Georgia Real-Estate Monitor — Q3 2026: Tbilisi New-Build Flat Prices Up 5.6% Year on Year

Key takeaways

• Geostat’s residential property price index for new Tbilisi flats reached 163.1 in Q3 2026 (2020 average = 100), up 5.6% year-on-year (YoY) and 1.2% quarter-on-quarter (QoQ).
• Price growth has accelerated for three straight quarters: the total index rose 3.0% YoY in Q1 2026, 4.9% in Q2 and 5.9% in Q3.
• Detached houses have pulled ahead of flats: +7.0% YoY and +2.7% QoQ in Q3, against +5.6% and +1.2% for flats, having lagged flats in Q1 (+1.8% against +3.3%).
• Registered purchases in Tbilisi reached 62,670 by 22 October 2026, 89.5% of the 70,039 recorded for the whole of 2025 (NAPR, annual series).
• First registrations of flats and units in Tbilisi reached 39,388 by 22 October, 89.7% of the 2025 total of 43,928 — the public proxy for new supply reaching the register.
• Construction permits in Tbilisi stood at 2,653 for January–June 2026 covering 3.14 mn m², 59.4% of national floor space; the permits series has not been refreshed since 30 July 2026 and the nine-month update is due 30 October.
• Asking prices and rental yield are not reported this quarter: the listing site the script reads is behind a bot-protection challenge, which this project does not work around.

1. Prices

Geostat’s Residential Property Price Index (RPPI) for flats stood at 163.1 in Q3 2026 (2020 average = 100), up 5.6% YoY and 1.2% QoQ, after 161.2 in Q2 2026 and 154.5 in Q3 2025. The total index reached 166.2, up 5.9% YoY and 1.5% QoQ. The index covers the market of new residential property in Tbilisi and is measured in national currency, so these are GEL price changes, not dollar ones.

The three-quarter sequence is the quarter’s main fact. The total index rose 3.0% YoY in Q1 2026, 4.9% in Q2 and 5.9% in Q3; flats 3.3%, 4.8% and 5.6%. Each step is larger than the one before.

Detached houses and flats have changed places. Houses rose 1.8% YoY in Q1 2026, below flats at 3.3%; by Q3 they were at 7.0% against 5.6%, and their QoQ rate, 2.7%, was more than double the 1.2% for flats. The houses index now stands 72.7% above its 2020 average, against 63.1% for flats.

2. Asking prices by district — not gathered

No district table, no city-wide asking medians and no gross rental yield this quarter.

The script reads home.ss.ge. In this run that host answered HTTP 403 with a Cloudflare “Just a moment…” interstitial on both the flats listing path and robots.txt itself, so the robots file could not even be re-read before a crawl. This project’s hard rule is that bot protection is never circumvented, so the block is recorded as not gathered rather than worked around. The gross yield is derived from the two asking series and falls with them.

ss.ge/robots.txt itself still returns 200 and still disallows only /images, /user/* and the /en/ and /ru/ listing sections. The obstacle is a challenge on the home.ss.ge host, not a change to that file.

This removes the only non-official source from the monitor. What remains — the price index, the register and the permits workbook — is entirely official statistics. It also narrows the public picture in a specific way: the banks' reports take asking prices from myhome.ge, which blocked automation earlier; with ss.ge now challenged too, no asking-price series for Tbilisi is readable by script from a public page.

3. Transactions

The register’s public statistics are annual, with the current year as a year-to-date figure. Every 2026 row in this run carries an as-of date of 22 October 2026.

Purchases in 2025, the last full year, were 164,403 nationally (+4.7% YoY) and 70,039 in Tbilisi (+2.3% YoY). By 22 October 2026 the Tbilisi count stood at 62,670, which is 89.5% of the 2025 full-year total; the national count, 140,932, is 85.7% of its 2025 total. Those are ratios of two figures in the table, not annualised estimates — and no pace comparison with 2025 is possible from this series, because the public page carries neither a monthly breakdown nor a 2025 year-to-date figure at the same date.

First registrations of flats and units — the public proxy for new supply reaching the register — reached 39,388 in Tbilisi by 22 October, 89.7% of the 2025 total. The series peaked in 2024 at 48,939 and fell to 43,928 in 2025.

Monthly transaction counts are not on the registry’s public page, and its API accepts no parameters. The monthly counts in the banks’ watch reports come from register data that is not published there.

4. Supply pipeline

Permits fell in 2025: nationally 11,578, down 4.2% on 2024, over 11.40 mn m², down 5.3%; in Tbilisi 5,345, down 5.4%, over 6.42 mn m², down 5.2%. For January–June 2026 the workbook gives 2,653 Tbilisi permits over 3.14 mn m², which is 49.2% of the national permit count and 59.4% of national floor space — Tbilisi takes half the permits and nearly six-tenths of the area. Its share of floor space was 56.4% for full-year 2025.

Two limits on this block. The workbook holds full years plus the current part-year only, so there is no January–June 2025 column and no YoY for the half-year. And the file has not been updated since 30 July 2026: its latest period is January–June, one quarter behind the price index this monitor is built around. Geostat’s calendar lists the nine-month permits release for 30 October 2026, six days after this monitor.

5. Calendar

• RPPI, Q4 2026 — Geostat has published the index on the 23rd of the month following each quarter through 2026 (Q4 2025 on 23 January, Q1 on 23 April, Q2 on 23 July, Q3 on 23 October). The January 2027 calendar is not yet published.
• Permits, nine months of 2026 — listed on Geostat’s calendar for 30 October 2026. That will give the first nine-month figure and the first comparable period for a YoY reading.
• Register statistics — refreshed continuously. In this run the 2026 rows carried an as-of date of 22 October 2026 and were written on the morning of 24 October.

6. Sources

QoQ is the change on the previous quarter; YoY the change on the same quarter a year earlier. No figure on this page was entered by hand.

How it was made

Gather. The routine fired at 15:00 Tbilisi, confirmed that Q3 2026 was in Geostat’s price-index table, and then pulled every series itself: the three RPPI tables from the PX-Web API, the Q3 release PDF, the registry’s statistics endpoint, the construction-permits workbook, and Geostat’s own release calendar for the dates in part 5. The asking-price crawler was not run: the mandatory robots and terms re-check that precedes it could not be completed, because the host now answers that file with a bot-protection challenge. The collected figures were written to the input pack, one table per source, with the endpoint and the time of access.

Write. Part 2 of prompt 02 was then run on that pack and nothing else, producing the monitor reproduced above. The prompt offers two routes — Option A for a model that can execute code, Option B for one that can only read pages in a browser and be handed a pasted pack. This run used Option A for everything it gathered. The asking-price block is the one place where Option B would now get further than Option A: a person reading the listing site in an ordinary browser still sees the prices that a script no longer may.

Score gate. Every number in the monitor was matched against the input pack by script, and every pack figure was re-confirmed against its source in this same run — nothing was carried over from the September worked example, even where the figure turned out to be identical. Score gate: 82 figures checked against 5 sources, 0 mismatches. Four of the 82 were roundings or differences of pack figures rather than literal matches (three floor-space totals in millions of m², and the six days to the next permits release); each was recomputed from the pack and agreed. The price index was also cross-checked two ways: the API’s QoQ and YoY tables against the ratios of the levels, and all nine Q3 figures against the PDF release. They agree.

Run facts. 24 October 2026; Claude Opus 5; 36 fetches and one web search; about 80 minutes wall-clock from the first request to the published page; run inside a scheduled routine, so no per-run API bill.

One finding worth recording for anyone repeating this: Geostat’s PX-Web installation needs a trailing slash after a table’s .px name, or it answers any request — metadata or data — with an IIS-level “404 — File or directory not found” that looks exactly like a wrong address. And the folder listing’s updated field still read 14 July while Q3 data was already in the table, so that field cannot be used to tell whether a quarter has landed. The check that works is a data query, or the release itself.

Known limits

From the prompt’s own list:

• Monthly transaction counts are not public. The registry’s statistics endpoint is annual, with the current year as a running year-to-date total. This is why the public version of this report is quarterly with a dated snapshot, and not the monthly transaction watch the banks publish.
• The year-to-date figure moves. It is as of 22 October 2026 here and will be larger on any later reading, so it is not comparable with a full year except as a share of one.
• Asking prices are not transaction prices — and this quarter they are missing altogether, so the monitor carries no price-per-square-metre figure and no rental yield at all. The RPPI is an index of change, not a level in lari or dollars per square metre.
• The primary-market developer survey in one bank’s residential report — sales by developer, absorption rates — is not reproducible from public data and is not attempted here.

Added by this run:

• Asking prices and rental yield: not gathered. The listing host is behind a bot-protection challenge. Not worked around, by rule.
• The permits block is a quarter behind. Its latest public period is January–June 2026, from a workbook last updated 30 July 2026, and it carries no comparable half-year for a YoY reading. The nine-month release is due on 30 October 2026, six days after this page.
• The price index covers new residential property in Tbilisi only, in lari. It is not a whole-country index, not a resale index, and not a dollar one.
• No comparator was gathered. This routine never quotes or paraphrases the banks’ real-estate reports. They are named only in the numbers-only comparison described below.

What this release is

This is a one-off test of a published prompt, run on official data on the morning after Geostat’s Q3 price index appeared, and before the banks' quarterly views of the same quarter are out. The question is not whether Tbilisi property is a good buy. It is narrower and checkable: given one prompt written in advance and only public sources, how much of a professional quarterly property report can an AI agent rebuild, and where exactly does it fail? This quarter the answer includes a clean failure — a whole block lost to a bot-protection wall — which is a more useful result than four blocks of numbers would have been.

When the banks’ quarterly residential reports for Q3 2026 are public, a numbers-only comparison will be added to this page: their figures beside ours, series by series, with no commentary on their analysis. That comparison is the experiment’s result.

It is not a quarterly series. GeoResearch publishes each report type once, because the question is whether AI can reproduce that report at all — and a copy every quarter would add risk without adding knowledge. There will be no Q4 edition of this page.