What GeoSeer and GeoAxis Report That Oceanir Does Not

GeoSeer and GeoAxis both publish more numbers about their own tools than Oceanir does. They give an accuracy figure within one kilometre, a median time for each mode, a cost per search and a count of cities with street-level coverage. Oceanir publishes prices, and benchmark tables for its Manta model on public datasets, and not much else. That is a fair criticism. This page answers it directly: what they report, what each number is worth, and what Oceanir can say today.

On GeoSeer's own table, Oceanir is third on accuracy, first on speed and last on cost, and the cost figure uses prices that changed on October 29. This page does not treat that single-pin test as the measure for Oceanir. Oceanir's published evidence is Manta on shared public benchmarks. The GeoSeer vs Oceanir comparison goes through GeoSeer's results measure by measure.

Research benchmarks report accuracy at fixed distances, so the number says what it means: the share of photos placed within 1 km (a street), 25 km (a city), 200 km (a region), 750 km (a country) and 2,500 km (a continent), plus the median error in kilometres. Anyone can compare two models by reading across the row.

Read across, a vendor row with a single number at 1 km tells you the least a result can. It says nothing about how far off the misses are. A tool that lands within 1 km on most photos and is wildly wrong on the rest looks the same in that one number as a tool that is a few kilometres off on the rest. The ladder and the median error show which one you have.

A named benchmark also has a published range. On Im2GPS3k, the other non-pipeline models in Manta's table score between about 8% and 23% within 1 km. Picarta reports 44.8% on the same set and says its conditions differ from other methods, so that figure needs its protocol before anyone compares it.

Every one of these numbers is useful and every one has a limit. Read the limit before you read the number.

• Accuracy within 1 km. It depends on the photos. On GeoSeer's 62 images, 94%, 87% and 79% are roughly 58, 54 and 49 correct. With only 62 images, the 95% intervals are about 85 to 97, 77 to 93 and 67 to 87. They overlap, so this test does not settle the order of the top three. It also scores one threshold, so it cannot tell a pin on the right block from one 900 metres away.
• The same tool scores differently on different sets. Picarta reports 44.8% within 1 km on IM2GPS3k, and says the comparison with other methods may not be entirely fair. GeoSeer's page puts Picarta at 62% on its own set. Neither need be wrong. They measure different photos.
• Speed. A median hides the slow end, and each vendor times a different mode. A fast mode and a high-accuracy mode are different products. Ask for the speed of the mode that produced the accuracy figure.
• Cost per search. It depends on the plan, the mode and how much you use. GeoSeer averages its modes, and averages Oceanir's three depths equally, as if one search in three were a detailed analysis.
• City counts. An indexed city means there is street-level imagery to match against. It says nothing about accuracy in that city, or outside it.
• Who ran the test. A number a vendor measures about itself is a claim. A number someone else reproduces is evidence. A number one vendor measures about another vendor's tool is a claim about a tool the measurer does not control, and the tool's owner cannot check it without the logs.

Oceanir's prices are public, and they are in credits. A basic check costs 1 credit, a standard check 2 and a detailed analysis 10. This is what that comes to.

Two comparison pages are out of date for that reason. GeoSeer's cost comparison prices Oceanir Pro at $29 for 100 credits and averages all three depths to about $1.30 a search. On the current allowance, a Pro subscriber who uses all of it pays about $0.10 for a detailed analysis. GeoAxis's pages list Starter at 30 credits and Pro at 100, and give 102 cities from March 2026 as Oceanir's latest coverage.

Oceanir's published evidence is Manta on shared public benchmarks. It does not publish a single-pin accuracy score for the app's modes, a latency table or a current street-level city count. Manta's benchmark tables are for a different product, the organization model, and should not be read as a score for the app's detailed analysis.

A distance table scores one output. Most of what people use Oceanir for is not in it.

• Three depths. A broad estimate, candidates with the visual evidence, and a detailed analysis with the supporting and contradicting evidence and the alternatives it considered.
• Evidence you can review. The leading candidate and its location precision, the stated clues beside the original image, alternative candidates and contradictions.
• Cases. Related photos and notes kept together in Work, with unlimited cases on Pro and saved places on your map in World.
• Photo queues and claim checks. Review many photos with notes, add a location claim to compare each against, and check a photo against a claimed place. Agreement is not confirmation, and Oceanir says so.
• Reports and exports. PDF reports and CSV exports, and a ruled-out state for places you have excluded.
• Ways to work. The app, a platform with a playground and a workspace, a REST API, and an MCP server that connects Claude, ChatGPT, Cursor and your own agents.
• Manta for organizations. The organization model, with a reviewed access process and benchmark tables in kilometres on shared public datasets.
• Control over your data. Saved history can be switched off for your account, and the server enforces it.
• A rule about people. Oceanir answers where an image was taken, never who is in it, and has no face recognition.

• Name the inputs. The dataset, the exact split, and where the reference coordinates came from.
• Name the configuration. The version, the plan and effort level, and whether metadata or a city hint was allowed.
• Keep the outcomes. The top answer, the timestamp and an identifier for every run, including errors, retries and exclusions.
• Use one scoring rule. Publish the distance calculation and the denominators. Within 1 km does not mean an exact address.
• Separate operating points. Tie accuracy, speed and cost to the configuration that produced each, and date the prices.

Oceanir welcomes comparable evaluations on those terms, including ones where another tool scores higher.

The AI tools comparison has a ten-photo test and a scoring sheet. It takes an afternoon, and it uses photos whose location you know, which is the one input no vendor can choose for you.