/v1/ai-answers/*. See the API Reference for the full endpoint list. Credit rates are in Credits.
This Source is the one you ask rather than search. You write the question a customer would ask, name the brands to look for in the reply, and Nephia puts that question to each engine on a schedule — recording the answer, the sites it cited, and which of your brands made it in.
Where the answers come from
Answers come from the engines’ consumer products — the same chatgpt.com and gemini.google.com a person reaches without an account — through a scraping vendor, not from the vendors’ own developer APIs. That is a deliberate choice and it is worth stating plainly, because the two are not the same thing. An API answer is what the model says to a developer, with the developer’s own settings. What a brand is paying to find out is what the model says to a customer. Those differ, and only one of them is the marketing question. What it costs you in exchange: answers are not deterministic. Asking the same question twice minutes apart can return a different set of sources and a differently worded reply. Nephia is built around that — see What counts as a change.The unit
Every run becomes an AI answer: the engine, the prompt, the answer text, the sites it cited, and one row per tracked brand saying whether it was mentioned, cited, both, or neither. Mentioned and cited are different questions, and the difference is the point:
A brand can be discussed at length and never linked. A brand can be linked in a list it is never named in — and that second case is worth knowing about precisely because it is invisible to anyone reading the answer.
searched tells you whether the engine actually browsed for this answer. An answer given from training data alone is a fact about the model’s priors, not about what is on the web this week; zero citations with searched: false is a complete, honest reading, and the same list with searched: true is a finding.
Engines
chatgpt and gemini.
Perplexity is not offered. Its anonymous endpoint is closed at our vendor and needs a Perplexity account session, which is a dependency we are not taking on for you. Claude is not offered either — reaching it would mean a second, differently-shaped provenance in the same table, and an answer measured one way sitting beside an answer measured another with nothing on screen saying so.
Watches
A watch takes aprompt and, optionally, the brands to look for.
There is no query and no terms here, and the absence is deliberate: on every other Source those mean “search text”, and this Source is asked rather than searched. Passing them is rejected with 400 rather than ignored.
Interval: 21600–604800 seconds — six hours to seven days. Cap: 10 active watches per Account. Both are lower and slower than every other Source for one reason: a tick here is a purchase, not a feed fetch.
Each engine costs 100 credits per run. Three runs a day on two engines is 600 credits a day. An engine that errors is not charged — the others still answer, and a failed engine simply keeps its previous answer as the baseline.
domains is what lets a citation of your own site count as cited even when the answer never names you. Without it, a brand is only cited when its name appears in a citation’s title.
country is the country the question is asked from. It is not a filter — the answer to “the best X” is genuinely different in different places.
What counts as a change
A prompt asked every six hours always answers, so “an answer exists” is never news. Only transitions produce events:
There is no silent baseline on this Source, unlike every other: the first run emits
answer.created. The first answer is the information, and a reader who just asked “what does ChatGPT say about us” is owed it rather than a six-hour wait.
The thresholds behind answer.changed are deliberately loose, because these engines reword themselves constantly. A rerun that reorders the same recommendation, or swaps one source in four, does not fire. A rule keyed on being named would fire on almost every run, so the change rule is keyed on being cited — which is stable. mentioned is still recorded on every run and visible in the history; it is a fact about an answer, just not one worth waking you up for.
Reading the history
GET /v1/ai-answers/watches/{id}/runs returns every stored run, newest first — including the ones that changed nothing and emitted no event, which is most of them and exactly what a trend is made of.