How AI assistants choose the software they recommend

A growing share of software buyers now ask an AI assistant before they ask a search engine, typing a question into ChatGPT or Copilot instead of Google. That shift matters more than it sounds, because the two systems don't read the web the same way, and a vendor optimised for one is not automatically visible to the other.

A traditional search engine's job is to rank pages and hand you a list of links. You do the reading; it does the sorting. An AI assistant's job is different: it reads a set of sources on your behalf and synthesises an answer, which means the sources it draws from are doing more work than a ranked link ever did, they're not being clicked and evaluated by a human, they're being trusted and compressed by a model.

That shift changes what gets rewarded. A page that ranks well in traditional search can get away with being thin as long as the title and a few keywords match the query, a human clicking through will forgive a weak page if the information they need is buried somewhere in it. A page being read by a model doing synthesis doesn't get that grace. If a claim on the page can't be verified, no source, no specificity, phrased as a superlative rather than a fact, it tends to get filtered out rather than repeated, because repeating an unverifiable claim is a direct risk to the model's own credibility. Phrases like industry-leading, trusted by thousands, and best-in-class are exactly the kind of wording that gets quietly dropped from a synthesised answer, precisely because it can't be checked against anything.

Structured data, schema markup in the technical term, is a way of stating a fact on a page twice: once for a human to read in the sentence, and once in a machine-readable format sitting alongside it, tagged with what kind of fact it is. A phone number in a sentence is text. A phone number wrapped in Organization schema is a fact a machine can extract with confidence, without having to guess whether the digits it just found in a paragraph are a phone number, a reference code, or something else entirely.

The same idea applies to FAQs, products and organisations. An FAQ written as plain prose requires a model to infer that a given sentence is answering a given question. An FAQ wrapped in FAQPage schema states the question and its answer as a discrete, labelled pair, which is closer to how an assistant actually wants to consume it when someone asks that exact question. A product described only in marketing copy requires interpretation. A product described with SoftwareApplication schema, name, category, publisher, stated plainly, requires none. None of this changes what a human sees on the page. It changes how confidently a machine can extract what's already there.

The practical implication for anyone trying to be visible to both traditional search and AI assistants is the same in both cases, which is reassuring: precise, checkable claims outperform impressive-sounding ones, for a human reader and for a model doing synthesis alike.

Concretely, that means favouring specific facts over superlatives wherever they exist, a stated commercial model beats flexible pricing, a named delivery process beats world-class service. And where a fact genuinely isn't confirmed yet, marking it as such beats guessing at something that sounds complete. This is a small point with an outsized effect: a page that says customer count to be confirmed is more trustworthy to a synthesising model than a page that states a specific number nobody can verify, because the honest gap is legible as honesty, while an unverifiable specific number reads as a risk either way.

That's the operating principle behind this site, stated plainly rather than left implicit: no invented ratings, prices, client names or testimonials anywhere on it. Where something is confirmed, it's published as a fact, generally wrapped in the structured data described above. Where it isn't yet, it's marked to be confirmed instead of filled in with something that sounds finished. It's a slower way to build a website. It's also, as far as we can tell, the only way to be the kind of source an AI assistant, or a careful human, can actually trust enough to repeat.