GEO: getting cited by ChatGPT, Perplexity and AI Overviews

A growing share of searches no longer ends in a click. People ask ChatGPT, Perplexity or Google a question and get a written answer that cites three or four sources. If your site is not among those sources, you do not exist for that query, even if you rank first on the classic results page.

GEO, for Generative Engine Optimization, is the work of making a site readable and quotable by these answer engines. It does not replace SEO. It is a layer on top, with its own rules and its own measurements.

What GEO changes compared with SEO

SEO optimises for a ranking of pages. GEO optimises for passage extraction. That difference drives everything else.

A search engine returns a list and the user chooses. An answer engine composes text from fragments taken across several documents, then attributes sources. The unit that matters is no longer the page, it is the paragraph. An excellent article in which no paragraph stands on its own will be read by the machine and dropped at composition time.

Three practical consequences:

• Ranking first no longer guarantees a citation. Answer engines regularly reach outside the top three when a cleaner passage answers better.
• Traffic can fall while visibility rises. You are read and cited, but the user no longer needs to click.
• The brand becomes a measurable asset. Being named in an answer, even without a link, steers the rest of the conversation.

How an answer engine picks its sources

The exact mechanisms are not public and differ between engines. Three stages are nevertheless common to every retrieval-augmented architecture.

1. Access

The engine has to be able to read the page. That means its crawler is not blocked, the content is present in the served HTML, and the page responds quickly. Answer-engine crawlers are less patient than Googlebot and do not all execute JavaScript. A site whose content only appears after client-side hydration is, to them, an empty page.

2. Retrieval

The user question is turned into one or more queries, then passages are retrieved from an index. At this stage what gets compared is not your whole page but fragments. Hence the value of a structure where each section handles one question and only one.

3. Composition

The model writes the answer from the selected passages and attributes citations. It favours fragments that answer directly, carry a verifiable data point, and need no outside context to make sense.

The six levers that matter

Make content readable without JavaScript

This is the prerequisite, and it is where most sites disqualify themselves. Check what a crawler sees by fetching your page without running any script. If the text is not there, no other work will help.

Write self-contained passages

Every paragraph should survive extraction and still make sense alone. That means restating the subject instead of writing “it”, defining a term before using it, and putting the answer in the first sentence rather than the conclusion. A section heading phrased as a question, followed by a direct one or two sentence answer, is the format most often quoted.

Give verifiable data

Models favour fragments containing a number, a date, a name, a value. “The lead time is short” is not quotable. “The lead time is fourteen working days” is. If you state a figure, give its source and its date: that is what lets an engine reuse it without risk.

Structure your data

Schema.org markup will not magically place you in an answer, but it removes ambiguity: who publishes, when, on what topic, with what expertise. The useful types are few: Organization, Article or BlogPosting, FAQPage, Product, LocalBusiness. Link them through a stable identifier instead of redeclaring the company on every page.

Publish an llms.txt file

The /llms.txt file is a proposed convention, not an enforced standard, and adoption remains partial. The cost is nil and the value is real: it gives a plain-text description of what the company does, what it sells, and which pages matter. Treat it as an identity card written for a machine.

Decide which crawlers you allow

GPTBot, ClaudeBot, PerplexityBot, Google-Extended: each declares itself, and each can be allowed or refused in robots.txt. The calculation is simple. If your business sells advertising audience, blocking is defensible. If you sell a service and want clients, blocking means leaving the directory. For a services business, allowing is almost always right.

What to measure

GEO has a measurement problem: there is no Search Console equivalent for answer engines. Three indicators remain available without paid tooling.

The first is the most useful and the most tedious. It takes an hour to build: thirty questions a client would ask, asked monthly, in a spreadsheet. It is crude, and it is already more than most companies measure. If you would rather be handed the baseline, we produce it free of charge: see our packages.

Where to start

In order, because each step gates the next:

1. Check that your content is readable without JavaScript and that answer-engine crawlers are not blocked.
2. Ask your thirty questions and record who is cited today. That is your baseline.
3. Take the five pages that carry your offer and rewrite them as self-contained passages with dated data.
4. Add structured markup and the llms.txt file.
5. Ask the same questions again six weeks later.

None of this needs a marketing profile: these are technical decisions, taken by whoever writes the code. That is precisely the Forward Deployed Engineer role, which starts from the problem rather than the ticket.

None of this requires a rebuild. It requires precision, and some consistency in measurement. On a product still being built, these choices belong in the architecture rather than in a later pass, which is what we cover in our ninety-day delivery method.

The four engines do not behave the same

Talking about “answer engines” in the singular is convenient but misleading. The four main ones have different architectures, and what works on one does not transfer mechanically.

The practical consequence: classic SEO work remains essential for AI Overviews, while Perplexity and ChatGPT offer a way in for young or low-authority sites, provided the content is accessible and clean. That is where the opening sits for a smaller company starting out.

Five mistakes that cost you

Believing a block protects your content

Blocking GPTBot in robots.txt prevents crawling, not citation. An engine can still reuse your content if it finds it elsewhere, republished on a third-party site, without the link back to you. You lose attribution without gaining protection.

Confusing the crawling bot with the answering bot

At OpenAI, GPTBot feeds training, OAI-SearchBot serves search, and ChatGPT-User acts when a person explicitly asks to open a page. Blocking the first to keep training out while staying visible in search is coherent; blocking all three as a precaution removes you from the circuit entirely.

Optimising the home page rather than the answer pages

Answer engines rarely cite a home page, because it answers everything and therefore nothing. They cite the page that addresses the exact question asked. One page per question beats one page covering them all.

Burying the answer under the introduction

The usual narrative structure, context then development then conclusion, is exactly the inverse of what is needed. Put the answer in the first two sentences of the section, then expand. The extracted fragment will be the opening, not the closing.

Publishing numbers without a date or a source

A model choosing between two contradictory claims favours the one carrying a date and an origin. A bare figure is a figure nobody reuses, or worse, one that gets reused and attributed to someone else.

Building your question set

The measurement described above only works if the questions resemble what your clients actually ask. Here is the method we use, in about an hour.

1. Ten definition questions. What your business does, phrased by someone who does not know the industry vocabulary.
2. Ten comparison questions. Your solution against the alternatives, including the alternative of doing nothing.
3. Five selection questions. How to pick a supplier, on what criteria, at what price.
4. Five local questions. The same, with your city or region, if you serve a local market.

Each question goes to all four engines, and you record three things: are you cited, who is cited instead, and is the answer accurate. That third point is often the most instructive: a wrong answer about your company is fixed by publishing the correct information in an extractable format, not by complaining about it.

Repeat every two months. A single measurement says nothing; the gap between two measurements is what carries meaning.

Key points

• GEO optimises passage extraction, where SEO optimises page ranking. The unit that matters is the paragraph.
• A site whose content only appears after JavaScript runs is, to most answer-engine crawlers, an empty page.
• A fragment carrying a dated, sourced number gets reused more often than one describing the same thing with adjectives.
• There is no official console for measuring AI visibility. The workable method is a fixed question set asked at regular intervals.
• For a services business, blocking GPTBot or PerplexityBot means leaving the directory its prospects consult.