How AI assistants choose sources

Many business owners think AI answers are random. They aren't. Behind an answer from ChatGPT or Perplexity lies a selection process you can track — one that has surprising similarities to how a search engine works.

In this article, I explain how that selection works in practice, which signals carry the most weight, and where most SME websites fall short. I test this monthly with set prompts for clients, so the examples come from real answers, not theory.

An assistant rarely searches using your exact words. 'Which SEO specialist helps my webshop in Utrecht?' becomes an internal series of search queries: 'seo specialist utrecht', 'seo webshop utrecht', 'seo specialist sme netherlands'. It searches for those queries, not your prompt.

That explains why classic keyword research is still the foundation. If you don't rank in the results for the underlying search queries, you won't feature in the rest of the process.

ChatGPT and Microsoft Copilot rely largely on Bing's index. Gemini and Google's AI Overviews use Google's index. Perplexity searches live and displays its sources alongside the answer. Claude works with search results and with what the model has previously learned.

The practical takeaway: Bing Webmaster Tools is not an afterthought. Many Dutch SME websites are well indexed in Google but only half in Bing, thereby missing the exact index ChatGPT pulls from. Submitting a sitemap and checking your indexation in Bing takes an afternoon and yields disproportionate results.

From the retrieved results, the model selects the text snippets it can use. These are passages that make sense on their own: an answer in two or three sentences, a price with its conditions, a step-by-step list, a table comparing two options.

What gets dropped: texts where the answer is scattered across multiple paragraphs, pages focused primarily on the provider rather than the question, and content hidden behind a cookie banner or a click.

A model recommending a business takes a risk: is this company genuine, and is it suitable? That risk decreases when multiple independent sources say the same thing. A Google Business Profile with genuine reviews, a Chamber of Commerce registration, a LinkedIn profile, a trade journal mention.

This is precisely where SMEs have the most to gain. The website itself is often fine; what's missing is the second and third source confirming you exist, what you do, and who you do it for. A handful of consistent listings does far more than ten extra blog posts.

Models link information by entity name. If your website says 'Websucces Marketing', the business directory says 'Websucces', LinkedIn says 'Lubna Ouarghi Marketing', and your business profile lists a different address, the picture fragments into pieces too weak to cite.

So create a single summary line that is identical everywhere: company name, the person behind it, the core service, and the service area. Use this exact line on your website, across your profiles, and in every listing. It is the cheapest form of AI optimisation out there.

There is no way to buy a placement in an AI answer, and there is no sign-up form. Tricks like hidden text or on-page prompts aimed at the model don’t work either; they are simply ignored or actively arouse suspicion.

What remains is solid craftsmanship: being discoverable in both indexes, stating facts explicitly, adding machine-readable structure, and ensuring third parties validate you. That is precisely the work that improves your rankings in Google, too.