How AI search actually works

In classic search the user types a query, the engine ranks ten links, and the user decides where to click. In AI search that chain collapses: the user asks, the model writes one answer, and often nobody clicks anything. The decision point is no longer the ranking — it is the content of the answer.

One question, many queries: query fan-out

Modern AI search systems do not search for your question verbatim. They break it apart first. "What is the best accounting software for a small business?" quietly becomes several separate searches: software comparisons, user reviews, pricing, small-business-specific recommendations, alternative lists. This decomposition is called query fan-out.

The consequence: your visibility is decided in each of those sub-queries, not in the original question. If you show up in comparison queries but not in pricing queries, you make it into part of the answer at best — or none of it.

The model draws on two sources

• Training data: whatever was on the web up to the model's cutoff. Your presence here shifts slowly; it is your brand's long-term footprint.
• Live retrieval: real searches run at question time and pages fetched right then. Perplexity and ChatGPT's search mode lean heavily on this. It moves fast — new content can matter within days.

That distinction has a practical consequence: a page you publish today can surface quickly on platforms that retrieve live, yet leave answers grounded in training data unchanged for months. Knowing which platform does which keeps you from misreading your own measurements.

How a model picks a brand

Platforms do not publish their algorithms, but examine enough answers at scale and recurring signals emerge:

• Source diversity: a brand named across several independent, credible sources outweighs a hundred mentions on a single site.
• Contextual clarity: flat descriptive sentences like "X is an accounting tool" make it far easier for a model to file the brand under the right category.
• Answerability: content that answers the question directly, summarisable in one paragraph, is far more quotable.
• Technical access: no crawl, no citation. A page blocked in robots.txt never reaches the answer, however well written it is.
• Freshness: on retrieval-heavy platforms, dated and current content wins.