What this release is

GeoResearch asks one narrow, checkable question: can an AI agent rebuild a professional Georgian research report from public data alone? The first release answered it for this prompt sequence by running five other AI models through it. This page is the other half of that experiment — the note itself, run by us today on the live pages, so that the thing the study scored exists as a published document rather than a test fixture.

This release is unusual in the series in one way: there is no official edition to beat. A polling institute’s topical note appears when its author decides it should, not on a calendar, so nothing is being raced and nothing is being waited for. The value is of a different kind — it shows what a topical survey note is actually made of, and that anyone can rebuild this one from CRRC’s public pages in half an hour, with no licence, no microdata and no subscription. It is published once. There will be no second edition.

EU and NATO membership support in Georgia, 2009–2024: what the Caucasus Barometer shows

All percentages are CRRC’s published weighted figures, rounded to integers. Differences between percentages are in percentage points (pp). Read from caucasusbarometer.org on 3 October 2026.

Overview

In the 2024 Caucasus Barometer, 48% of adults in Georgia said they fully support the country’s membership in the European Union and a further 22% rather support it — 70% combined, against 10% who rather or fully oppose it and 5% who did not answer. Combined support for NATO membership in the same wave was 57%, which puts the two series 13 pp apart: in the nine waves that carry both questions, the gap had never exceeded 7 pp, and in 2019 and 2020 it was 2 pp. Both series moved the same way since 2021 — “fully support” up, “rather support” down, outright opposition up — but only the EU series gained on balance, recovering to 70% from 63% while NATO slipped from 60% to 57%. The sharpest internal divide is by age: 58% of 18–34-year-olds fully support EU membership against 43% of those aged 55 and over.

Key figures

2024 distribution (EUSUPP): fully support 48%, rather support 22%, partially support / partially don’t 14%, rather not support 5%, don’t support at all 5%, DK/RA 5%.

Lowest combined support: 2015 (42%). Highest: 2020 (73%); 2024, at 70%, is the third-highest reading in the series, behind 2020 and 2012 (72%), and 1 pp above the first wave with the question (69% in 2011).

What changed between 2021 and 2024 is the shape of that support, not only its level. “Fully support” rose 14 pp, from 34% to 48% — its highest reading in the series, 7 pp above the 2020 peak and 9 pp above 2011 — while “rather support” fell 7 pp to 22%, its lowest reading in the series, and the middle category fell 6 pp to 14%. Opposition rose from 7% to 10% and non-response fell from 10% to 5%, the lowest in the series (it was 15% in 2015). The 2024 wave therefore holds both the series’ strongest endorsement and more respondents willing to state an outright position, in either direction, than any wave before it.

Breakdown by age, 2024

Full support is highest among 18–34-year-olds (58%), 12 pp above the 35–54 group (46%) and 15 pp above those aged 55 and over (43%). Combined support narrows that distance to 8 pp — 74% among 18–34, 72% among 35–54, 66% among 55+ — because older respondents more often pick “rather support” (23–26% against 16%). Opposition is flat across the three groups at 9–10%, so the age difference in this cross-tabulation is about intensity and about answering at all, not about opposition: non-response is 10% among the 55+ group against 3% in both younger groups, and that 7 pp accounts for most of the gap in combined support.

Comparison: NATO membership

The NATO question runs two waves longer than the EU question, from 2009. For most of the record the two series move together and close: both peak in 2010–2012 and again in 2020, both bottom out in 2015 (38% NATO against 42% EU), both recover from 2017, and in every one of the nine waves that carry both questions NATO support is the lower of the two — by 2 pp in 2019 and 2020, by 3 pp in 2011 and 2021, and by no more than 7 pp (2013) until this wave. In 2024 they separate: EU support rose 7 pp from 2021 while NATO support fell 3 pp, opening a 13 pp gap, the widest in the record. The internal movement is the same in both series — “fully support” up (NATO +6 pp, to 37%), “rather support” down (NATO −9 pp, to 20%, its lowest reading) — but on the NATO question outright opposition rose from 10% to 16%, where on the EU question it rose from 7% to 10%. These are distributions from two questions in the same questionnaire; nothing here identifies a cause for either path.

Methodology

Caucasus Barometer 2024 Georgia: 1,509 completed face-to-face CAPI interviews, 16 April to 13 May 2024, adults aged 18 and over, excluding populations living in territories affected by military conflict (South Ossetia and Abkhazia). Multi-stage stratified cluster sample: electoral precincts as primary sampling units from 2020 parliamentary election frames, allocated to three strata (capital cities, other urban, rural), urban and rural strata further split into four geographic quadrants; PSUs drawn probability-proportional-to-size, households by random walk from a predefined starting point, direction and step size, and the respondent within the household by Kish grid with no replacement. Sample size was calculated independently at a 95% confidence level for each stratum. Interviews were conducted in Georgian (89.8%), Azerbaijani (8.7%) and Armenian (1.5%); minimum response rate (AAPOR RR1) 23%; 1,224 interviews were completed at first contact, 238 on a second attempt and 47 on a third, and CRRC-Georgia re-checked 10% of them face-to-face after fieldwork. Weights adjust for selection probability (from precinct-level voter numbers) and for non-response, then post-stratify by gender, age and geography against the 2014 National Census of Georgia. The methodological report states no margin of error — it names a “targeted margin of error” only as one input to the sample size — so differences of 1–2 pp in this note should not be read as changes. The 2024 Georgian wave was self-funded by CRRC-Georgia; the survey has been supported since its inception by the Carnegie Corporation of New York. Earlier waves were fielded in the year of the dataset name, and the 2015, 2017, 2019, 2021 and 2024 waves covered Armenia and Georgia only.

Question wording and answer labels changed over the run, which matters for any comparison across waves. The EU question was “Please tell me to what extent would you support Georgia’s membership in the European Union?” in 2011–2012, “To what extent would you support Georgia’s membership in the EU?” in 2013 and 2015, an explicit 1-to-5 scale in 2017, 2019, 2021 and 2024, and in 2020 a read-out variant listing the six answers in the question itself. The NATO question was asked without a scale read out in 2009–2015, with the 1-to-5 scale from 2017, and in the same read-out variant in 2020. The middle category is labelled “Equally support and don’t support” in 2011–2015 and “Partially support, partially don’t support” from 2017; in 2009 and 2010 the three middle NATO categories are printed as the bare codes 2, 3 and 4. “Don’t know” and “Refuse to answer” are reported as one cell in some waves and separately in others, and the 2013 NATO page prints a DK row with no RA row at all. All figures are CRRC’s published weighted percentages.

Sources

Citation, in the form CRRC requests: The Caucasus Research Resource Centers. (2009, 2010, 2011, 2012, 2013, 2015, 2017, 2019, 2020, 2021, 2024) “Caucasus Barometer”. Retrieved from http://www.caucasusbarometer.org/ on 3 October 2026.

How it was made

Six numbered prompts, run in order in one session, each linked to its text in the published sequence. Nothing was read from an archive, a mirror or a text-extraction proxy: the site itself, or a stated gap.

Prompt 0 — the capability check, before any gathering. Four probes (A–D), one per source type, each requiring pasted evidence rather than a claim: A an HTML page, B a JSON API, C a PDF attachment, D a binary .xls. All four passed this run, which is why the methodology section below carries the five fields that live only in the PDF. The study found this probe to be the single most useful prompt in the sequence — ten consumer AI products failed the PDF probe in the same way, and the limit was each product’s fetch tool, not the model behind it.

Prompt 1 — which waves exist. Read from the “EUSUPP in other datasets” list on the 2024 page itself, discarding Armenia, Azerbaijan, the regional and time-series datasets and two non-Barometer surveys: nine Georgian waves for the EU question, eleven for NATO. No wave list was taken from memory.

Prompt 2 — the EU series. The frequency table, question text and answer labels copied from each of the nine wave pages, with the support sums added from the printed cells.

Prompt 3 — the NATO series and the age cross-tab. Eleven NATO tables and the 2024 EUSUPP-by-AGEGROUP cross-tabulation.

Prompt 4 — the methodology. Ten fields from the factsheet, then five more from the methodological report, since probe C had passed. The sequence forbids copying personal contact details: the factsheet prints a researcher’s e-mail address and it is not in our pack, which is a rule the study added after two models copied it.

Prompt 5 — write the note. Given the input pack and nothing else: no new fetching, no memory, every percentage carrying its wave, every change naming the two waves compared, and no causal claim beyond what a cross-tabulation shows.

The input pack is the full record of the gather, published as it was written: 2026-10-03-survey-synthesis-inputs.md. Our run of the write step is published beside it: 2026-10-03-survey-synthesis.md.

How to run it yourself. Option A — give Prompts 0–5 in order to any AI product with browsing, filling in today’s date at the top of Prompt 0, and compare each printed table with the URL beside it; this is the honest test, and the whole sequence needs HTML pages only unless you want the five PDF fields. Option B — paste our input pack under Prompt 5 in any model, with no tools at all; the prose comes out different and the numbers come out the same. The gap between A and B is the measure of what tool use adds: in this prompt type it is the entire difference between a note and a formatting exercise.

The score gate. Every number in the note was checked by script against the input pack, and every cell of the pack was then re-confirmed on the live page it cites, re-fetched for the purpose. Score gate: 172 cells checked against 23 pages, 0 mismatches. The gate caught two defects before publication, both ours: the note carried the interview-language split, which the first pack had not recorded, and it called 2024’s 70% the second-highest reading in the series when it is the third, behind 2020 and 2012. Both were corrected against the pages.

Compared with the first gather. All 134 table cells were also compared with the 21 September pack, which the model study scored against: not one cell has changed. CRRC revisions do happen, and the sequence assumes they can, so the comparison is run and reported rather than assumed; twelve days apart, these tables are identical, down to the site’s own “pardially” typo on the two 2020 pages.

Run facts.

• Date: Saturday 3 October 2026, from 12:00 Tbilisi.
• Model: the site’s own operating agent — one general-purpose model with a shell, HTTP and PDF text extraction. No second model was used, and no human edited a figure.
• 50 HTTP requests in total, 3 seconds apart, none failed: 26 in the gather, 22 in the re-confirmation pass, and 2 more during the Georgian proofread, probing whether the site’s Georgian locale serves these variable pages. It does not — both /ka/cb2024ge/EUSUPP/ and the NATO equivalent return 200 but render the ODA landing page — so the Georgian renderings of the question wordings on the Georgian edition of this page are ours, stated as such there rather than presented as CRRC’s own Georgian questionnaire. Every request went to caucasusbarometer.org except the two Geostat probes.
• Wall-clock: about an hour end to end, gather through proofread.
• Cost: run inside a scheduled routine; no per-run API bill. For the same sequence on other models, the study measured $0.08 to $1.23 per complete run — about fifteen times between cheapest and dearest — and one further run that spent $5.09 without finishing. Those figures are not re-measured here.

Known limits

From the sequence’s own list, plus what this run showed:

• This synthesis needs no microdata — but only because the published tables already carry the numbers. Every figure here is a weighted percentage CRRC has already computed and published. The moment a question needs a regression, a significance test, a multi-way cross-tabulation or any breakdown the site does not publish, the public pages stop being enough and the SPSS/STATA files on the downloads page become the route. That is the real boundary of this prompt type, and it is a boundary of the method, not of the model.
• Rounded integers, and no margin of error. CRRC publishes these percentages as integers, so support sums carry rounding error of a point or so, and the methodological report states no margin of error at all. Differences of 1–2 pp are therefore not findings. The 2024 EU–NATO gap of 13 pp and the 15 pp age gap are well outside that; the 1 pp between 2024 and 2011 is not, and is reported as a level, not a change.
• Wording and labels changed across the run. Four EU wordings, three NATO wordings, two different middle-category labels, and DK/RA sometimes one cell and sometimes two. Every one is listed in the methodology above because a time series assembled across such changes is not a clean instrument, and a reader should be able to see exactly where the seams are.
• Nothing here is explained. A survey distribution shows what people answered, not why. The note names no cause for the 2024 EU–NATO divergence, and a cross-tabulation by age identifies an association, not a mechanism.
• One cross-tab, one dimension. Age only, because that is what the sequence asks for. Settlement type, education, ethnicity and income are all published on the same site and none is in this note.
• Reproducibility is at the level of numbers and structure. Two people running this sequence on the same day should get these figures and these sections; the sentences will differ.