How a website becomes a source cited by ChatGPT and Google AI Overviews

What generative search changes for a small business

For twenty years the game was simple: show up in the list of links, then convince the visitor to click. Generative search rewrites that contract. ChatGPT, Perplexity and Google AI Overviews compose an answer directly, and the visitor only sees your website if the engine cites it as a source.

For a small business, this creates an unexpected opening. Generative engines do not rank sources by brand fame: they look for the page that answers the question most precisely. A local company that explains its subject clearly can be cited where a large player with vague content will not be.

This discipline is called GEO, for generative engine optimization. It does not replace classic SEO, it extends it: the same pages now have to convince a human reader, an indexing crawler and a language model.

How AI engines choose their sources

When a generative engine receives a question, it usually starts with a classic web search. It retrieves a batch of candidate pages, extracts the relevant passages, then writes its answer based on those passages. The pages kept as sources are the ones where a single excerpt answers the question directly, with no detour.

This means the unit of competition is no longer the page but the passage. A self-contained paragraph that states a clear claim and backs it up has a better chance of being quoted than a long text where the information is diluted. Engines also favour pages whose author is identifiable and whose claims are consistent with what they read elsewhere.

Freshness matters, but less than reliability. A stable, precise, dated page that says who is speaking and where the information comes from earns more trust from a model than an anonymous page updated yesterday.

The content structure that makes a page quotable

A quotable page reads like a good answer, not a brochure. It states its subject in the title, answers the main question in the first paragraphs, then develops. The hurried reader and the language model do the same thing: they look for the answer at the top of the page.

Each section must stand on its own. A subheading phrased as a question or a claim, followed by a paragraph that fully answers it, forms a block the engine can lift as is. Vague wording, padding and content-free promises are exactly what models discard.

• One subject per page, stated in the title and the first paragraph
• Subheadings that mirror the questions your customers actually ask
• Self-contained paragraphs that make sense out of context
• Precise claims rather than cautious generalities
• An FAQ with short answers that can be quoted verbatim

Structured data that speaks to AI engines

Structured data, written as JSON-LD using the schema.org vocabulary, describes your page in a language machines read without ambiguity. It states explicitly: this is an article, written by this person, published on this date, by this organisation. The engine no longer has to guess.

For a small business website, a few types are enough. Organization and LocalBusiness establish who you are, Article or BlogPosting describe your content, FAQPage links each question to its answer. That last type is especially useful in generative search, because it delivers ready-made question and answer pairs.

The absolute rule: structured data must describe what is actually visible on the page. Markup that announces content which is absent or different erodes trust instead of building it.

Making your site accessible to AI crawlers

Before citing a page, an engine has to be able to read it. Generative search crawlers, such as GPTBot for OpenAI, PerplexityBot or Google-Extended, obey your site’s robots.txt file. An overly broad block, often inherited from a default setting or a security plugin, is enough to make you invisible.

Rendering matters as much as permission. Content that only appears after JavaScript runs, behind a click or inside a popup, is out of reach for some of these crawlers. Important text must be present in the HTML served on first load.

Running our own platforms in production taught us to verify this in the server logs rather than trusting the configuration: the logs show which crawlers actually visit, what they request and what they are served.

Classic SEO and GEO work together

GEO is not built against SEO, it is built on top of it. Generative engines almost always start from a classic search index: a page that ranks poorly there has little chance of entering the candidate set the model examines. The groundwork stays the same, fast pages, sound internal linking, genuinely useful content.

The difference lies in the goal. SEO optimises for a click, GEO optimises for a citation. The first polishes the titles and snippets shown in the results list, the second polishes the passages a model can reuse in its answer. A well-built page serves both goals at once.

Measuring your visibility in AI answers

Measurement is the weak spot of GEO: there is no equivalent of Search Console for generative answers yet. You have to combine several partial signals to form an honest picture of your presence.

The simplest method is still the manual test: ask the engines the questions your customers ask, in their words, and note which sources get cited. Repeated regularly with the same questions, the exercise reveals trends, who is gaining ground, who is losing it, and where you are absent.

• Ask ChatGPT, Perplexity and Google your customers’ questions and record the cited sources
• Track referral traffic from answer engines in your analytics
• Spot AI crawler visits in your server logs
• Watch for brand mentions inside answers, even without a link

The mistakes that make a website invisible to AI

The first mistake is accidental blocking: an overly strict robots.txt, a firewall rejecting unknown bots, a host filtering by default. The site works perfectly for humans and does not exist for answer engines. Nobody notices, because nothing visibly breaks.

The second is substance-free content. Pages that talk about you rather than to your customers, texts that skim ten subjects without covering one, claims nobody can verify. A language model looks for answers: a page that contains none will never be cited, whatever its design.

The third is inconsistency. A company name spelled differently across pages, information that contradicts what directories say, missing dates. Engines cross-check sources, and a blurry entity is an entity nobody cites.