How AI Overviews, ChatGPT search and Perplexity actually select sources

Every one of these systems, despite different branding, follows a similar two-stage pattern: retrieval, then generation. When a query comes in, the system does not ask its language model to invent an answer from memory alone. It first runs a search-like retrieval step — often built on top of an existing search index (Google’s own index for AI Overviews, Bing’s index for ChatGPT search’s web mode, and a combination of its own crawler and Bing for Perplexity) — to pull a shortlist of candidate pages that appear relevant to the query. Those candidate pages are then fed into the language model as context, and the model composes an answer grounded in that retrieved text, attaching citations back to the specific pages it drew from.

This matters because it means you are not optimising for “the AI” in the abstract. You are optimising for two separate things in sequence: first, being retrieved at all (which behaves a lot like conventional organic ranking, because it typically is powered by the same underlying index), and second, being the source the model actually chooses to quote or paraphrase once it has your page in front of it. A page can rank well and still get skipped in the generated answer if a competing page states the same fact more directly, more concisely, or in a more extractable format.

Retrieval and grounding, in plain terms

“Grounding” is the term for tying a generated statement back to a specific retrieved document rather than letting the model answer from its training data alone. Grounded answers are generally more accurate and are what these products aim for, both to reduce hallucination and to give users a way to verify claims. From a practical standpoint, three things increase your odds of being part of the grounding set: being indexed and crawlable in the first place (robots.txt, canonicalisation and server response codes still matter enormously here), matching the retrieval query closely at the passage level rather than only at the page level, and containing the kind of self-contained factual statement that a model can lift cleanly without needing surrounding context to make sense.

Retrieval systems also weight freshness and update signals more heavily than static evergreen content assumptions suggest. Perplexity in particular favours recently crawled and recently updated pages for time-sensitive queries, which means a periodic content refresh cycle — updating statistics, dates and examples on your core pages — is not just a ranking-maintenance habit, it is an AI visibility habit too. Pages that visibly carry a “last updated” date, and that are actually kept current rather than just re-dated, tend to be favoured when a system has to choose between two otherwise similar sources.

Entity clarity and schema

Answer engines work with entities, not just strings of text. A page that clearly establishes who or what it is about, disambiguated from similarly named things, is easier for a retrieval system to match confidently to a query. This is where structured data still earns its keep even though Rank Math manages your schema output. Organization, Person, Article, FAQPage, Product and Review schema, applied accurately and consistently across your site, give models an explicit, machine-readable statement of entities and relationships that reduces ambiguity. Consistent naming, consistent NAP (name, address, phone) data, and a coherent internal linking structure between related entity pages all reinforce the same signal in a way plain prose alone does not.

It is also worth being deliberate about internal linking between your entity pages and your supporting content. A well-built cluster — a pillar page on a core topic linked bidirectionally to supporting FAQ, comparison and definition pages — gives both traditional crawlers and AI-oriented crawlers a clear map of how your entities relate to each other. This reduces the chance that a retrieval system pulls an outdated or thin page from your site over a stronger one you have already built, simply because the stronger page was harder to discover.

Content structures that actually get extracted

Animated flow: query to citation

Brand mentions and citation building

Answer engines weigh reputation signals that look a lot like traditional digital PR and link building, but the target is broader than backlinks alone. Unlinked brand mentions on reputable third-party sites, inclusion in “best of” and comparison roundups, coverage in trade press, and consistent presence on review platforms and Q&A sites (Reddit and Quora threads are frequently retrieved by Perplexity and appear in AI Overview source sets) all contribute to a model’s confidence that your brand is a credible entity to cite on a given topic. This is not a reason to abandon conventional link building; it is a reason to widen the definition of what counts as a citation-building activity to include earned mentions that never carry a hyperlink.

Measuring AI visibility

Implementation checklist

What not to do: scaled content abuse

The temptation with AI-driven search is to flood a site with large volumes of thin, templated pages designed purely to catch long-tail queries, on the assumption that more surface area means more citation opportunities. Google explicitly classifies this as scaled content abuse and treats it as a spam violation regardless of whether the content was written by a human or a model. The same principle applies with answer engines more broadly: low-effort, interchangeable pages rarely get cited because they lack the specificity and clear entity grounding that retrieval systems reward, and they carry real risk of manual action or ranking suppression. The durable approach is fewer, deeper, more clearly structured pages that a model can extract from confidently, not more pages.

Measurement discipline matters more here than in traditional SEO because the feedback loop is noisier. Rather than treating a single missed citation as a failure, track directional trends across a fixed sample of thirty to fifty priority queries checked on a consistent schedule, and log whether your brand appears, whether it is cited by name, and how accurately the summary reflects your actual offering. Misrepresentation in an AI-generated answer, such as outdated pricing or a discontinued service, is worth correcting on-page immediately, since these systems tend to re-surface the same phrasing across multiple queries once it has been grounded once.

Frequently asked questions