Generative Engine Optimisation (GEO): How to Optimise for ChatGPT, Claude and Perplexity

What is generative engine optimisation?

Generative engine optimisation, usually shortened to GEO, is the work of making your content and your brand easy for AI assistants to find, understand and cite. When a customer asks ChatGPT, Claude, Perplexity or Gemini a question, the assistant often writes a short answer and may name a few sources. GEO is about being one of the sources, or at least being described accurately in the answer.

It is a young field, and honesty is useful here. The companies behind the assistants do not publish the full rules for choosing sources, and the behaviour changes as products change. What we can do is separate what is documented, what is widely observed and what is guesswork. This article tries to stay on the first two.

The good news for anyone who already does SEO is that most GEO practice is familiar. The less comfortable news is that no technique guarantees a mention. If someone sells you a guaranteed place in AI answers, treat the promise with suspicion. Part of the work happens off your site, in credible publications, and that is the side our link building campaigns address.

How AI assistants find and use information

For practical purposes, an assistant draws on information in two ways, and they call for slightly different thinking.

What the model learned in training

A language model is trained on a very large body of text. What it “knows” about your company reflects how often and how consistently you were described in the material it learned from. You cannot edit that after the fact, which is why long-term, consistent and accurate descriptions of your brand matter. This route changes slowly.

What it fetches when you ask

Many assistants can also search the web at the moment of the question, read a handful of pages and cite them. This is the route where a new page can matter within days or weeks, because it depends on the page being reachable, indexable and clear enough to quote. Perplexity is built around this behaviour, and ChatGPT and Claude can do the same when web search is available and switched on.

Which crawlers are involved

Companies document the user agents their systems use. Examples include GPTBot and OAI-SearchBot from OpenAI, ClaudeBot from Anthropic and PerplexityBot from Perplexity. Google offers a separate control token, Google-Extended, for some of its AI products; it is distinct from the crawler used for Search. Names and purposes change, so check each company’s current documentation instead of relying on a list, including this one.

What stays the same as SEO, and what changes

Google’s published guidance for AI features in Search points back to the usual fundamentals: helpful content, accessible pages, good page experience. The same logic probably applies to other assistants, although they do not document it in the same way. Pages that are clear, accurate and well structured tend to serve both a search engine and an assistant.

The practical consequence is that the answer has to be easy to lift out. A page that makes the reader scroll through four paragraphs of preamble before stating the point serves nobody, including an assistant looking for a quotable passage.

Seven steps on your own site

None of these steps is exotic. They are the kind of work that also helps human readers, which is a good sign that they are the right priorities.

1. Answer first. Open important pages with a direct two- or three-sentence answer to the question in the title, then expand. A definition, a price condition or a process can usually be stated in a few lines.
2. Use clear structure. A sensible heading hierarchy, numbered steps for processes and tables for comparisons make passages easy to extract.
3. State facts that can be checked. Replace “extensive experience” with the year you started, the number of projects, the terms of your service. Where you cite data, name the source and the date. If you want help with sourcing, see how to cite sources and data credibly.
4. Describe yourself consistently. Use the same company name, description and contact details on your site, in your profiles and in published articles. Contradictions make it harder for any system to form a stable picture of who you are.
5. Mark up what you can. Structured data such as Organization and Article does not promise inclusion anywhere, but it states facts in a machine-readable way. The schema.org vocabulary is documented at schema.org.
6. Keep the technical basics healthy. Pages should load, be indexable and be reachable without logins. Check your sitemap and your internal links.
7. Review robots.txt. Open the file and look for rules that block the crawlers you want to allow. A blanket rule written years ago against scrapers may be keeping assistants out.

Mentions beyond your own site

An assistant that is asked about a company has two kinds of evidence: what the company says about itself and what others say. For commercial or sensitive questions, third-party sources are usually more credible than the company’s own pages. That is the point where GEO meets public relations and link building.

In practical terms, that means being described accurately in publications that readers and systems consider credible: news sites, sector blogs, trade media. Press releases and advertorials are among the controlled ways to do it, because you decide the facts and the wording that go into the text. They work only when the text is genuinely informative, and a stream of thin promotional posts can do more harm than good. If you are planning a campaign, our link building packages are organised as monthly campaigns across suitable categories of sites.

Diversity matters for the same reason it matters in SEO. Mentions from a single type of site look narrow, while mentions across national news, business media and sector publications build a fuller picture. The rules of a healthy link profile still apply, including a natural pace and varied anchors. If you are worried about the risks of overdoing it, read our article on the signs of a risky backlink profile.

A word on languages. If you want to be understood by assistants answering in Romanian, publish in Romanian. If you operate across Europe, keep the key facts consistent across language versions: same company name, same description, same service names.

How to check where you appear

There is no Search Console for AI answers, so measurement is partly manual and always approximate. A workable routine looks like this:

• write a list of 10 to 15 questions your customers would ask, in the languages they use;
• put them to the assistants you care about once a month, in a fresh session, with web search on and off if the product allows it;
• record whether your brand appears, how it is described, whether the description is correct and which sources are shown;
• look in your analytics for referral visits from the assistants’ domains;
• watch branded searches in Search Console, as a rise can be an indirect sign of growing awareness.

Remember that answers differ from one session to the next and from one country or setting to another. A single run proves nothing; a pattern over several months is informative. Keep notes, because you will otherwise forget what you saw.

What to avoid

Shortcuts are tempting in a new field, and most of them cost more than they return.

• Mass-produced pages. Hundreds of thin pages made only to catch questions can fall foul of Google’s spam policies, described in its spam policies documentation.
• Hidden text or instructions aimed at assistants. Text placed in a page to manipulate a model is an attempt at manipulation, not optimisation, and may be filtered or penalised.
• Invented reviews or statistics. They are wrong in themselves, and an assistant that repeats a false claim about you will damage your reputation.
• Promising results. Nobody outside the companies can tell you exactly why a source was chosen on a given day.

Sensitive subjects deserve extra care. On health, finance or legal topics, assistants and search engines are generally more cautious about what they show, and the rules for what you may claim are strict. Our article on advertorials for clinics and medical practices shows how those limits work in a regulated sector.

A realistic plan for the next quarter

If you want a simple sequence, this one is proportionate for a small team. In the first month, audit robots.txt, fix any blocking, and rewrite the openings of your five most important pages so that each starts with a direct answer. In the second month, align your company description across your site and profiles, add structured data where it is missing, and start your monthly question log. In the third month, review the log, identify the questions where you do not appear, and decide whether the gap is content on your own site or a lack of credible mentions elsewhere.

For a brand-new site, the order is slightly different, because the foundation has to exist before mentions can help. The article on link building for a new website sets out the first 90 days.

Next step

GEO rewards patience and accuracy. Make your own pages clear, keep the facts consistent and earn credible mentions at a natural pace. If you decide that published articles are part of that plan, you can plan a campaign with us, or simply write to us with your goal and we will tell you what is realistic.

Frequently asked questions