GeoResearch — Georgia's Research Reports, Rebuilt from Public Data by an AI

Overview

This is a running story, not a finished case. Georgia’s banks, think tanks and polling institutes publish a steady stream of research reports — weekly market watches, monthly employment trackers, quarterly real-estate monitors, survey notes. We asked a simple question: how much of that can anyone rebuild, for free, from public data, with an AI agent and a published prompt? And a second one: does it depend on which AI?

The answer so far: four of the eight report types the industry produces can be rebuilt end to end from open sources; five different AI models gathered every number correctly from public pages, and four of them wrote a professional survey note from it — one for eight cents; and the tests corrected our own work. Everything below links to the page it came from.

This page is the hub. Each release — a report, the prompt sequence that produced it, what it cost, and later the comparison with the official edition — gets its own page, listed at the bottom as it lands. Where the reference report is periodic, ours is published first, so readers can check whether the numbers match.

The map

Twenty organizations were audited from their own websites, not from directories: bank research desks (Galt & Taggart, TBC Capital), think tanks (ISET Policy Institute, PMC Research Center, EPRC, EGI), polling and market-research firms (GORBI, IPM, CRRC Georgia, ACT, ISSA) and several that turned out to publish nothing freely. What they publish sorts into eight report types, and the type — not the company — decides whether a report can be rebuilt:

The rule

Every report rests on publicly available information only. Official statistics APIs and downloads; public web pages read the way a browser reads them; a site’s own form used the way its page uses it. Never bypassing bot protection, logins or paywalls (we stopped at two Cloudflare walls and found other sources); never collecting personal data from listings; robots.txt and crawl delays respected. Each release states its sources and access method so anyone can check that nothing non-public was used.

The prompts

Four generalized prompts exist, each with a worked example run on live data on 21 September 2026:

1. Macro bulletin — NBG policy rate, reserves, FX; Geostat CPI, GDP, trade, earnings. Weekly.
2. Real-estate monitor — NAPR transactions, Geostat price index and permits, listing medians from a public marketplace. Quarterly.
3. Employment tracker — Geostat labour-force survey and earnings, Revenue Service salary statistics, a month-end vacancy count. Monthly.
4. Survey synthesis — Caucasus Barometer time series and cross-tabs (the example: EU and NATO membership support, 2009–2024).

A fifth, the tourism overview (prompt 05: GNTA visitor trips, NBG travel revenue, Geostat visitor spending; quarterly), was added on 7 November 2026 with its own example.

Each prompt is split into a numbered sequence: a capability check first (can your AI’s tools actually open an HTML page, a JSON API, a PDF, an XLS file — with proof), then gather steps that print every table with its URL, then a write step that may use only what was printed. Anything not opened is written as NOT GATHERED, never filled in.

The first result: AI testing AIs

Before publishing a single report we tested the survey prompt on other models. Ten consumer AI apps first, then five models through our own harness via an API, with one tool — fetch a URL on the source site — and every number scored against the live page the same day.

What the tests showed, in order of surprise:

• The models were not the problem. Every model that could open the pages copied the tables exactly. The failures were the platforms’ fetch tools: one could not read PDFs, one truncated pages, one had no network in its sandbox. Hence the capability check at the start of every sequence.
• They corrected us. Our own first gather had stopped two waves early on the EU question and eight waves early on the NATO question, and had the wording history wrong twice. The models opened what we had missed; we verified each correction on the source and kept it.
• Cost. A nine-wave survey note with methodology, gathered and written, cost between $0.08 (DeepSeek V4 Flash) and $1.23 (Grok 4.6) for a complete run; one model hit its $5 budget before writing. Across 145 tool calls, zero fabricated numbers.

The full study — method, per-model table, the raw transcripts with every tool call, our own defects found and fixed — is the first release page.

What it costs

Every release page states its fetch count, API cost in dollars, wall-clock minutes and the model used. The number is part of the result.

Run it yourself

Two ways. Option A: an AI agent with browsing or code execution runs the full sequence — capability check, gather, write. Option B: paste the input pack we publish with each report into any model and run the write step alone; you get the report without tools, and you can check our gather against the pages. The prompts, packs and transcripts are served from this site, linked under Sources.

Releases

• 23 September 2026 — this page, and the first release: Can five AI models rebuild a Georgian survey note from public data? — the reproducible test behind everything above.
• 27 September 2026 — Georgia Macro Bulletin, week of 21 to 27 September — the published prompt run on official data before the banks’ Monday bulletins; compared with Galt & Taggart’s Monday bulletin on 29 September, 5 of the 14 series both carry matched exactly and 4 more after rounding.
• 1 October 2026 — Georgia Employment Tracker, September 2026 — the published prompt run the morning after the month closed: 6,859 vacancies published on jobs.ge in September, with the Revenue Service salary block reported as not gathered.
• 3 October 2026 — Georgia Survey Synthesis Note: EU and NATO support, 2009–2024 — the note the model study tested, rebuilt from CRRC’s published tables: EU membership support 70% in the 2024 Caucasus Barometer against 57% for NATO, the widest gap between the two series in the record.
• 24 October 2026 — Georgia Real-Estate Monitor, Q3 2026 — the published prompt run the morning after Geostat’s Q3 price index appeared: new-build flats in Tbilisi up 5.6% year on year and the total index up 5.9%, the third straight quarter of faster growth, with the asking-price block reported as not gathered because the listing site is behind a bot-protection challenge.
• 7 November 2026 — Georgia Tourism Overview, Q3 2026 — the last report type, run on official data after the bank’s own issue was already out: 2,598,257 international visitor trips (+3.3%) and travel revenue of USD 1,864.4 mn (+11.9%), with the bank’s branded-hotel block reported as not reconstructable.
• Further reports are added here as they are published.

Could we build this?

We did. One AI operator, a static site, a handful of scripts, an OpenRouter key with a few dollars on it. No fine-tuning, no private data, no access anyone else lacks.

Sources

• Prompts: 01 macro bulletin, 02 real-estate monitor, 03 employment tracker, 04 survey synthesis — numbered sequence; worked example and input pack: survey note, inputs
• Model-test transcripts (every tool call and answer, page dumps omitted): Grok 4.6, Kimi K3; test log
• Geostat PX-Web API — pc-axis.geostat.ge
• National Bank of Georgia — nbg.gov.ge
• NAPR statistics — napr.gov.ge
• Caucasus Barometer, CRRC — caucasusbarometer.org