How does AI search work and how is it different from Google? What we tell clients

Clients usually put it this way: "I get Google. I know how to rank. What actually happens when someone asks ChatGPT?" We explain it the same way every time, because a clear mental model saves a lot of wasted effort later.

The short version

Google returns a list of pages ranked by relevance and quality. An AI search product returns an answer written by a model, with a few sources attached. The model does not know your website. It looks things up at the moment of the question, reads what it finds, and writes from that.

The steps, in the order they matter to us

Interpretation. The engine reads the question in context. People ask longer, more specific questions in AI tools than they type into Google, and they follow up.

Query fan-out. This is the piece clients find most surprising. The engine rarely searches for the exact prompt. It splits the question into several sub-queries and runs them. A question about the best accounting software for a small design studio might turn into searches on pricing, invoicing features, integrations and reviews. If your pages answer none of those sub-questions, you will not be retrieved, whatever you rank for on the head term.

Retrieval. The engine pulls pages from a search index or from crawler data. This lookup step is why the pattern is called retrieval-augmented generation. The model does not rely on memory alone.

Synthesis. The model reads the passages it has and writes an answer. It prefers text that is explicit, factual and easy to lift.

Citation. Some of the sources that were read get shown. We separate fetched from cited, because an engine can read many pages and show few. A page that is fetched but never cited is a different problem from a page that is never fetched.

Variation. The same prompt can give different answers on different runs. We never judge a result on one run.

What differs from Google, and what does not

The differences that change our work:

The unit of success. In Google it is a position. In AI search it is whether you are named and cited in the answer. Fewer people click through, so we judge success on inclusion, and on the traffic and conversions that do arrive.
Sources outside your site. Engines draw on review sites, forums, news and directories heavily. Our client's own pages are only part of the picture.
Query shape. Conversational prompts, sub-queries and personas replace short keywords.
Instability. Rank trackers give a stable number. Prompt trackers give a distribution.

What stays the same:

Crawlability. The engines still need to find and read your pages. Different bots do different jobs: some fetch on demand to answer a question, some feed a search index, some collect training data. Robots rules and site infrastructure that block the wrong ones can remove a client from answers.
Clear, accurate content. Pages that state things plainly do well in both worlds. Google's guidance to site owners says its AI features rely on the same search fundamentals.
Authority. Being referenced and trusted by other sites still helps.

How we measure it for clients

We run client programs on Promptwatch, and the stages above map onto its features.

Fan-out: query fan-outs per model show the sub-queries an engine ran for a prompt. The GEO Inspector Chrome extension, free and local to the browser, shows ChatGPT fan-outs and fetched versus cited sources. We use it in client workshops because it makes the mechanism visible in a few minutes.
Retrieval and citation: citation analytics down to page and domain level, Page Tracker for specific URLs, offsite mentions where a brand is named without a link, and Reddit, YouTube and LinkedIn reports.
Access: Agent Analytics logs real-time, IP-verified AI crawler visits by their real names, such as GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot, and shows a citation rate per page with failure explanations. Crawler logs start on Professional ($245 a month).
Outcome: visitor analytics attribute AI-referred traffic and conversions.
Fixes: content gap analysis and the Action Board give a prioritized list of what to publish or repair.

Explore is free (10 prompts, ChatGPT only), Essential is $95 a month, Business is $579, and agency plans are $199, $399 and $799.

Mistakes we see

Treating one answer as the truth. A single run proves nothing. Track a set of prompts over time.
Optimizing only the head keyword. If fan-out sub-queries go unanswered, the head term will not save you.
Blocking bots without knowing which. A blanket rule meant for training crawlers can also block the ones that serve live answers.
Ignoring third-party sources. If review sites and forum threads drive the answer, the fix may not be on your site.
Expecting a guarantee. Nobody can make an engine recommend a brand. We can show where a client stands and what is most worth fixing.

What we do first

For a client new to this, we run a set of buyer questions, show the fan-out and the cited sources, and check crawler access. The first workshop usually ends with a list of five or six concrete fixes: pages that answer missing sub-questions, a robots or CDN rule that needs changing, one or two third-party sources worth working on. We then re-measure in a month. The model in the head is simple enough to remember: the engine looks up, reads, writes and cites, and our job is to be easy to look up, easy to read and worth citing.