Generative Engine Optimization: being the source an AI answer quotes

What actually changed

Classic SEO competes for a position in a list of ten links. Generative engines do not produce a list — they produce an answer, and they name a handful of sources. The competitive unit moved from "rank for a query" to "be the passage that gets quoted".

That has a practical consequence most sites have not absorbed: an assistant does not scroll, does not click a tab, and does not wait for your bundle to execute. It fetches, reads and extracts. Anything that is not in the response body may as well not exist.

Prerequisite 1 — can they even fetch you?

Test this before anything else, because it invalidates all other work. Fetch your home page with each AI crawler user-agent and read the status code.

This is not a theoretical check. The host this site launched on returned HTTP 403 to GPTBot, ClaudeBot, PerplexityBot, Applebot, Amazonbot, meta-externalagent and Bytespider — at the edge proxy, before the server ever read a robots.txt that explicitly allowed all of them. Shared hosts and some CDN bot-management defaults do this silently, and nothing in your dashboard mentions it.

for ua in GPTBot OAI-SearchBot ChatGPT-User ClaudeBot PerplexityBot Applebot CCBot; do
  printf "%-16s " "$ua"
  curl -s -o /dev/null -w "%{http_code}\n" -A "$ua/1.0" https://example.com/
done

# Every line must read 200.
# Any 403 means you are invisible to that engine no matter what you publish.

Prerequisite 2 — is the content in the HTML?

Google renders JavaScript on a deferred queue. AI crawlers, as a rule, do not render at all. A client-rendered site is not slow for them — it is empty.

View source and count the words. If the number is a fraction of what you see on screen, that fraction is what an assistant reads about your business. Static generation or server rendering is the fix, and it is the same fix as for classic SEO, which makes it easy to justify.

Prerequisite 3 — does anything else corroborate you?

Answer engines are built to avoid confidently citing an entity they cannot verify. If the only source claiming you exist is you, a model has no basis to name you specifically — particularly with a common name, where several other people hold similar sites.

Corroboration means independent surfaces that agree on the same facts: profile pages, directory listings, code repositories, marketplaces, and links from sites you do not control. Two profiles is not corroboration. Ten consistent ones, with identical name, location and URL, is.

What makes a passage quotable

Once you are retrievable, extractability decides whether you are quoted. The formats that survive extraction share one property: they can be lifted out of the page and still make sense.

• A direct answer in the first 40–60 words under the heading, before any preamble
• Self-contained sentences — no "as mentioned above", no pronouns pointing at earlier paragraphs
• Specific, checkable facts: numbers, names, dates, versions, prices
• Question-shaped headings that match how a person would ask
• Tables, definition lists and numbered procedures, which extract far more reliably than flowing prose
• A clear author and a date, so the model can weigh currency and expertise

Atmospheric marketing copy is the opposite of all of this. "We craft bespoke digital experiences" contains no extractable fact, so it is never the sentence that gets quoted.

Structured data does double duty

Structured data was built for search features, and answer engines increasingly parse it because it removes ambiguity. A graph that states who published this, what they sell, where they operate and what it costs saves the model from inferring any of it from prose — and inference is where mistakes about your business come from.

Measuring it, honestly

There is no Search Console for AI answers. What is available is indirect, and worth setting up anyway because the alternative is guessing.

1. Server log filtering by user-agent — the only direct evidence that GPTBot, ClaudeBot and PerplexityBot fetched you, and how often
2. Referrer tracking for chat.openai.com, perplexity.ai and similar, which captures the visits that convert
3. A fixed prompt set: ten buyer questions, asked monthly across ChatGPT, Perplexity, Claude and Google AI Mode, with the cited sources recorded
4. Brand-mention monitoring, because a mention without a link still contributes to corroboration

Questions people ask about this

Is GEO different from SEO?

It shares most of its foundations — crawlability, rendered HTML, structured data, real content. It diverges on what wins: classic SEO optimises for a position, GEO optimises for being the extractable passage, which rewards direct answers and specificity over keyword placement.

Do I need to write for AI instead of for people?

No, and trying to is the failure mode. The formats extractors prefer — a direct answer up front, checkable facts, clear structure — are the formats impatient humans prefer too. Content written to game a model reads as it is, and is exactly what these systems are trained to discount.

How long does it take to appear in AI answers?

Retrieval-based systems, such as Perplexity and ChatGPT Search, can pick up new content within days of a crawl. Anything sourced from training data moves on a timescale of model releases, which nobody outside the labs controls.

Should I block AI crawlers to protect my content?

If your revenue depends on people arriving to read the page, that is a genuine trade-off. For a business selling services, blocking removes you from the answer at exactly the moment a buyer is asking who to hire.

Who wrote this

Anas Bin Masud builds e-commerce sites and does technical SEO for businesses in the UK, Canada and Pakistan — fifteen live client sites, six of them stores taking real payments. The examples in these guides come from those builds and from the audit that rebuilt this site, not from a content brief. More about how I work, or read the case studies.