject instead of using a pronoun, answer in the first sentence, move context after the answer. Then check whether anything you would not want quoted out of context should be wrapped in data-nosnippet.
Ship a page that a simple fetcher can read and an entity resolver can identify, and open the pull request yourself.
Structured data and server rendering are the two technical levers with actual platform confirmation behind them. They are also the two that content teams cannot implement without engineering help, which is exactly why knowing them well changes what you are worth in a room.
Publish one page and verify it end to end
Build a page with linked JSON-LD entities, validate it in both tools above, then fetch it with curl and a crawler user agent and confirm the content is in the raw response. Not the rendered DOM. The raw response. Most pages fail this and their owners have no idea.
Report AI visibility with numbers that survive being challenged by someone who did not want to hear them.
This is where careers are made or quietly ended. Anyone can screenshot an assistant naming their brand. Producing a repeatable measurement, with a stable prompt set and honest variance, is rare enough to be a differentiator on its own.
Build a prompt set and run it three times
Take the prompt from stage two and write twenty questions around it that a real buyer would ask. Run each three times in fresh sessions across two assistants. Record presence, position, tone and every cited URL. The variance between identical runs will tell you more about this field than any article, including this one.
Build your own measurement instead of paying a vendor for a number you cannot audit.
Every AI visibility tool on the market is a prompt loop, a parser and a dashboard. Once you can write those yourself you stop being a customer of other people's methodology, and you can answer questions no product has a screen for. This is also the stage that most cleanly separates senior practitioners from the rest.
Script your prompt set end to end
Take the twenty prompts from stage seven and automate the whole loop: run them against two models through an API, parse presence and cited URLs, append to a file with a timestamp. Schedule it weekly. From this point on your measurement runs whether or not you remember to do it, which is the only kind that survives a busy quarter.
Run AI visibility across many countries and languages without the markets quietly cannibalising each other.
Almost everything written about AI search assumes one site, one language, one market. Real enterprise work is seven domains and thirty-five markets, where answers differ by region and a single global score is meaningless. This is the least covered area in the field and one of the most valuable to be good at.
Run one prompt set in three markets
Take five questions and run them from three different countries, using a VPN or a proxy if you have no other route. Record which brands are named in each. The differences will be larger than you expect, and they are the argument for why a single AI visibility score should never be reported to anyone.
Get products, prices and comparisons represented correctly when an assistant is asked what to buy.
Commercial queries are where AI search has the most direct revenue consequence and the least published guidance. Product data is structured, feeds already exist, and assistants are increasingly answering buying questions directly. This stage exists because most GEO advice is written for blogs and quietly fails on a catalogue of two million items.
Ask an assistant to shop
Pick a product category you know well and ask an assistant for a recommendation, then for a comparison, then for the cheapest option. Record which retailers are named and trace where each answer came from. You are looking for whose product data is being read and whose is invisible, which is a different question from who ranks.
Operate at a scale where crawl budget, migrations and governance decide outcomes more than any individual page does.
On a small site you optimise pages. Past a certain size you manage systems, and the failure modes change completely: a rendering regression, a migration nobody told you about, a robots rule shipped on a Friday. Everything in this stage is about the failures that erase a year of work in an afternoon.
Write the one page nobody wrote
Document your site's crawl and indexation rules in a single page an engineer can act on: what is disallowed and why, which templates are noindex, who owns robots.txt, what happens on deploy. Almost no organisation has this written down anywhere, and producing it is often the fastest way to be taken seriously by a platform team.
Reason about where this is heading without repeating whatever was posted this week.
Assistants are moving from answering to acting, and the interfaces they use to reach the outside world are being standardised in public. Read the protocols rather than the predictions. You will be years ahead of people forming opinions from conference talks.
Write down what you expect, then date it
Write three predictions about AI search over the next year, with the reasoning for each, and put the date on it. Come back in six months. This is the single fastest way to find out whether you understand this field or are reciting it, and almost nobody in it does the exercise.
75 resources, roughly 61 hours, all links checked in August 2026. This path is deliberately short. There is far more written about AI search than is worth reading, and the filtering is most of the value.
For the foundations of SEO itself, start with LearningSEO.io by Aleyda Solis, which remains the best free roadmap in the field. This path covers the AI search layer on top of it rather than repeating it.