COMM7035 Chap.3 Search and AI Answer Systems
Search and AI Answer Systems
Follow the source through the pipeline
A source does not travel directly from publication to visibility. It must be discoverable, fetchable, eligible for an index, competitive for serving, available to retrieval and useful enough to influence a response. At every gate, passage is possible without success at the next.
Pipeline reasoning turns a vague complaint such as we are invisible into a testable question about where the source stopped and what observation would prove it.
Search selection is not one operation
Conventional systems discover links and submitted locations, crawl accessible resources, interpret content, build indexes and select results for a query.
Ranking is conditional on the query, context and available corpus. Index presence therefore does not guarantee a prominent result, while absence from a result page does not prove a crawl failure.
Logs, inspection tools and controlled result records answer different questions and should not be collapsed into one visibility score.
Generated answers add retrieval and synthesis
A search-enabled answer system may reformulate a question, retrieve candidate documents or passages, rerank them and provide selected context to a language model.
The model then composes a response from that context and its other capabilities. This broad retrieval-augmented pattern helps explain why sources can enter the process without receiving a link, and why a citation may support only part of a sentence.
It is a conceptual model for diagnosis, not a claim that every product exposes or implements the same internal architecture.
Citation and representation differ
Citation asks which source is visibly attributed. Representation asks what the answer says and how the entity is positioned. A response can cite an owned page yet distort an important fact, or describe an organisation accurately while citing a third party.
Measurement should therefore preserve the answer text, visible sources and coding decision. Counting links alone discards the communication outcome the course cares about.
Use observed behaviour without inventing algorithms
Search and AI products are proprietary and change over time. A responsible analysis distinguishes documented mechanisms, conceptual models and direct observations.
It can state that a tested answer cited particular sources on a date, compare recurring source types and form a hypothesis about source selection. It cannot reverse-engineer a stable ranking rule from a small prompt sample. The practical response is reproducibility: fixed questions, recorded settings, timestamps, screenshots or exports and explicit coding rules.
What this chapter covers
- 01
Discovery, crawling, indexing and serving as separate gates
- 02
Query-conditional ranking and the limits of index evidence
- 03
Question expansion, retrieval, reranking and synthesis
- 04
Retrieval-augmented generation as a diagnostic model
- 05
Citation separated from factual representation
- 06
Owned, earned and third-party source roles
- 07
Reproducible observation without algorithm speculation
Worked example · free
Locate a failure without guessing the algorithm
- 3State what the exact-title result establishes about the early pipeline.
- 3Separate presence, owned citation and accuracy observations.
- 3Propose a reproducible next check and one alternative explanation.
Key terms
- Discovery
- The process by which a system first learns that a URL or resource exists.
- Crawling
- Automated fetching of a resource so its content and links can be processed.
- Indexing
- Storing and organising interpreted content so it may be considered for later requests.
- Reranking
- Reordering an initial candidate set using an additional relevance or quality procedure.
- Grounding Context
- Retrieved information supplied to a model to support the construction of a response.
- Source Selection
- The choice of documents or passages used, surfaced or attributed for a response.
Search and AI Answer Systems FAQ
Does a conventional ranking prove that an AI system can retrieve the same page?
No. The products may use different indexes, partnerships, retrieval timing and selection procedures. A ranking can establish conventional visibility for a defined query, while AI retrieval and citation require their own observations. Keep the result as useful context rather than treating it as downstream proof.
Why save the complete answer instead of only cited links?
Because the communication outcome includes claims, omissions, sentiment and positioning. A source list cannot show whether the organisation was accurately described or recommended. Saving the answer, prompt, date and visible citations allows another reviewer to reproduce the coding and challenge an interpretation.
What can a small prompt test legitimately support?
It can describe what occurred under the recorded conditions, reveal recurring differences and justify a follow-up hypothesis. It usually cannot establish a universal platform rule or causal effect. Strength comes from transparent sampling, repeated observations and qualified language, not from presenting a small test as exhaustive.
Assessment move
Draw the full pipeline from source existence to answer representation without looking at the chapter. Under each arrow, write one item of evidence that shows passage and one failure that could stop progress. Then take three public webpages and test an exact-title query, a broad conventional query and two fixed AI prompts. Save dates and source links. Code presence, owned citation and factual accuracy separately.
Finish by writing a six-sentence diagnosis in which every claim names its evidence layer. Circle any sentence that guesses at an internal algorithm and rewrite it as an observation plus hypothesis.
Working through Search and AI Answer Systems in COMM7035? Sia is AskSia’s AI Marketing tutor — ask any COMM7035 Search and AI Answer Systems question and get a clear, step-by-step explanation grounded in how COMM7035 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
