What this release is

A one-off test of a published prompt, run the morning after the month closed. The question GeoResearch asks is narrow and checkable: can an AI agent rebuild a professional Georgian research report from public data alone? So the prompt is published, the input pack is published, and this page is the output — dated 1 October, about six weeks before PMC Research Center’s Employment Tracker for the same month is due. When that tracker is public, a numbers-only comparison will be added to this page, and that comparison is the experiment’s result. This is not a monthly series: the tracker is published once, compared once, and that is the whole of it. A monthly copy of someone else’s report would add cost and risk without adding knowledge.

Georgia Employment Tracker — September 2026

jobs.ge data for the full month 1–30 September 2026, counted 1 October 2026. Salary-band data (Revenue Service) not gathered this run. Labour-force indicators are quarterly: the latest published quarter is Q2 2026.

Changes are labelled MoM (month on month), YoY (year on year) and QoQ (quarter on quarter). Percentage changes are to one decimal; changes in a share are in percentage points (“pp”).

1. Persons receiving a salary

Not gathered. The Revenue Service’s statistics form could not be read in this run: it answers only a browser session carrying the page’s own antiforgery token, and the browser available here could not establish a trusted connection to rs.ge. No figure is given, and none is estimated. On the Revenue Service’s own calendar the block could at best have carried August 2026, since data for a month appear during the following month.

2. Vacancies on jobs.ge

Employers published 6,859 vacancies on jobs.ge in September 2026, counted on 1 October 2026 across all 17 of the site’s categories, job type “Job” only. Four categories account for two thirds of the month: Sales/Procurement is the largest with 1,390 vacancies (20.3% of the month), followed by Administration/Management with 1,202 (17.5%), Logistics/Transport/Distribution with 1,031 (15.0%) and Finance/Statistics with 934 (13.6%). IT/Programming, which the official-style tracker reports as a named category, is small on this board at 150 vacancies, or 2.2%. The smallest categories are Media/Publishing with 4 and Beauty/Fashion with 9. There is no MoM or YoY change here: this is the first full-month count in the series, and the only earlier run covers 1–21 September only.

The largest components of “Other” are Food (465), Medicine/Pharmacy (348) and General Technical Staff (281).

jobs.ge, job type “Job” only; 6,867 listings were live at the crawl, of which 6,859 were published in September 2026. Listings live about 30 days, so a month must be counted at its close. Counted 1 October 2026; no history before 21 September 2026.

3. Labour market indicators

The unemployment rate was 13.8% in Q2 2026, 0.6 pp lower QoQ and 0.5 pp lower YoY — the lowest of the three quarters shown. It fell while the labour force shrank by 14.9 thousand QoQ to 1,584.3 thousand, so the fall is not the product of more people working: employment was 1,365.2 thousand, 4.0 thousand below Q1 2026 and 1.1 thousand below Q2 2025. What did change inside employment is its composition — hired employees rose 19.2 thousand QoQ to 940.7 thousand while the self-employed fell 22.3 thousand to 422.3 thousand. Average monthly nominal earnings were GEL 2,389.3 in Q2 2026, 1.1% higher QoQ and 8.0% higher YoY.

Geostat labour force survey and earnings statistics, population aged 15 and over, total for the country. Q3 2026 indicators are not published yet; they are due about mid-November 2026. Queried 1 October 2026.

By economic activity, the highest Q2 2026 earnings were in information and communication at GEL 4,400.4 (+5.4% YoY), financial and insurance activities at GEL 4,074.1 (+9.2%), construction at GEL 3,496.6 (+0.8%) and professional, scientific and technical activities at GEL 3,304.8 (+10.3%).

4. Sources

No figure on this page was fetched manually.

How it was made

Two routines and two steps, all published, in the order they ran.

Step 0 — the count, the day the month closed. A separate cloud routine, openclaw-ge-georesearch-jobsge-monthly, runs at 07:00 Tbilisi on the 1st of each month and does one thing: run scripts/jobsge_count.py against jobs.ge’s own scroll endpoint, tally the listings by category and publish month, and commit the table to the repository. It stores counts only — never a job title, an employer name or a contact — and it honours the 5-second crawl delay jobs.ge asks for in its robots.txt. This exists as its own job because jobs.ge publishes listings, not counts, and listings expire after about 30 days: a month can only be counted at its close, so the count has to be taken before anything is written.

Step 1 — gather. Part 1 of prompt 03 names three blocks and where each lives: the Revenue Service’s statistics form for persons receiving a salary by band, the jobs.ge count for vacancies, and Geostat’s quarterly labour force survey plus its earnings release for the indicators. This routine, openclaw-ge-georesearch-employment-monthly at 10:00 Tbilisi, read the September table the 07:00 job had left, queried Geostat’s PX-Web API with explicit period codes, downloaded the earnings release and extracted its text, and attempted rs.ge the only way the prompt allows. The result is an input pack — every figure with its source, URL and date — published in full: 2026-10-01-employment-tracker-inputs.md.

Step 2 — write. Part 2 of the same prompt is given the pack and nothing else. It fixes the four sections, the table columns and the register, and forbids any number that is not in the pack. That is the tracker above.

How to run it yourself. Option A — give a model with browsing and code execution Part 1 and Part 2 together, and it does both steps, except that the jobs.ge count will give you only the trailing 30-day window unless you run it at a month’s close. Option B — paste our input pack under Part 2 in any model, with no tools at all; the writing comes out the same, only the gathering differs. Option B is the honest demonstration that the hard part of this work is tool use, not prose.

The score gate. Before this page was written, every number in the tracker was checked by script against the input pack, and every figure in the pack was re-confirmed against the source it cites in this same run — both PX-Web tables re-queried, the earnings PDF text re-read, the jobs.ge table re-read and re-totalled from its 17 category rows. Score gate: 77 figures checked against 4 sources, 0 mismatches. All 75 distinct figures in the tracker were found in the pack.

Run facts.

• Date: Thursday 1 October 2026, 10:00 Tbilisi, the morning after the month closed.
• Model: the site’s own operating agent — one general-purpose model with a shell, HTTP, a browser and PDF text extraction. No second model was used, and no human edited the figures.
• 17 HTTP requests during the gather and the checks that followed it, of which 3 failed — all three the rs.ge browser navigations. The jobs.ge crawl itself belongs to the 07:00 routine, not to this count.
• Wall-clock: see the commit; the gather, the score gate, the Georgian write-up and the proofread ran end to end in one session.
• Cost: run inside a scheduled routine; no per-run API bill. The only paid item on the page is the illustration.

Known limits

From the prompt’s own list, plus what went wrong this month:

• The salary block is not gathered. rs.ge answers only its own form, carried by a browser session with the page’s antiforgery token. In this run every browser launch failed at navigation with a certificate-authority error, because the egress path re-terminates TLS and the available browser build does not trust it. Getting past that would mean either turning off certificate verification or reaching rs.ge by some route other than its own form, and this series’ hard rule forbids both. A blocked source is reported blocked. It is worth being plain about the cost: this is the one block of the three that the public web does not hand over freely, and it is also the one an AI agent most needs a human-operated browser for.
• The jobs.ge series starts here. This is the first full-month count, so there is no month-on-month or year-on-year change to report, and none is inferred. The series’ only earlier run covers 1–21 September and is a partial month; it is not used as a comparison.
• A month counted at its close is a lower bound, not a census. Listings live about 30 days, so some vacancies published in the first days of September had already expired when the count was taken on 1 October. The figure is what was still visible on that date, which is also what makes it reproducible: anyone counting on 1 October gets 6,859.
• The quarterly block lags the monthly one by design. The labour force survey and the earnings release are quarterly, and Q3 2026 is not out yet, so a tracker dated 1 October carries Q2 indicators next to September vacancies. The official-style tracker has the same lag; it is a property of the data, not of the method.
• The official tracker may count more than we do. It may include listing types beyond jobs.ge’s “Job”, so the totals are not guaranteed comparable even when the method is. The comparison section, when it is added, will say which differences are revisions, which are listing-type choices and which are different viewing dates.
• Reproducibility is at the level of numbers and structure. Two people running this prompt on the same day should get the same figures and the same sections; the phrasing will differ.