Discoverability through generative search interfaces: a practical primer
Generative search interfaces are converging on a single mode of answer construction: ChatGPT search, Perplexity, Claude, Copilot and Gemini all behave similarly enough to talk about as a category. The implications for SMB websites are concrete, but the literature has not yet caught up. A short orientation.
A pattern is now stable enough to talk about cleanly. The generative search interfaces (ChatGPT search, Perplexity, the search modes of Claude, Microsoft Copilot, Google's AI Overviews and Gemini) have converged on a similar approach to answer construction. They retrieve, they read, they reason, and they cite. The cited sources are surfaced to the user, sometimes as inline footnotes, sometimes as a panel beside the answer.
For a small or mid-sized professional services firm, the practical question isn't whether to engage with this layer of the search market. It is whether the firm is currently discoverable in it, and if not, what to do.
This primer sets out, in plain terms, how generative search interfaces actually construct answers, where the SMB website fits in that process, and which characteristics of a site materially affect the firm's chances of being included in the cited material.
The shape of a generative search answer
A generative search answer is, in broad terms, the output of a retrieval-augmented generation pipeline. The user's query is parsed and rewritten, often into multiple sub-queries, and those queries are issued against a search index or a live web fetch. The retrieved material is then summarised and synthesised into a natural-language answer by a language model, with citations inserted to support specific claims.
Three properties of that pipeline matter for the SMB.
First, the model is selecting from a candidate set, not from the entire web. The candidate set is, in practice, a few dozen pages at most. A firm whose pages aren't in that candidate set will not be cited, regardless of the quality of its content. The question of whether the firm is in the candidate set is the question of whether the firm has been indexed and surfaced by the underlying search layer.
Second, the model prefers material it can quote. Pages that contain clear, declarative sentences with specific facts are noticeably more useful to the model than pages that hedge, qualify, or generalise. The model is looking for a sentence it can pull into the answer without the user having to question it.
Third, the model prefers material that is internally consistent with the other candidates. A firm whose own pages contradict each other, or whose own claims contradict third-party sources, is harder to cite. The model would rather cite a page whose claims are corroborated than a page whose claims are contested.
What the model looks for, in practice
Three signal categories drive inclusion in the cited material.
Authority signals. The model is, in effect, looking for evidence that the firm is the kind of source a trained researcher would cite. The supporting signals are the usual ones (a credible domain, an author with verifiable expertise, backlinks from other recognised sites, structured data declaring who the firm is), but the weighting is different from classical search. The presence or absence of schema.org/Organization or schema.org/ProfessionalService markup is, in my work, the single signal most strongly correlated with whether a firm is cited.
Topical depth signals. The model is looking for a site that clearly does this work, not a site that mentions this work. A firm whose homepage lists eight services but whose underlying pages are 200-word marketing summaries is harder to cite than a firm whose service pages are 1,500 words each, written with specificity. The model needs material to quote.
Recency signals. The model is, in most cases, biased toward recent material, particularly for queries that touch on regulation, market conditions or fast-moving technology. A firm that last updated its blog in 2022 is, in this layer of the search market, structurally disadvantaged.
The first two categories are addressable in a single engagement. The third requires a sustained publishing cadence. That's the reason the Authority package exists, and the reason an implementation roadmap treats publishing cadence as a separate workstream from the foundational structured-data work.
Why most SMB websites aren't in the candidate set
When I start an engagement, the most common finding is that the firm's website is structurally invisible to the generative search layer for reasons that have nothing to do with content quality. The frequent culprits:
- The site is built on a template platform that doesn't expose schema.org markup, or exposes it in a way that's inconsistent with the firm's actual positioning. See the vendor-by-vendor assessment.
- The content is too thin. Pages average 250 to 400 words and contain no quotable, specific claims.
- The site has no published canonical answers to the questions a prospect would ask. The FAQ page is absent or generic.
- The site has no
llms.txt, leaving the model to construct a summary from whatever it can find. See Understanding llms.txt. - There's no FAQ schema on the pages that do answer questions, which makes those answers hard for the model to extract reliably. See FAQ schema.
Each of these is independently addressable. None requires a rebuild of the site, though in some cases a rebuild is the faster route.
The longer-term consideration
The generative search layer is, at the time of writing, still maturing. The mechanisms by which individual models select and weight sources will continue to change. Any specific tactic described in this note is, accordingly, subject to revision as the field develops.
The underlying observation, however, is stable. The search market is shifting from a model in which the user picks from a list of links to a model in which the AI picks one or two sources to cite. For a firm whose business depends on being found at the moment a prospect is researching, the implication is that the firm now has to be the kind of source the AI will choose.
That's a higher bar than being the tenth result on a Google search results page. But it's also a more durable position. A firm that is cited as the canonical source on a question is, in practice, the firm the prospect contacts.
The work of getting there is methodical rather than mysterious. The next pieces in this series cover the audit framework, the six signals AI assistants evaluate, and the implementation roadmap. The services line covers delivery.