All Things AI · Season 3 · What You Can Build · Episode 9
People are asking AI who to work with instead of searching for it. Being the answer inside that response is a different discipline from ranking on a results page, it is barely contested, and the window where being early matters is open right now.
The question changed and almost nobody noticed
For twenty years the question was how do I rank. Somebody types something, gets ten links, and clicks one. Everything the industry knows about being found was built for that shape.
The shape is changing underneath everyone. A growing number of people now ask a model instead, in a full sentence, and get a paragraph back with a recommendation in it. There is no page of ten options. There is an answer, and either you are in it or you do not exist for that question.
That is a substantially harsher outcome than being fourth. Fourth is visible. Not being in the answer is invisible, and there is no way to tell it is happening from any dashboard you currently look at.
What GEO actually is
Generative engine optimisation. The practice of being the source a model uses when it composes an answer, rather than a link a search engine displays.
It overlaps with traditional search work but it is not the same discipline, and treating it as a rebranding of the old one is the mistake almost everyone is currently making.
There is no ranking, only inclusion
A model is not ordering results. It is assembling an answer from material it considers reliable and relevant. You are either drawn from or you are not. Effort spent on the difference between position two and position five has no equivalent here.
It reads for meaning, not for keywords
Stuffing a phrase into a page does nothing, because the model is working from meaning rather than matching strings. What it responds to is whether your page actually answers the question clearly, which is a considerably harder thing to fake and a considerably better thing to be true.
It rewards being unambiguous
Pages that state something plainly get used. Pages that circle a subject for eight hundred words before saying anything do not, because there is no clean sentence for the model to draw on. This is why so much traditional content marketing performs badly here: it was written to occupy space, not to answer.
The three layers, and why they are not the same job
It is worth holding all three in your head at once, because work on one does not automatically serve the others.
Search
The results page. Still the largest source of intent and still worth doing properly. Rewards structure, relevance, and authority in ways that are well understood by now.
Answers
The box at the top that resolves the question without a click. Rewards a direct answer placed early on a page that owns that specific question and does not compete with your own other pages for it.
Generative results
Being drawn from when a model composes a response. Rewards clarity, specificity, consistency across everywhere you appear, and being genuinely useful on a narrow question. Least contested of the three by a wide margin.
What actually works
This is early enough that a lot of what is written about it is speculation. What follows is the part that is straightforwardly sound because it follows from how these systems work.
- Answer the question in the first forty words of the page. Before the setup, before the context, before the story. Models draw from the top and so do answer boxes, and a reader who has to scroll to find out whether you answered their question usually does not.
- One page owns one question. Two pages competing for the same question split the signal and neither becomes the obvious source. This is the single most common mistake on agent websites and it is entirely self-inflicted.
- Be specific rather than comprehensive. A page that answers one narrow question completely gets drawn from. A page that covers a broad subject adequately gets ignored in favour of something sharper.
- Say the same thing everywhere you appear. Models assemble from multiple sources, and consistency across all of them is treated as a signal of reliability. Contradicting yourself across your own properties is worse than saying nothing.
- Write in complete, quotable sentences. A model needs a clean statement it can lift. Fragments, bullet-only pages, and copy that only makes sense with the heading above it are all hard to draw from.
How you tell whether it is working
Not from your analytics, which is the uncomfortable part.
When a model answers a question using your material and the person never clicks, nothing appears in any report you have. The traffic did not happen. The influence did.
So you check it directly. Ask the models the questions your clients actually ask, in the words they would use, and read what comes back. Do it monthly and write down what you see. That is a crude instrument and it is currently the honest one, and it is considerably better than assuming.
The second signal is qualitative and it arrives before any number does: people showing up already informed, referencing things they could only have got from your material, without being able to say where they read it.
Where this goes wrong
Three failures, and the first two are the ones that waste the most time.
Treating it as keyword work with a new name. Repeating a phrase does nothing here. The model is not counting words, it is evaluating whether your page answers the question, and the only way to do well is to actually answer it.
Producing volume instead of clarity. Forty vague pages perform worse than four sharp ones, because none of the forty contains a sentence clean enough to be used. Volume was a strategy that worked on the old layer and it is actively counterproductive on this one.
Waiting until it is measurable. The window is open precisely because this is hard to measure and therefore hard to justify, which is why so few people are doing it. By the time there is a dashboard that proves it, the advantage will have gone to whoever moved without one.
Why the timing matters more than the technique
None of the advice above is difficult. Answer clearly, own one question per page, be specific, be consistent. An agent could implement all of it in a month.
The reason it matters now is that almost nobody in this industry is doing any of it, and the systems that compose these answers are building their sense of who is reliable on a given subject right now, from whatever material exists.
Being early to a search layer has happened twice before in this industry, and both times the people who moved first held the position for years. This is the third time, it is happening at the moment, and the barrier to entry is a month of clear writing.
What it grows
The top of everything else.
Every other capability on this site makes your work better or faster. This one determines whether anyone encounters it. Being the recommended answer is worth more than a page-one ranking ever was, because a recommendation arrives with authority attached rather than as one option among ten.
It also compounds in a way rankings do not. Being consistently drawn from on a subject makes you more likely to be drawn from on adjacent subjects, because the reliability is attached to you rather than to the page.
What it costs
Nothing but the writing, which is the reason there is no excuse. There is no tool to buy and no service that does this for you, and anyone selling one is selling the old discipline with a new label. The work is clear pages that answer real questions, and it is the same work whether you do it for this layer or the other two.
Questions agents actually type
What is GEO and how is it different from SEO?
GEO is being the source a model draws from when it composes an answer, rather than a link a search engine ranks. There is no position two, only inclusion, and it responds to whether a page genuinely answers a question rather than to keywords or volume.
How do I get my business mentioned by AI?
Answer specific questions clearly in the first forty words of a page, let one page own one question, be specific rather than broad, write in complete quotable sentences, and say the same thing consistently everywhere you appear.
Can I measure whether AI is recommending me?
Not from your analytics, because an answer given without a click leaves no trace there. You check it directly by asking the models the questions your clients ask and recording what comes back monthly. It is crude and it is currently the honest instrument.
Is it too late to start with GEO?
No, and it is unusually early. Almost nobody in this industry is doing it, the techniques are not difficult, and the systems are forming their sense of who is reliable right now from whatever material exists.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.