AI for Site & Precedent Research

Early research is a mix of gathering, structuring and understanding, and AI helps most with the first two. Hand a capable assistant a site and a brief and it can produce, in one pass, a structured briefing: the broad climate character of the region, typical seasonal patterns, cultural and historical context, the kinds of regulations that usually apply to that building type, and a shortlist of relevant precedents to investigate. What would have been a scattered morning of searching becomes a single organised document you can react to.

Its real strength is structure and breadth. AI is excellent at turning a vague question into a checklist you had not fully written yourself - "what should I be asking about this site?" is a question it answers well, because surfacing categories of concern is pattern-work, not fact-work. It is also strong at synthesis: give it your own site notes, survey data, or a council document and ask it to organise, summarise and cross-reference them, and it stays anchored to material you supplied.

Where it is weak is precisely the part that looks most impressive: specific, checkable facts. Exact rainfall figures, precise setback distances, the year a building was completed, which architect designed what - these are the claims most likely to be confidently wrong. A useful rule from the start: let AI shape the questions and structure the search, but treat every specific number, name, date or rule it gives you as an unverified lead, not a finding.

It helps to name the three modes of help explicitly. AI can gather - pulling together context you would otherwise chase across many tabs; it can structure - organising that context and your own material into something navigable; and it can prompt your attention - reminding you of categories a thorough researcher would consider but you might forget under deadline. All three are about breadth and organisation, where the model is genuinely strong. None of them is about adjudicating truth, where it is not. Keep that split in mind and you will reach for AI at the right moments and hold it back at the dangerous ones.

A language model does not look facts up; by default it generates the most plausible-sounding continuation of your prompt. Most of the time plausible and true coincide, which is why it feels reliable. But when it does not know, the model does not fall silent or say "I am not sure" - it produces a fluent, specific, entirely fabricated answer. This is a hallucination, and research is where it does the most damage, because a made-up rainfall figure or a misattributed precedent looks identical to a real one and can quietly steer a whole concept.

The fix is not to avoid AI but to change what you ask it to be. A plain chat model working from memory is the riskiest configuration for facts. A search-grounded or retrieval-augmented (RAG) tool - one that actually queries live sources or a document set and cites them - is far safer, because you can follow the citation to the source. In 2026 the mainstream assistants offer this: web-search modes, tools that return links, and research features that assemble cited reports. Prefer these for anything factual, and always open the citations - a link is not proof that the linked page says what the AI claims.

Your default posture: assume nothing factual is true until you have seen it at a primary source. For climate, that means a meteorological dataset or a recognised climate resource, not the model's memory. For regulations, the actual authority (covered hard in the next lesson). For precedents, the architect's own documentation or a reputable archive. AI gets you to the door faster; it does not get to vouch for what is inside.

Good research prompting builds verifiability in from the start. Three moves matter most. First, ask for sources, not just answers - require the tool to cite where each claim comes from, and to flag anything it is inferring rather than sourcing. Second, separate fact from interpretation - ask it to label which parts are established facts (to be checked) and which are its own synthesis. Third, invite uncertainty - explicitly tell it that "I don't know" or "this needs verification" are acceptable answers, which measurably reduces confident fabrication.

Compare a weak research prompt with a framed one:

Weak:  "Tell me about the climate and building rules for Jaipur."

Better: "Using web search, brief me on the climate of Jaipur, India for a
         residential project. For EACH claim (temperature range, rainfall,
         humidity, prevailing winds) give the figure AND a source link.
         Separately, list the CATEGORIES of building regulation I should
         check with the local authority - do NOT state specific setback or
         FAR numbers, only what I must go and verify. Flag anything you are
         unsure of."

The second prompt does three things: it grounds the search, it forces citations for facts, and it deliberately stops the model inventing regulatory numbers by re-casting that part as a checklist. For precedents, prompt the same way: "suggest 6 precedents for a hot-dry courtyard house; for each give the architect, location and the specific move worth studying, with a source - and mark any you are not certain about." You then investigate each properly. The prompt is not fishing for finished answers; it is generating a well-organised, source-tagged map of what to confirm.

Two more habits sharpen research prompting. Iterate rather than expect one perfect answer - treat the first response as a draft to push on: "that figure has no source, find one or mark it unverified", or "you listed five precedents; for each, what is the single move worth studying and where did you confirm the architect?". The follow-up is where a vague first pass becomes a checkable one. Triangulate the claims that matter - for any fact that will shape a decision, ask the model to find a second, independent source, and be suspicious when it cannot. If two phrasings of the same question yield different numbers, that inconsistency is a strong signal the model is guessing rather than reporting, and the claim goes straight into the verify-manually pile.

Keep your prompts honest about what you actually want at this stage, too. Early research is about mapping the territory, not reaching conclusions, so ask for breadth and structure - "what are all the factors I should consider for this site?" - rather than premature answers. The model is excellent at widening your field of view; let it do that, then narrow it yourself with verified evidence.

Raw facts are not yet insight. Once you have verified material - checked climate data, real precedents you have actually looked at, your own site survey - AI becomes useful again in a different role: as a synthesiser and questioner. Feed it your confirmed notes and ask it to draw out implications: "given this climate and this site orientation, what passive strategies deserve early attention?" or "across these five precedents I have studied, what common moves address the courtyard-privacy problem, and where do they differ?" Here the model is reasoning over material you trust, which is far safer than asking it to recall facts.

This is also where AI can challenge your first read of a site. Ask it to argue the opposite of your instinct, to list what a sceptical planning officer or a cautious client might raise, or to name context you may have under-weighted - cultural, historical, environmental. It will not always be right, but it widens your field of view cheaply, and you remain the one deciding what matters.

Keep a clean separation in your own files between verified facts (with their sources), AI-suggested leads still to check, and AI interpretation. That discipline is what lets you move fast without ever presenting an unverified figure to a client or building it into a concept. Done this way, AI does not replace site research - it front-loads the gathering, sharpens the questions, and deepens the synthesis, while the responsibility for truth stays exactly where it belongs: with you.

There is also a quieter benefit worth naming: using AI this way makes your research legible to others. Because you have separated verified facts from leads and interpretation, a colleague, a reviewer, or your future self can see exactly what stands on evidence and what is still provisional. On a team, that transparency prevents the all-too-common failure where an unchecked AI figure travels from a scratch note into a report into a client presentation, accreting false authority at each step because no one remembers where it came from. Good research hygiene is not bureaucracy - it is what lets speed and trust coexist.