GEO in practice
Is showing up in the API the same as showing up in the ChatGPT interface?
No. And it isn't a minor difference: it's why the Dénia GEO study explicitly separates its API measurement phase from a separately planned user-interface phase.
By the ArribaIA team · Updated
Summary
An API and an AI assistant's consumer interface can share the same underlying model and still return different answers: the interface can apply personalization, conversation history, or default-enabled tools that a basic API call doesn't include unless turned on explicitly. Measuring via API alone is faster and more repeatable, but it describes the API's behavior, not necessarily the product a real person uses.
What an API is, in this context
An API (application programming interface) is the channel a program uses to send a question to an AI model and get an answer back, without going through the page or app a regular person uses.
It's the channel this study uses: it lets us automate sending thousands of queries, fix exact parameters (for example, whether web search is enabled or not), and keep an exact record of what was sent and what came back. Hard to replicate at this volume by opening the app by hand.
What the user interface is
The user interface is the app or website a person interacts with: ChatGPT.com, the ChatGPT app, the Gemini app. It's the product the vast majority of people actually use when they ask an AI where to eat.
The interface can include product behaviors not present in a bare API call: conversation history, device location, memory across sessions, or tools enabled by default that the API requires you to turn on explicitly.
What we've measured so far in the Dénia study
The current phase of the Dénia GEO study sends its queries against the analyzed providers' APIs, not their consumer interfaces. We cover this in detail in the full methodology.
A second phase is planned to repeat a representative sample of queries using each assistant's real interface, precisely so we can check whether the API and the interface agree in this specific case. That phase complements the API measurement, it doesn't replace or invalidate it.
Where they can diverge, aspect by aspect
This table summarizes differences documented by the providers themselves between programmatic use (API) and the consumer product, not an exhaustive comparison of every feature.
| Aspect | Via API | Consumer interface |
|---|---|---|
| Web search | A configurable tool, with filters and sources that must be explicitly enabled in the call. | Can be on by default depending on the product and model chosen, with behavior the user doesn't control in detail. |
| Personalization | No memory or account context unless the calling application adds it itself. | Can incorporate conversation history, saved preferences, and, in some products, device location. |
| Conversation context | Each call is independent unless the system using it manually resends the history. | Natively keeps the conversation thread within the same session. |
| Scalability | Built to launch thousands of automated, measurable queries. | Built for one person at a time; automating it at scale isn't the product's intended use. |
| Repeatability | Every call's parameters are fixed and logged, making it easier to repeat a query under the same conditions. | Different session and product factors can make the same question, typed twice, not behave exactly the same. |
What this comparison doesn't settle
- We don't know precisely which product settings each interface applies by default: providers don't publish that detail.
- A difference observed between API and UI in one sample doesn't prove that difference holds always, for any query, at any time.
- This table doesn't cover every product or every assistant that exists, only the general concepts documented by the providers cited below.
Documentation consulted for this article
- Official documentationOpenAIWeb search (Responses API)
Describes the API's web search as a configurable tool, with filter and source controls the consumer interface doesn't expose the same way.
- Official documentationPeec AIIntro to Peec AI
Official documentation for an AI visibility monitoring tool explaining why it uses interface automation instead of the API: "API responses often differ from what users see in the actual interface".
Frequently asked questions
Common questions about API and UI
So the study's API results are useless?
No, they're useful within their actual scope: they rigorously describe the API's behavior under controlled, repeatable conditions, which is exactly what a measurable experiment at this volume needs. What we don't do is present them as identical to what someone would see opening the app.
Why didn't you measure directly with the interface from the start?
Because automating thousands of queries against a consumer interface isn't viable and isn't what those products are built for. Starting with the API allows a volume and control the interface doesn't offer; the interface phase comes after, on a representative sample.
Back to the full study
This article is part of the Dénia restaurants GEO study. The full methodology details exactly what's been measured so far and what's still to be published.