The masthead's case for itself
Why CiteTrail exists[1][1] A narrower promise: the thesis below trades a bigger engine list and a fuller dashboard for numbers that survive being clicked on. See the six rules in chapter 02.
CiteTrail measures how AI assistants answer the questions your buyers actually ask: which of them mention you, who gets recommended instead, and which sources the model read to decide.
This page is not a founding story. There is a more useful thing to put here: the thesis the product is built on, and the rules we wrote into it to keep our own numbers honest.
01: The thesis
A channel that does not keep receipts
Search was legible. Someone typed a query, saw a list of links, clicked one, and arrived carrying a referrer that said where they had been. Every step left a record, and a whole discipline grew up around reading those records well.
AI-assisted discovery is not legible in the same way. A buyer describes a problem in their own words. A model reads sources it picked, composes an answer, and names two or three products. The buyer may not click anything. If they arrive later, they often arrive by name or direct, with nothing attached to say what shaped the decision. The step that built the shortlist is the step nobody logged.
That gap is the whole reason this product exists. You cannot manage a channel you cannot see, and you cannot see this one with analytics designed for clicks. The only honest way to observe it is to ask the models the same buyer questions, repeatedly, and record exactly what they say: the answer text, the products named, the sources cited, and how much all of that moves between runs.
That is a narrower promise than most of this category makes. It is also one we can keep.
02: the rules we hold our own numbers to
These are not values on a wall
Each rule is enforced in the product, and each one costs us something: a shorter engine list, an emptier dashboard, a more careful claim.
We query the models, not the apps
CiteTrail does not scrape a chat window. It asks provider APIs the same buyer questions your customers ask, on a schedule, and records what comes back. That is repeatable and defensible. It is also not the consumer surface, which is why we name the surface every time we show a number.
We name the surface, not the brand you expect
One engine we support is Microsoft's Azure-hosted OpenAI deployment, so we call it Azure OpenAI. Another is Gemini answering with live Search grounding, which approximates the shape of a search-grounded answer without being one. Printing the more famous name in either slot would be better marketing and worse measurement.
An engine ships when its adapter does
An engine appears on this site only when a production adapter exists and real prompt runs exercise it. We have left well-known assistants off the list for exactly that reason. A longer logo wall would be easy to build and would be selling a label rather than a measurement.
Observed, estimated and inferred are three different words
Most AI engines strip referrers, so attribution here is genuinely hard. CiteTrail separates traffic it directly observed from influence it estimated and baselines it inferred, and labels each. Blending the three into one confident number is the fastest way to make a dashboard look better and a decision get worse.
An empty state stays empty
When a workspace has no real mentions yet, CiteTrail shows nothing rather than something plausible. A seeded figure reads as data the workspace has not earned, and anyone who acts on it has been misled by us rather than by a model.
Every finding carries its evidence
Scores are arguments, not verdicts. Each one opens into the runs behind it, the sources the engine cited, and how confident we are given how much we sampled. A number that cannot survive being clicked on does not belong on the screen.
03: What we will not tell you
Four claims this page could make and does not
That an answer is the answer
We report what a model said on a given run, on a given surface, in a given region. Someone else asking the same question a minute later may get something different. We show you the spread instead of picking the flattering one.
That we watched your buyer
We query models. We do not observe a real person's chat session, and we never will. Anything framed as “what your buyer saw” would be a guess wearing a lab coat.
That every visit can be attributed
Most assistants strip the referrer. Where we can identify one, we label it Observed. Where we cannot, we label it Estimated or Inferred and leave it in a separate column. We do not add those columns together.
That other people's results predict yours
There are no customer logos on this site, no testimonials and no borrowed statistics. Your category, your competitors and your sources decide your numbers. Run it and find out.
All of it is written up properly on the methodology page, including how sampling works and what each score does and does not mean.
04: Part of the suite
Three products, one account, different questions
CiteTrail
AI-assisted discovery. Which buyer questions assistants answer with your name in them, who they recommend instead, and which sources decided it. You are here.
ClickTrail
The click side. Links, campaigns and the traffic that does carry a referrer.
Visit ClickTrailClickTrail Marketing Intelligence Suite. CiteTrail's job in that picture is AI-assisted discovery.
05: Fair questions about us
Fair questions about us
Is CiteTrail related to ClickTrail?
Yes. CiteTrail is part of the ClickTrail Marketing Intelligence Suite, alongside ClickTrail and NeuralEye. CiteTrail covers AI-assisted discovery specifically: how assistants describe your brand, which buyer questions you appear in, who they recommend instead, and which sources shape those answers.
What does CiteTrail actually query?
Provider APIs, the models behind the assistants, rather than scrapes of their consumer apps. We ask those models the same buyer questions repeatedly and record the answers, the products named and the sources cited. Results will differ from what you see in a chat window. We would rather explain that than imply we are reading over your buyer's shoulder.
Why are some well-known assistants missing from your engine list?
Because we only list engines we can query through a production adapter that real prompt runs exercise. Naming an assistant we cannot reach would be selling a label. When an adapter ships, the engine appears. If one is removed, it disappears.
Do you publish customer names, logos or case studies?
Not yet. Every example on this site is labelled illustrative, and no figure in our marketing is presented as a customer result. When there are results worth publishing, they will be published with the method attached, because a number without its method is decoration.
How can I check any of this?
Read the methodology page. It covers how prompts are sampled, how scores are computed, what confidence levels mean, and what the numbers should not be read as. If something there does not match what the product shows you, that is a bug and we want to hear about it.
- “A narrower promise” is not modesty. It is the trade this whole page describes: fewer engines listed, more caveats printed, in exchange for numbers you do not have to take on trust.
Judge it on your own category.
The argument on this page is worth exactly as much as your own results. Run one.