What has changed
Until recently, someone looking for a supplier opened Google, looked at the first three results and asked two or three for a quote. That path still exists, but another one has appeared alongside it: opening ChatGPT and typing "I need a company to set up a CRM for me, which ones would you recommend?"
The difference is brutal for whoever is selling. On Google you compete to be among ten links. In an AI answer you compete to be among two or three names. Either you are there or you do not exist. There is no second page to fall back on.
We call it GEO, for Generative Engine Optimization. It is an ugly name but it is the one that stuck.
First: make sure they can read you
This part looks obvious and it is where more people fail than you would think.
AI crawlers are not Google's. GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot and CCBot are different bots, and there are two very common ways to shut them out without realising.
The first is robots.txt. Plenty of sites block these bots, sometimes because someone copied a template that went around when everyone was talking about protecting content from AI training. If your business wants AI to recommend you, blocking them is shooting yourself in the foot.
The second is quieter and affects a huge number of modern sites: if your site is built in React, Vue or Angular without server-side rendering, the HTML reaching the crawler is essentially empty. Googlebot executes JavaScript and eventually sees the content. Most AI crawlers do not. To them your site is a blank page.
It is easy to check. Open a terminal and run `curl -s https://yourdomain.com | wc -c`, looking at how much real text is there, or disable JavaScript in your browser and reload. If you see nothing, you have a problem, and it is not a content problem.
Second: make sure they know who you are
Models do not reason about websites, they reason about entities. They need to be able to say "Elevatec is an artificial intelligence agency based in Madrid that does X, Y and Z" with some confidence.
That confidence is built with structured data. A JSON-LD block of type Organization declaring your name, description, services, areas served, founder and, above all, the `sameAs` field with links to your external profiles: company LinkedIn, directory listings, a Clutch profile.
That field matters more than it looks. It is what lets a model cross-reference sources and confirm that the company your site talks about is the same one appearing on LinkedIn and in the industry directory. Without it, your site is an isolated claim.
Third: answer before you explain
This is where a writer's instinct works against you.
A good traditional blog post opens with context, frames the problem, develops it and concludes. Content that gets cited by AI does the opposite: it answers in the first two sentences and develops afterwards.
Models extract the most informative self-contained fragment they can find. If your first hundred words are introduction, there is nothing to extract. A 600-word article that opens with the answer gets cited more than a 3,000-word one that opens with "in today's landscape of digital transformation".
Two things help a lot here. Writing H2s as the actual question someone would type into a chat. And adding a frequently asked questions block with FAQPage markup, which is the format models process best for pulling out definitions.
Fourth: get talked about elsewhere
This is the uncomfortable part, because it is not solved by touching your own site.
When a model has to recommend suppliers, it leans heavily on sources that are not the company's own website: comparison articles, industry rankings, directories like Clutch or GoodFirms, press coverage, threads where someone asks and others answer.
It makes sense when you think about it. A model trained to be useful is not going to recommend a company based only on what that company says about itself.
So the hardest part of GEO is also the most decisive: earning mentions on sites you do not control. Complete listings in your sector's directories, presence in the comparisons where your competitors already appear, content published in trade media.
How to know whether it is working
There is no Search Console for ChatGPT. Not yet. So measurement is more manual than anyone would like.
What we do is a fairly simple monthly protocol: a fixed list of fifteen questions a real prospect would ask, run through ChatGPT, Gemini, Perplexity and Claude, always in a fresh session with no history. You log the date, the exact question, the engine and whether the brand appears. It is a spreadsheet, nothing more.
The second indicator is referral traffic. Analytics starts showing visits originating from `chatgpt.com` and `perplexity.ai`. Usually low volume, but it converts far better than average, because whoever arrives comes with the recommendation already made.
How long it takes
Less than classic SEO, especially if the site started from a poor technical base.
The technical side (crawlability, structured data, content format) is fixed in weeks and shows up as soon as the crawlers come back round. The external mentions side is slower, three to six months, because it depends on other people.
What is worth understanding is that this is not a campaign with an end date. Models get retrained, the search engines feeding them change, and your competitors are moving too. It is maintenance, not a project.
Frequently asked questions
Four things are needed: allow AI crawlers in robots.txt and serve content in HTML rather than JavaScript alone, declare a clear entity with Schema.org Organization structured data including the sameAs field, write content that answers your customers' real questions in the first few sentences, and earn mentions on third-party sources such as industry directories and comparison sites.
No, but they overlap. SEO optimises to rank among a search engine's results. GEO optimises so a language model extracts your content and cites you in its answer. They share the technical foundation (indexing, speed, structured data) and diverge on content format: GEO prioritises direct, self-contained answers over keyword density.
It stops OpenAI using it for training, but it also stops ChatGPT citing you when someone asks about suppliers in your sector. For a paywalled publisher that may make sense; for a company that wants to be recommended it is counterproductive.
Only if the content arrives in the HTML. A single-page application without server-side rendering or prerendering delivers empty HTML, and most AI crawlers do not execute JavaScript. It is solved with server-side rendering or by generating static HTML per route at build time.
The technical side is fixed in weeks and shows up as soon as crawlers revisit the site. The external mentions side takes three to six months because it depends on third parties: directories, media and comparison sites.