Not whether a page ranks, but whether you get named at all, and what gets said about you when you do. AISurface asks the questions your buyers ask, across five assistants, on a schedule, and records the answers.
A monitoring app for AI answer surfaces. You give it a domain, it works out the brand, the specific product, the competitors and the topics people ask about in that category. You then define a query set, connect assistant API keys, and run.
Every run asks each query multiple times across every connected channel and stores what came back:
Collection happens on the server rather than in the browser, so closing the tab does not stop a run, and the three channels no browser is allowed to call still work.
Keys are encrypted before they touch the database and are never sent back to the browser. There is no endpoint that returns one. Once a key is saved it can be replaced or deleted, never read.
Rank tracking answers a question nobody is asking anymore. When the answer is synthesised, there is no position ten to climb to. There is one paragraph, and either you are in it or you are not.
Worse, being in it is not automatically good. Assistants confidently state pricing, features and guarantees that are out of date or simply wrong, and no rank tracker will ever tell you that. Sentiment tools will not either, because a warm paragraph containing a wrong price is still a warm paragraph.
AISurface measures two things a rank tracker cannot:
Whether you are in the paragraph at all, by name or not.
Whether the claims it makes about you are actually true.
Four situations account for almost every run I make of it.
Six steps to a dataset, and the fourth one is the one people skip. The loop below plays one run for one product across five assistants, then starts again.
Five platforms: Claude, ChatGPT, Perplexity, Gemini and Google AI Overviews. Anthropic is required, because it does the extraction and verification work for every channel. The other four are optional channels, and a run with none of them still measures Claude.
Enter a domain and, optionally, the specific product. The app returns the brand, the product, the competitors and the topics, so the query set has something to be built out of rather than a blank field.
Correct the analysis by hand. Aliases, competitors that are not competitors, topics that belong to a different product. This is a starting point, not an answer, and every later number inherits whatever is left uncorrected here.
Verified facts about the product, one per line: pricing, features, guarantees. Optional, and the highest-return five minutes in the setup, because it is what later claim verification is checked against.
The questions asked on every run. About sixty percent should never name you, because that is where the real result is: an answer that recommends three tools and none of them are yours is invisible to anything that only looks at citations.
Every query is asked repeatedly on every channel in scope, so answer volatility is visible rather than averaged away. Cited claims are then optionally verified against the pages they cite, deduplicated so a claim recurring across runs is checked once.
Then the explorer: six ways into the same dataset, filterable by platform, source type, sentiment band, head-to-head, topic and query intent, with the underlying citation records downloadable as records.json. Coverage answers whether you showed up at all; the other five describe what happened when you did.
Anyone with a category worth being named in and the patience to define a real query set. Signup is open, sign-in runs through GitHub, and you pay your own assistant API usage rather than a subscription to me.
Coming soon. It will be open to signup, and free to use as long as you bring keys.
Status and access
A measurement is where this starts, not where it ends. Knowing that four of five assistants recommend a competitor on your best topic, and that the one citing you is citing a page you deprecated in 2024, is the beginning of a sequence of work. Deciding what that work is, and in what order, is what an engagement is for.