How to show up in ChatGPT and Perplexity: the complete GEO guide
More and more customers ask an AI before they search Google. «Which accountant would you recommend in Valencia?», «which ERP suits a small business?». If your company does not appear in those answers, you lose customers who never ran a search at all.
This is called GEO, generative engine optimisation. It is not hype or a new label for the same old thing. There is published research measuring what makes a model cite you rather than someone else, and the conclusions are fairly concrete.
This guide covers what that research says, the crawlers you have to let in so you can be read, how to structure your pages, and what is actually true about the llms.txt file everyone keeps talking about.
1. What changes when an AI gives the answer
In a classic search engine you compete for a position in a list and for the click. In a generative engine there is no list. There is a written answer mentioning a handful of sources, and either you are one of them or you do not exist.
The detail that explains almost everything is that these systems do not answer purely from what they memorised during training. When you ask about something current, they search the web right then, retrieve a handful of pages, summarise them and link to the sources they used. That retrieval step is your opening.
Classic search engine
SEO- You compete for a position in the list
- The goal is the click
- Ten results per page
- Authority and links win
Generative engine
GEO- You compete to be one of the cited sources
- The goal is the mention, click or no click
- Two or three sources per answer
- Density of verifiable data wins
2. What the research says about getting cited
The reference work is the study GEO: Generative Engine Optimization, by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at the KDD conference in 2024. They tested different ways of rewriting the same content and measured how much the odds of being cited changed. The gain reaches up to 40%.
| Technique | Effect | What it means on your site |
|---|---|---|
| Including expert quotes | The largest gain measured | One quoted sentence from someone with a name and a role is worth more than three of your own paragraphs. |
| Adding data with its source | High gain | Concrete figures, with the origin stated and dated. |
| Referencing your sources | High gain | Linking to studies, official documentation or legislation. |
| Writing clearly | Medium gain | Clean, well-ordered sentences. Clumsy writing is penalised. |
| Explaining things simply | Medium gain | Explaining the complex without unnecessary jargon. |
| Using industry terms | Small gain | The precise word where it belongs, not as decoration. |
| Repeating keywords | Reduces visibility | The old SEO tactic makes results worse here. |
| Adding obscure synonyms | No effect | Swapping words without adding information does nothing. |
Results from the study published on arXiv (2311.09735) and presented at KDD 2024, measured across a broad set of queries and knowledge domains.
What this means in practice
The three techniques that lift the most are the three that make content verifiable. Named quotes, sourced data and checkable references. That makes sense, because a model about to assert something to a user would rather lean on something it can attribute.
And the bad news for anyone still doing what they have always done, repeating the keyword everywhere makes things worse. It is not neutral, it actively costs you.
3. First of all, make sure you can be read
Before writing a single line, check that the crawlers can get in. If you block them there is nothing to optimise, because your site is not even in the draw.
Here is the distinction hardly anyone makes. Each company runs several crawlers with different jobs, and lumping them together is the expensive mistake.
Training
Blocking is optional- Collects content to train future models
- GPTBot, ClaudeBot, Google-Extended
- Blocking it does not affect your rankings
Search
Never block these- Builds the index the answers come from
- OAI-SearchBot, Claude-SearchBot, PerplexityBot
- Block it and you vanish from citations
User-initiated
Let them through- Arrives when someone asks for something specific
- ChatGPT-User, Claude-User, Perplexity-User
- It is a visit requested by a real person
| Crawler | What it does | What to do |
|---|---|---|
| `GPTBot` (OpenAI) | Trains OpenAI's models | Optional. Block it only if you do not want your content used for training. |
| `OAI-SearchBot` (OpenAI) | Indexes for ChatGPT search | Always allow. Without this you are not cited in ChatGPT. |
| `ChatGPT-User` (OpenAI) | Opens a page a user asked for | Allow. Note that OpenAI states robots.txt rules may not apply here, because a person initiated the action. |
| `ClaudeBot` (Anthropic) | Trains Claude's models | Optional, same as GPTBot. |
| `Claude-SearchBot` (Anthropic) | Indexes for Claude's search | Always allow. |
| `Claude-User` (Anthropic) | Opens a page requested by the user | Allow. Anthropic confirms it does respect robots.txt. |
| `PerplexityBot` | Indexes for Perplexity | Always allow. |
| `Googlebot` | Indexes for Google, including AI answers | Never block this. |
| `Google-Extended` | Permission to train Gemini | Optional. Blocking it does not affect your Google rankings. |
| `Bingbot` | Indexes for Bing, which feeds Copilot | Allow. |
Names and functions according to OpenAI's and Anthropic's official documentation, checked in September 2026.
4. The robots.txt we recommend
If you want to be cited but would rather not hand over your content for model training, this is the configuration. It blocks training and lets everything else through.
# Model training: out
User-agent: GPTBot
Disallow: /
User-agent: ClaudeBot
Disallow: /
User-agent: Google-Extended
Disallow: /
User-agent: Applebot-Extended
Disallow: /
User-agent: CCBot
Disallow: /
# Search indexes: in, this is where citations are won
User-agent: OAI-SearchBot
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Googlebot
Allow: /
User-agent: Bingbot
Allow: /
# Requests made by a user at that moment
User-agent: ChatGPT-User
Allow: /
User-agent: Claude-User
Allow: /
Sitemap: https://yourdomain.com/sitemap.xmlIf you would rather also allow your content to be used for training, which is perfectly reasonable when what you want is exposure, drop the first block entirely.
If you use Cloudflare, this changed in September 2026
Cloudflare retired its single block-AI-bots switch and split it into three independent controls, one for search, one for training and one for agents. Since 15 September 2026 you can refuse to have your content used for training without dropping out of the search index, which used to be all or nothing.
In the same update, Cloudflare introduced an accountable-crawler designation for operators that meet its transparency requirements. Apple, Google and Microsoft got it from the start. If your site sits behind Cloudflare, it is worth reviewing those settings, because an older configuration may be shutting out the very crawlers you want.
5. How to write so you get cited
This is where the study's findings apply. It is not about writing for a machine, but about writing with the precision you would use for a report someone is going to check.
- 01Answer in the first lines. Right below the heading, two or three sentences that directly answer the question. That is the fragment a model can lift as is. Open with three paragraphs of preamble and you are just making extra work for nothing.
- 02Give figures and say where they come from. «A click on Meta costs around $0.63 according to Triple Whale's 2026 report» works. «Clicks are cheaper on social media» does not.
- 03Quote someone by name. A quoted sentence from an identifiable person, even yourself with your job title, is the technique that lifts the most.
- 04Link to your sources. To the official documentation, the study, the legislation. A figure with no origin is indistinguishable from an invented one.
- 05Date anything perishable. «Prices checked in September 2026» tells the model the figure is recent and bounded.
- 06Cover the whole topic. Models prefer one source that settles the entire question over five that each settle a piece of it.
- 07Do not repeat keywords. It is measurably counterproductive. Write naturally.
6. How the page itself has to be built
Writing well is not enough if the page makes life hard for whoever reads it. These are the technical points that weigh most.
- Keep the text in the HTML. Many crawlers do not run JavaScript. If your content only appears once the browser code loads, the page is empty as far as they are concerned. Check what shows with JavaScript disabled.
- Nothing hidden in tabs and accordions that only load on click. If the content matters, put it in the HTML from the start.
- Real tables, with proper table, row and cell tags. A table laid out with boxes collapses into a run of text when processed, losing the link between rows and columns.
- Headings in order, without jumping from the second level to the fourth. Systems that split content into chunks use headings to know where each idea starts and ends.
- Structured data for your organisation, articles and FAQs, with the `sameAs` field pointing at your profiles. That is what connects your site to the record the search engine holds about your company.
- Questions as headings. An `h2` reading «How much does an Odoo implementation cost?» matches what someone types into a chat far better than one reading «Our pricing».
7. The truth about the llms.txt file
This is the most talked-about and most overstated topic in the field. The llms.txt file, proposed by Jeremy Howard, is a plain-text summary of your site, meant to let a model understand what your site is about without digesting the whole HTML.
So is it worth having? Yes, but for a different reason. It works as a map for agents and coding assistants that need to navigate your site without burning resources, and it costs very little to generate. We publish one for exactly that reason, not because we expect it to rank us.
What you must not do is add it and assume you have now done GEO. The real work is in the earlier sections.
8. Where to start and how to measure it
| Phase | What to do | How you know it is done |
|---|---|---|
| 1. Access | Review robots.txt and your CDN settings so search crawlers get in. | Search bots get a clean response in your server logs. |
| 2. Content | Rewrite your main pages with a direct answer up top, sourced data and quotes. | Every page answers its question in the first lines and holds quotable data. |
| 3. Structure | Text in the HTML, real tables, headings in order and structured data. | The page reads in full with JavaScript disabled. |
| 4. External presence | Reviews, industry directories, mentions and comparisons on other sites. | Your company shows up across several independent sources. |
| 5. Measurement | Track visits arriving from assistants and the mentions themselves. | You know how many enquiries come from there. |
How you measure this, which is the awkward part
There is still no equivalent of Search Console for generative engines. What you can do is this.
- Ask them yourself. Write twenty questions your customer would ask and run them monthly through ChatGPT, Perplexity and Gemini. Note who shows up. It is manual, but it is the most direct data available.
- Watch where traffic comes from. Your analytics already show visits originating from chatgpt.com or perplexity.ai. Group them so you can see them together.
- Check your server logs. They show which crawlers come in, how often and which pages they take.
And it is worth saying plainly, nobody can guarantee an AI will cite you. What you can do is make it easy and raise the odds considerably. Anyone promising you a position in ChatGPT is selling something they do not control.
Sources
Questions about how to show up in ChatGPT and Perplexity: the complete GEO guide
No, it builds on it. Generative engines retrieve pages from search indexes, so ranking well still helps. What changes is how you write and how you structure what you already have.
No, and be wary of anyone who promises it. Nobody controls what a model answers. What you can do is meet the conditions that make citation likely, starting with letting its crawlers in.
That is your call and it does not affect your rankings. If your content is your product, blocking makes sense. If you want exposure, leaving it open is reasonable too. What matters is not confusing those crawlers with the search ones.
It works as a map for agents and coding assistants, and it costs little. What it is not, despite what gets repeated, is a ranking factor. Google has said no AI system currently uses it to decide what to surface.
Changes on your site are picked up by the indexes in days or weeks. Building presence across external sources, which weighs most in recommendations, takes months.
Especially for a small company. In a classic search engine you compete against the accumulated authority of the big players. In a generated answer, whoever answers that specific question best competes, and there a small, precise site has plenty to say.
Let's talk about your
next project.
No sales pitch. A direct conversation about how automation and web development can transform your operation in the age of AI.
Support across Spain, LATAM and the USA
Spain · Colombia · Latam · USA
A direct reply and assessment in under 24 hours.