Finding a photo's location is not the same as obtaining a plausible place name. A reverse-image result may identify where a copy was published. EXIF may contain a coordinate that has been edited or inherited. An AI model may recognise a country while inventing an exact pin. A defensible result records how each claim was produced and stops at the level of precision the evidence can support.
This workflow is designed for ordinary travel, archival and verification tasks. It deliberately separates discovery from proof: first preserve the evidence, then generate candidates, then try to disprove them. Do not use it to identify a private person's home, routine or live location.
1. Preserve the best available file
Work from the highest-resolution version you can legally access. Save the original without editing and create a separate working copy. Record the source URL, acquisition date, dimensions and whether the image is a screenshot, social-media download or forwarded file. These details determine which later conclusions are possible.
Inspect EXIF, but keep metadata in its own evidence column. GPS coordinates are a strong lead when they come from an original camera file and agree with the scene. They are not self-authenticating: metadata can be removed, changed or copied. Conversely, missing metadata is normal after screenshots and messaging apps.
2. Write a clue inventory before searching
| Clue family | Examples | Typical precision |
|---|---|---|
| Direct text | Place names, route numbers, domains, transit operators | Country to street, but text may name a destination |
| Infrastructure | Road markings, driving side, bollards, poles, drains | Country or region |
| Built environment | Roof forms, setbacks, façade rules, street geometry | Region to exact viewpoint when geometry matches |
| Natural environment | Terrain, geology, coast, vegetation, sun | Climate zone to valley or coastline |
| Context | Caption, uploader, nearby images, claimed date | Useful lead, but independent of the pixels |
For every clue, write what it supports and what it cannot prove. A beer named after a city may be sold nationwide. A bus displaying an airport destination may be travelling away from it. A tropical plant can be ornamental. This small discipline prevents associations from becoming facts.
3. Generate candidates without asking for false precision
Use an AI photo location finder to turn the inventory into a short candidate set. Ask for observed clues, alternative explanations and confidence at country, region and city level. Save the complete answer. Do not ask for a single coordinate until the system can explain which fixed feature would make that coordinate unique.
Maintain at least one alternative. A useful working table contains candidate, supporting clues, contradictions and the next test that could reject it. If all candidates merely repeat the same famous-city bias, deliberately add a control location with similar climate or architecture.
4. Search text in layers
- Search exact distinctive phrases in quotation marks.
- Search partial text with the likely country or language.
- Try spelling variants, transliterations and OCR corrections.
- Search a logo or object crop separately from the full image.
- Record queries that failed; repeated unstructured searching is not evidence.
Reverse-image search is best for finding earlier copies, alternate crops and context. It may reveal a higher-resolution original with new clues. It cannot by itself prove the capture location, because articles often reuse stock or unrelated images.
5. Convert combinations of objects into map queries
A single church, bridge or supermarket is usually too common. A church within 100 metres of a curved tram line and a small square is much rarer. Search the relationship between objects. OpenStreetMap and Overpass are especially useful when visible objects correspond to mapped tags, but coverage and freshness vary by region.
[out:json][timeout:25];
(
nwr["amenity"="place_of_worship"](area.searchArea);
nwr["railway"="tram"](around:150);
);
out center;Treat this as a pattern, not a ready-to-run universal query. Define the area, choose tags that match the image and narrow progressively. A timeout means the request or server failed; it does not mean the location hypothesis failed.
6. Verify geometry rather than resemblance
Once a candidate district is found, match durable relationships: road curvature, intersection angle, window order, roofline, slope direction, distance between structures and the sequence of objects from the camera. Satellite imagery tests road and building geometry. Street-level imagery tests viewpoint. Historic imagery can explain changed shop names or demolished objects.
Resemblance is cheap. Many buildings share colour and style. Verification becomes strong when several independent fixed features align from one camera position and a plausible feature would be visible exactly where the photograph shows it.
7. Report the right level of confidence
- Country: national clues agree but local evidence is absent.
- Region: environment and infrastructure narrow the search.
- City: several clues support one urban area, but the viewpoint is unmatched.
- Coordinates: fixed geometry and camera position are independently verified.
Include alternatives, contradictions and the next verification step. If the evidence supports “central Siem Reap near Pub Street,” do not manufacture six decimal places. A broad, honest answer is more useful than a precise wrong one.
A reusable investigation record
Save the original source, file hash, visible clues, metadata, AI response, search queries, map links, rejected candidates, final claim, confidence level and unresolved questions. This case file lets another person reproduce or challenge the result and protects against hindsight rewriting.