AI Search — AI Glossary

Reference details

Related terms

• RAG (Retrieval-Augmented Generation)
• Research Agent
• AI Summarization

How it works

1. The user's phrasing is rewritten into one or more retrieval queries, since a conversational question is rarely a good keyword query.
2. Those queries are issued against a web index, and the returned results are fetched so their actual content is available rather than only their snippets.
3. Fetched pages are chunked, ranked for relevance to the original question, and the top passages are assembled into a context window.
4. A model generates a prose answer from that context, with instructions to ground each claim in the supplied passages rather than in parametric knowledge.
5. Citations are attached by mapping spans of the generated answer back to the passages that supported them, which is what makes the answer auditable.
6. Freshness is handled at the retrieval stage rather than in the model, so a question about today's events depends on the index's crawl latency and not on a training cutoff.

Example: Why answer synthesis changes search economics

Price one answer. Retrieving 8 pages at 2,500 tokens each is 20,000 tokens, plus a 1,500-token instruction block, so 21,500 input tokens, and the answer itself runs 400 output tokens. At $3 per million input and $15 per million output that is 0.0215 x $3 = $0.0645 plus 0.0004 x $15 = $0.006, so about $0.0705 per query. Classic keyword search serves a results page for a small fraction of a cent, and the gap is what reshapes the business: at 1,000,000 queries a day, answer synthesis costs $70,500 a day, or 70,500 x 365 = $25,732,500 a year in inference alone. That number explains most of what the products do. Trimming retrieval to 4 pages at 1,200 tokens each gives 4,800 + 1,500 = 6,300 input tokens and a 250-token answer, costing 0.0063 x $3 = $0.0189 plus 0.00025 x $15 = $0.00375, about $0.02265 — so the trimmed configuration is $0.0705 against $0.02265, roughly 3.1 times cheaper. Every term in that reduction is a quality decision: fewer sources to corroborate against, shorter excerpts that may cut the supporting sentence, and a shorter answer with less room for qualification. The visible differences between AI search products are largely this trade being made at different points.

Commonly confused with

AI search is not RAG over a private corpus, although the pipeline is nearly identical, and the difference is adversarial. A private corpus is content you control, so retrieved text can be treated as reliable; the open web contains pages written specifically to be retrieved and quoted, so AI search is a RAG system whose retrieval layer is under active attack. That is why prompt injection through retrieved content is a live production concern here and a theoretical one in most internal deployments. It is also not a replacement for a search engine in the general case. A synthesised answer forecloses the browsing that exploratory questions need, which is why the products retain a link list beside the answer rather than replacing it.

When to use it

Reach for it when:

• For specific factual questions where synthesis across several sources is the work the user wants done.
• On questions requiring current information, since freshness is a retrieval property rather than a training-cutoff one.
• When citations are carried through, which is what makes a synthesised answer checkable.
• Where the per-query cost is justified by the answer's value, given it priced at about $0.0705 against a fraction of a cent for a results page.

Reach for something else when:

• For exploratory or ambiguous intent, where a ranked list lets the user redirect and one answer does not.
• Without treating retrieved web content as untrusted, since pages are written to be quoted and can carry instructions.
• At query volumes where the economics do not close, as 1,000,000 daily queries reached $25,732,500 a year above.
• For navigational queries, where the user wants a specific site and a synthesised paragraph is pure overhead.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Search — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path