Engines & Regions

Aigeo probes nine assistants. They differ in how they retrieve information, which is why the same question gets different answers from each — and why “which market am I measuring?” has a different answer per engine.

The engines

Engine	Retrieval	Region signal
ChatGPT	Live web session	Adjustable
Google AI Overview	Live search results	Adjustable
Google AI Mode	Live search results	Adjustable
Gemini	Live web session	Adjustable
Perplexity	Live web session	Adjustable
Grok	Live web session	Adjustable
DeepSeek	Live web session	Adjustable
Claude	Official API with its own retrieval backend	None
Doubao	Official API, Chinese-language retrieval	Fixed (CN)

Every engine counts the same against your allowance: one answer is one probe, with no cost weighting for expensive models. See Plans & usage.

Region is a property of retrieval, not of the model

A model’s weights have no location. Calling the same model from Tokyo or from New York produces the same distribution of outputs. Location only enters through the retrieval layer — what the engine finds when it searches.

That is why the major LLM APIs have no country parameter at the top level. (Anthropic’s inference_geo governs data residency, not the content of an answer.) The things that can accept a country are search and browsing tools.

So engines fall into three classes, and the app reports each class differently:

Adjustable Real geographic signal The session runs from an exit point in the target country, so the answer reflects what a local user actually sees. Only this class is used for cross-region comparison.

Fixed Locked to one market Retrieval is inherently tied to one market — Doubao searches the Chinese web. Region is a constant, not a parameter, so results are grouped under that market.

None No location declared A native API with its own retrieval that declares no location. A valid visibility signal, just without a geographic dimension — reported under "Global engines".

How a run expands across markets

Region-aware engines run once per selected market. Region-agnostic engines run once, in the primary market — sending them the same prompt several times would produce the same answer and consume allowance for nothing.

probes = prompts × Σ_engine (region-aware ? regions : 1)

Worked example — 10 prompts, ChatGPT + Gemini + Claude:

Markets	Naive 3 × N	Actual	Difference
1	30	10 × (1+1+1) = 30	—
2	60	10 × (2+2+1) = 50	−10
3	90	10 × (3+3+1) = 70	−20

The run dialog shows the exact count before you start, and it is the same number used for progress and for allowance.

What the panel does with each class

Scope selector	Shows
All	Every probe in the run, all three classes together
A country code	Adjustable engines probed in that market, plus fixed-region engines whose home market it is
Global engines	Engines with no location signal

Two rules keep the comparison honest:

1. Cross-region tables use adjustable engines only. Every row then has the same prompts and the same engines behind it, so the denominators match and the comparison means something. Mixing in Claude or Doubao would produce a table that looks comparable and isn’t.
2. Fixed-region engines are shown under their real market. Doubao is executed once in the primary market for cost reasons, but it is displayed under CN, because that is the market its answers describe. Showing it under, say, US would be misleading in exactly the way the region split exists to prevent.

The panel adds a footnote naming which engines fell into each bucket, so a missing engine in a region table is explained rather than mysterious.

Choosing markets

• Start with one. The primary market is where your first run happens, and one market keeps the probe count low while you tune the prompt library.
• Add markets you actually sell into. Each additional market multiplies the run cost by roughly the number of region-aware engines you’ve selected.
• Compare like with like. If a market looks weak, check the region table’s engine footnote before concluding anything — a region with fewer engines behind it is a different sample, not necessarily a worse market.

Region-only differences that are real usually show up in two places at once: a lower mention rate and a different top competitor. A local player dominating one market and absent from others is the clearest signal the region split produces.

Averages & denominators → How region scores and the cross-region baseline are computed. Plans & usage → What a run costs once markets multiply.