--- title: "You ask one question. The engine asks five. How AI answers are built" canonical: "https://www.symios.ai/en/blog/how-ai-answers-are-built" lang: "en" --- # You ask one question. The engine asks five. How AI answers are built Author: Luca Di Cesare · Published: 2026-09-28T08:00:00+00:00 · Modified: 2026-09-25T09:53:30+00:00 · Language: en **TL;DR** - An AI engine does not always search. Sometimes it answers from what it learned in training, sometimes it searches live. When it searches, it breaks the question into what are often called grounding queries: smaller searches, invisible to the customer, and rarely the same twice. - Each grounding query runs on an index, and the index is not the same for every engine. Some engines then fetch the pages live, with only seconds to spare per page; others read from what they have already indexed. - Then the engine reads passages, not pages. A page competes through its best passages for each grounding query, and most of what is on it never enters the answer. - A selection of sources, from a few to a few dozen depending on the engine, reaches the model that writes the answer. The brands it names are mostly the ones whose passages won those searches. If the sources are wrong, the answer is usually wrong too, and the model rarely notices. In a year, at Symios we have asked AI engines tens of thousands of questions and read what came back. We observed, hypothesised, verified, corrected, confirmed. What follows combines what the engines document about themselves with what we have seen, again and again, in their answers: the few seconds between a customer's question and the reply. One caution before we start. This is the typical mechanism. Every engine does it its own way, and few of the details are confirmed by the companies that build them. Where something is documented, we link to it. Where it is not, we say what we observe. ## Does the engine always search? No. An AI engine has two ways to answer. It can draw on what its model learned during training, or it can search live and build the answer from what it finds. Most engines decide case by case: a question about something recent or specific triggers a search, a general question often does not. [OpenAI describes this](https://help.openai.com/en/articles/9237897) for ChatGPT, and the user can force a search or leave the choice to the system. This matters more than it seems. When the engine answers from memory, the brand it describes is the brand as it appeared in the training data, months or years ago. Prices, ranges, ownership, stores: all frozen at a date nobody knows. When it searches, the picture can be current. Which of the two happens is not the brand's choice. ## What are grounding queries? When the engine searches, it does not search for the customer's question. It rewrites it into several smaller searches and runs them in parallel. Google calls this [query fan-out](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) and describes it openly: the question is broken into subtopics, and multiple queries are issued at once. Microsoft shows the same mechanism in its own webinars and calls those searches grounding queries. We use that name. Take "which sofa should I buy for a small living room". The customer typed one sentence. The engine may run "best sofas for small spaces", "two-seater vs loveseat dimensions", "sofa brands UK reviews", "compact sofa bed options", "small sofa delivery times UK". Nobody sees these. Nobody typed them. And they are the searches that decide which brands appear in the answer. Two things follow. First, these searches do not look like the keywords anyone optimises for. [SALT.agency compared](https://salt.agency/blog/microsoft-clarity-ai-citations-correlations/) the grounding queries Microsoft Clarity reports for a set of websites with the keywords those sites rank for: only 2% matched a tracked keyword exactly. Most were longer, more specific rewordings of the same topic. A snapshot, not a census, but the direction is clear. Second, they are rarely the same twice. Ask the same question again and the engine may fan out differently. [Researchers at the University of St. Gallen](https://arxiv.org/abs/2604.07585) measured it: the sources cited by identical prompts on the same day overlapped by only a third to two fifths. It is why Symios never measures once: one answer is a sample, not a result. ## Where does the engine look? Not in the same place. Each engine searches its own picture of the web: Google's AI features use Google's index, ChatGPT uses Bing plus an index of its own, Perplexity has built its own, Claude relies on a third party. The same question, on four engines, is searched in four different webs. They also differ in how they read. Some open the pages at the moment of the question and give each a few seconds, so a slow site is left out. Others read the copy they already hold, which may be days or weeks old. Engine Search behaviour Where it searches How it reads pages Sources cited per answer ChatGPT Decides case by case Bing plus OpenAI's own index Fetches live, seconds per page Around ten Google AI Mode / AI Overviews Search-native Google Search index From the index A few Gemini (app) Answers from the model by default Google Search index, when it searches From the index A few, when it searches Perplexity Always Its own index plus third-party data Fetches live Twenty or more Claude Decides case by case Third-party search index Fetches live A few Based on provider documentation and independent observation, September 2026. Exact pipelines are proprietary and change without notice. ## What does the engine read? Not pages. Passages. Once the pages are in, the engine does not judge them as documents. It splits them into passages and asks, for each grounding query, which passages answer it best. A page competes through its best passages. Most of the rest of the page, however good, never enters the answer. This is the part that overturns twenty years of habit. A long, complete, well-written page about sofas is not "the best page about sofas" to an engine. It is a set of passages, and only the ones that answer a specific search count. A short page that answers "two-seater vs loveseat dimensions" in its first sentences can beat the encyclopaedic page every time, on that query. Google has described [ranking passages rather than whole pages](https://blog.google/products/search/how-ai-is-powering-a-more-helpful-google/) since 2020; in AI answers the same logic decides what is quoted. ## Who gets named? From all the passages retrieved, the engine keeps a selection, from a few to a few dozen depending on the engine, and hands it to the model that writes the answer. That model works with what it has been given. The brands it names are mostly the brands present in the winning passages. Sometimes it adds one from memory, or from the conversation. Mostly, it names who was on the page that answered. Two consequences, both uncomfortable. If the passages contain a wrong price, a discontinued product, a store that closed, the answer will usually repeat it. The model that writes rarely notices, because it does not know how the sources were chosen. It trusts them. [OpenAI itself warns](https://help.openai.com/en/articles/9237897) that results and citations can be incomplete, out of date or wrong. And the link shown next to a sentence does not tell the whole story. In Symios measurements we see brands named with no source attached, and sources that never mention the brand they sit next to. The citation shows what the engine chose to display, not everything that shaped the answer. What a brand can actually see is the answer itself: which brands are named, in which order, how often, on which engine. ## What changes for a brand? A brand is no longer competing for a position on a page of results. It is competing to be the passage that answers a search it never saw, on an index it does not choose, inside an answer written by a model that trusts whatever it is given. That sounds like a loss of control, and partly it is. But it also has an edge that ranking never had. Ranking rewarded age, links, authority accumulated over years. Passages reward whoever answers the question best, today, in clear sentences. At Symios we have watched two-month-old domains become the most cited source in their category because they did exactly that. The answer is the only place where all of this becomes visible. That is where we measure, and where we optimise. Symios works on the answers, not on the pipeline. --- Luca Di Cesare, Founder of Symios. Twenty-five years in digital advertising, now working at the frontier of Generative Engine Optimization.