What GEO is in SEO, and what it isn't
GEO stands for generative engine optimization, a name that tells you what it aims at and nothing about what it involves. It sounds like a category someone invented to sell you something, and part of that is fair. The problem underneath is real though, and it fits in one sentence. When someone asks ChatGPT which agency to hire, the model replies with two or three names, and you are either in that answer or you don't exist for that person.
The difference with classic SEO is in the shape of the answer. On Google you compete for a slot in a list of ten blue links and the user picks one. In an AI assistant there is no list, there is a written answer that mentions a handful of names, and no page two to fall back on.
Now for what it isn't, because there is plenty of smoke around this. GEO is not a plugin you install. It is not writing your copy "with AI in mind", whatever that means. It is not filling the site with questions and answers to see if any of it sticks. And it is certainly not a separate product with its own fee sitting next to your SEO, although half the market is packaging it that way.
If someone sells you GEO as a separate line on the quote, they are charging you twice for the same work.
GEO and SEO are not separate things
This is the part almost nobody explains properly, and it is the one that saves you the most money. A language model does not have its own crawler running on magic rules. It reads what a normal crawler reads, and in many cases it uses the index of a classic search engine directly. ChatGPT leans on Bing's index, and AI Overviews are literally Google.
So the website a model can read, understand and cite is, ninety per cent of it, the same website that ranks well. What changes is the last stretch.
| Aspect | Classic SEO | GEO |
|---|---|---|
| What you win | A slot in a list of links | Being mentioned inside an answer |
| How many people get through | Ten results share the click | Two or three names, sometimes one |
| The unit that competes | The whole page | The single paragraph that can be lifted out |
| What weighs most | Inbound links and domain authority | Being named on sites the model already reads |
| How it gets measured | Average position, impressions, clicks | Mentions in answers and referral visits |
| Technical base | Crawling, speed, structured data | Exactly the same |
Look at the last row, which is the conclusion of the whole table. The technical base is identical. That is why GEO doesn't get hired separately but goes inside the monthly SEO, since whoever is fixing your crawling, your markup and your structure is already doing half the work, and charging you a second fee for the other half is charging you to split one invoice in two.
How a model decides who to cite
There is no official documentation explaining it, so anyone handing you an exact algorithm is making it up. What can be observed are four conditions that hold in almost every website that ends up cited. They go in order, because if the first one fails the other three don't matter.
That it can read you, which is less obvious than it sounds.
Plenty of websites block GPTBot, ClaudeBot or PerplexityBot in robots.txt without anyone having decided to, because it came that way in the template or a security plugin put it there. And if your content gets painted with JavaScript after load, some crawlers are left looking at a blank page. Before optimising anything, check that the door is open.
That it understands what you are without having to work it out.
Structured data stopped being a trick for pulling stars in Google and is now the cheapest way to tell a machine what you sell, where you are, what it costs and who stands behind it. A model that finds a block of markup with your service, your price and your city has nothing left to guess, and what isn't guessed can't be got wrong.
That it can lift out one sentence and use it on its own.
Here is the biggest practical difference with SEO as we knew it. A model doesn't cite pages, it cites fragments. If your answer to a specific question is spread over four paragraphs that each depend on the others for context, there is nothing to extract. If it sits in one sentence that stands on its own, with the figure inside, it can be quoted exactly as it is.
That somebody else talks about you.
What you say about yourself counts for little, while what others say about you on sites the model already reads counts for a lot, and that is where sector directories, business listings, forums and the local press come in. It is the equivalent of link building as we have always done it, except that a mention with no link counts too, because the model reads text and not only tags.
How I found out this actually works
On an ordinary discovery call, one of the twenty minute ones. I asked the client how they had found me, which is what I always ask, expecting the usual answer of Google, a mutual contact, someone who had passed on my number. They told me they had asked ChatGPT who could build their website and that my name came up there, alongside two others. They had not opened a search engine once in the whole process.
That was not entirely luck. This website has been built around the four signals from the previous chapter since day one, and the good thing about telling it with my own site as the example is that you don't have to take my word for it. You can check the lot by opening two tabs and it takes less than a minute.
- The robots.txt gives explicit permission. GPTBot, ClaudeBot, PerplexityBot and Google-Extended are allowed in writing and by name, not inherited from a template. It is at burguerastudio.com/robots.txt and it reads in five seconds.
- There is an llms.txt at the root. A plain text summary of what this studio does, which services it has, what each one costs and where everything lives, so a model can work it out without crawling a hundred pages. You have it at burguerastudio.com/llms.txt.
- Every price is written down, and in a sentence that stands on its own. There isn't a single "ask us" anywhere on the site. If a model has to answer what a landing page costs, it finds the figure and the unit on the same line, without stitching three paragraphs together to work it out.
- Nothing is painted with JavaScript. The HTML comes out of the server finished, so what a crawler sees is exactly what you see. That is the difference between being read and being a blank page.
- The FAQs are marked up one by one. Not as decoration for pulling stars in Google, but because a question with its answer next to it is the easiest unit to cite there is. That is where the 323 pairs come from.
When I open somebody else's website I look at those same five things and nearly always in that order, because they run from cheapest to most expensive to fix. The robots.txt and the render are an afternoon. Structured data is a week. Rewriting the answers so they can be cited on their own is real work, and building mentions elsewhere takes months. Starting at the end is the most common way to spend money without moving anything.
And there is one detail about this channel worth knowing before you get into it. There is no console where you can check whether you are showing up, and no report that warns you. This website gets visits arriving from chatgpt.com and that much shows in Analytics, but knowing which answers you appear in and which you don't means asking the question yourself, by hand, over and over.
Nobody tells you that you are showing up in those answers. You find out when a client mentions it on a call.
What actually gets touched
Down to earth, and in the order I would do it if you started tomorrow. The first two points are an afternoon's work. The last two take months, like everything that works.
- Open the door. Go through the robots.txt and take out the blocks on AI crawlers that are sitting there for no reason. If your site gets painted with JavaScript, look at what a crawler really receives before assuming anything.
- Mark up what you already have. Structured data for service, local business, FAQs and breadcrumbs. It is not new content, it is labelling what is already written so a machine doesn't have to interpret it.
- Rewrite the answers so they can be lifted out. Every question a client asks you deserves a paragraph that answers it whole and on its own, with the figure or the timescale inside. It is the most profitable rewriting you can do and the one fewest people bother with.
- Work on the entity. Google Business Profile up to date, consistent presence in your sector's directories, and the business name and address written the same way everywhere. A model that finds three different versions of your details distrusts all three.
- Get mentions elsewhere. Local press, trade associations, comparison pieces, forums where people ask about your sector. They work even without a link, which is the big difference from classic link building.
None of those five things is exclusive to GEO. All five also improve your position on Google, and that is precisely why this work goes inside the SEO for AI service and not on a separate invoice.
How it gets measured, which is what nobody covers
This is where most articles about GEO fall apart, including the ones sitting above this one on Google. They explain what it is and hand you a checklist, without telling you how you will know whether it worked. And without measurement this is faith rather than work.
There are three ways and none of them is perfect, so it is worth using all three at once.
- Referral visits, which is the hard data. In Google Analytics you filter by source and look for chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com. It won't tell you how many answers you appeared in, although it does tell you how many people came after reading them, which is what you are paying for in the end.
- Asking the models directly, with a method. Write ten questions your ideal client would ask, run them through ChatGPT, Perplexity and Gemini once a month and note whether you come up and in what position. It is manual and the answers vary between sessions, so it works for reading the trend rather than for showing off a single figure.
- Bing Webmaster Tools, which almost nobody looks at. ChatGPT leans on Bing's index, so if Bing hasn't indexed you properly there is a part of the channel closed off underneath you. It is free and it imports from Search Console in one click.
If your provider talks to you about GEO and doesn't show you any of those three things in the report, they are not measuring anything.
Where a small business with no time for this starts
With the boring part, which is also the cheap part. Before thinking about language models, look at whether your website loads fast, whether the HTML arrives finished, whether you have structured data in place and whether your prices are written down anywhere. If those four are failing, GEO is not your problem yet.
With that sorted, the next step is rewriting the answers so they can be cited and starting to build mentions away from your own website. That is ongoing work of the same kind as SEO and on the same timescales, since the first movements show at three or four months and the serious conclusions at six.
And a warning for when someone rings to sell it to you. This market is full of people who discovered the term six months ago and already offer it as a speciality. Always ask two things, which project they have applied it to and what metric they use to know whether it worked. If there is no clear answer to both, it is a new label on the same old work, or on no work at all.
What I get asked most
GEO stands for generative engine optimization. It means preparing a website so that artificial intelligence assistants such as ChatGPT, Perplexity or Gemini cite it when they answer a question. The difference with classic SEO is that you are not competing for a slot in a list of ten links, you are competing to be one of the two or three names that appear inside a written answer.
No, it sits on top of it. AI models read the web with crawlers similar to those of a search engine, and several use a classic index directly, since ChatGPT leans on Bing's and AI Overviews are Google. That means the technical base is the same, which is crawling, speed, server-rendered HTML and structured data. What GEO adds is that the content can be extracted in standalone fragments and that your brand gets mentioned away from your own website.
In my case it costs nothing extra, because it goes inside the monthly SEO service, whose guide range for a small business is €600 to €2,000/month. The reason is that ninety per cent of the work overlaps with classic SEO, so charging for it as a standalone product would mean invoicing the same thing twice. If a provider hands you a separate GEO fee, ask them to spell out what they do there that they are not already doing in the SEO.
With three checks combined. In Google Analytics you filter visits by source and look for chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com, which is the hard data. Then you write ten questions your ideal client would ask and run them through the three assistants once a month, noting whether you come up. And finally you check Bing Webmaster Tools, because if Bing isn't indexing you properly you have the part of the channel that comes from ChatGPT closed off underneath.
It is a plain text file at the root of the domain summarising what your business does, which services it offers and where each piece of content sits, meant so a model can read it without crawling the entire site. No vendor has declared it mandatory and adoption is still patchy, so don't expect a jump from putting it up. It costs half an hour, it can't do any harm and it tidies your own architecture along the way, which is the part you do notice.
The same timescales as SEO, between three and six months to see movement and around a year for serious conclusions. The technical part gets fixed in weeks, although the mentions away from your website and your brand entity are built slowly. Be wary of anyone promising you will appear in ChatGPT within a month, because there isn't even a console where a request can be submitted.
It is especially useful, because local questions are the ones these assistants answer best and the ones people ask most. When someone asks for a physiotherapist in Palma or a restaurant in Ses Salines, which is the kind of question I see from where I work in Mallorca, the model leans on business listings, directories and reviews rather than only on your website. That is why entity work weighs more than technical work in local, and I cover it inside SEO for small business.
