One photo. No metadata. No caption. A wall, a strip of road, a sliver of sky. That's the entire case file — and somewhere on Earth there is exactly one spot that matches it. Manual geolocation is the discipline of finding that spot before the news cycle moves on.
Forget the marketing copy about "AI that finds anything." The verified geolocations that hold up in court rooms, in newsrooms and in war-crimes dossiers still come from analysts who treat an image like a crime scene — every pixel is evidence until proven otherwise. The tools have multiplied. The method hasn't budged.
The mental model: stop looking, start decomposing
Beginners scan an image looking for "the answer." Operators do the opposite. They break the image into a checklist of independent features and let each one shrink the search space, continent → country → city → block:
- Architecture and roof geometry — narrows continent and often country
- Vegetation type and density — narrows climate zone
- Road markings and signage standards — narrows region
- License-plate format, font and color — narrows country (and often year)
- Electrical sockets and power-line geometry — narrows region
- Language and script on signage — usually narrows to country
- Vehicle makes, models and bodywork — confirms market
- Sun azimuth, shadow length and time of day — narrows latitude
None of these alone is an answer. Stack five of them and you are often inside a 5 km radius. That is the methodology Bellingcat formalised a decade ago in its beginner's guide to geolocating videos, and it hasn't aged because the physics hasn't either. AI assistants like Picarta and GeoSpy accelerate the triage. They don't replace the chain of evidence.
The map stack: where the work actually happens
No serious geolocator runs on one tool. The map stack is layered on purpose, because each platform sees a different planet.
Google Earth Pro remains the workhorse for satellite imagery, historical layers and 3D terrain. Use the historical-imagery slider to walk a building backwards through time; use the ruler to measure shadow length; use the path tool to overlay road geometry on candidate locations.
Google Street View is the western default, but the date selector is non-negotiable. A single drive-past from 2014 can decide a case the 2023 imagery can't.
Yandex Maps and Panoramas is the answer for the post-Soviet space. As OSINT Industries puts it, investigators geolocating images from conflict zones or criminal marketplaces often achieve hits on Yandex that fail elsewhere. Two caveats from Bellingcat's toolkit: Yandex's panorama coverage inside Ukraine is largely frozen around 2011, so anything newer needs cross-checking, and Yandex's reverse image search is still better than Google's at matching architecture and signage in the same region.
Mapillary and KartaView are the crowdsourced street-level layers that fill the holes Google won't drive into. KartaView's coverage in Southeast Asia is particularly strong; Mapillary lives wherever amateur drivers have a phone and a dashboard mount. Both are open-licensed, both are fair game, both are quietly indispensable on rural roads.
Bing Maps Bird's-eye — underused. Oblique aerial views expose building façades that straight-down satellite never shows. When you are trying to match a three-storey shopfront, Bing often wins where Google shrugs.
OpenStreetMap and Wikimapia are the metadata layer. OSM tags every road, hydrant, mast and bench — query it with Overpass Turbo and you can ask the planet questions like "show me every roundabout within 30 km of this point with exactly five exits." Wikimapia adds amateur annotations: often messy, regularly the only place an obscure landmark is named in the local language.
Sun, shadows and the geometry of light
If you can see a shadow and you can guess a season, you can constrain a location to within a few kilometres. This is the technique that turned SunCalc into a verb in OSINT circles.
The mechanics are clean: the sun's position is two angles, azimuth and elevation. Pick a candidate location and a candidate time, and SunCalc returns exactly where shadows should fall and how long they should be. If a shadow in your image points north-east at 4 pm and the candidate spot puts it due south, the candidate is dead. Bellingcat's writeup on the technique remains the canonical reference, and tools like suncalc.org and ShadowCalc collapse the trigonometry into one click.
The change in azimuth across an hour varies by latitude. That means a single video showing a shadow rotating across a wall is its own latitude estimator. For mountain skylines, PeakVisor matches a horizon silhouette against a 3D model of every named peak on the planet. If the photo shows ridges and the photographer can't fake ridges, PeakVisor gives you the camera's pointing direction within a degree.
The AI layer: useful, not magic
Two tools matter here, and the rest is noise.
Picarta runs vision transformers trained on geo-tagged imagery and returns city- or country-level predictions with confidence scores. It will not solve a hard case. It will absolutely tell you in three seconds that an image is most likely Eastern Mediterranean rather than Latin America — and that triage alone saves an hour. The public API returns city, province, country, GPS coordinates and a confidence score per prediction.
GeoSpy, built by Graylark Technologies, plays in the same space and is now marketed around law-enforcement deployments. Pair it with Bellingcat's OpenStreetMap-based search tooling and the AI gives you a region while OSM gives you the matching geometry. Operators have written entire public workflows around that pairing.
Two ground rules for AI assistants. First, they never replace verification — every "high confidence" hit needs a satellite or panorama match before it leaves your draft folder. Second, they fail hardest in the places that matter most: low-data conflict zones, rural villages, generic interiors. Use them for triage, never for evidence.
Techniques the pros actually use
The toolset is half the job. The other half is knowing which trick to pull when. The shortlist that comes up in nearly every confirmed case:
- Skyline silhouette matching — feed the horizon shape into PeakVisor and read off the camera angle.
- Shadow-azimuth and length — convert sun geometry into latitude and time of day; eliminate candidates that don't fit.
- License-plate format — every country has a distinctive font, color, aspect ratio and digit pattern. A blurred plate still routinely leaks the country.
- Road-line standards — line widths, dash lengths and color combinations are codified by region. A road shoulder is effectively a passport.
- Power-line and antenna geometry — pylon spacing, insulator count, voltage class, antenna spacing. Telecoms hardware is fingerprint-able if you know what you are looking at.
- Mapillary corridor matching — pull the imagery sequence along a candidate road and step through it frame by frame; you are effectively comparing two videos.
- Reverse architectural detail search — feed unique window shapes, mosaic patterns or signage details into Yandex and Bing reverse image search. Google is the last place to try, not the first.
- The #geoconfirmed methodology — show the evidence chain publicly: original frame, candidate location, satellite overlay, panorama match. GeoConfirmed built an entire conflict-monitoring infrastructure on exactly this principle: every verification is reproducible by anyone with a browser.
The community: who to actually follow
Manual geolocation is a contact sport. The serious work happens on X (Twitter) and Telegram, in real time, with frame-by-frame discussions in the replies. A short list of accounts worth a permanent column in your dashboard:
- @geoconfirmed — the canonical conflict-geolocation aggregator, run by @gocha. Daily verified incidents from Ukraine, Gaza, Sudan and elsewhere.
- @bellingcat — the methodology benchmark; the resources section is the textbook.
- @benjaminstrick — Bellingcat investigator and trainer; tight, reproducible writeups of real cases.
- @aric_toler — long-form geolocation case studies, formerly Bellingcat.
- @sector035 — runs the weekly "Week in OSINT" digest; consistent teaching on shadow analysis.
- @nrg8000, @hatless1der, @ericdeshazo — operators who post puzzles and walk through their solutions in public.
- GeoGuessr World — yes, the game. The top players do in thirty seconds what most analysts manage in thirty minutes. Their YouTube explanations are a free masterclass in pattern recognition.
The reality check
Geolocation is not pattern-matching against a global database — that database does not exist, no matter what an AI startup's pitch deck says. It is hypothesis generation, hypothesis elimination, and one hypothesis surviving the cross-check.
The discipline that separates a confirmed location from "looks about right" is the same one Bellingcat published in 2014: every claim shows the evidence, every evidence chain is reproducible, every candidate location is killed before it's accepted. The tools change. The method doesn't.
If a workflow can't survive being posted publicly with the original image and the satellite overlay side by side, it isn't a geolocation. It's a guess.