A Basano mode
Lanthe (LAN-thee): AI citation measurement. A Basano mode.
Why AI answer surfaces cite some brands and not others, measured across the ways real buyers actually ask, and honest about what a measurement like this cannot settle.
What it answers
AI answer surfaces recommend some brands and pass over others. Lanthe measures which brands an AI answer cites for the questions real buyers ask, and where a brand is absent, which sources hold the slot in its place.
The catalog's SEO skills can audit a page for schema, entity coverage, and answer-engine readiness. They cannot tell you whether that work moved you into the answer set on the surfaces buyers actually use. Lanthe measures the outcome rather than the readiness.
The measurement
Two sources of variance, and only one is small.
A brand's citation share carries two independent sources of noise. The first is within-prompt: ask the same question twice and the answers differ, because the models sample as they generate. That source is small.
The second is across-prompt: a brand's citation rate depends heavily on how the question is phrased. "Best trail running shoes" and "what shoes for a first ultramarathon" surface different brands. The number worth having is a brand's share across the whole range of ways buyers ask, and the spread between phrasings is where most of the variance lives.
So Lanthe samples the prompt space broadly and runs each prompt once, instead of re-running a handful of prompts many times. At a fixed budget, covering more of how buyers ask beats re-asking the same question, because across-prompt variance is the larger of the two.
The validity bottleneck
The prompt set is where the measurement is won or lost.
Sample size buys precision, not accuracy. A large set drawn from a skewed pool gives a precise estimate of the wrong number. The whole validity rests on the prompt set matching how buyers in a category actually ask, which makes its construction the part to get right, and the part to disclose.
Seed from real demand. Start from the category's real search demand, the closest proxy there is for what buyers want to know. It also carries the frequency weights used further down.
Anchor in real human phrasing. Pull the question forms from where people ask in the open: People Also Ask, autocomplete, and category threads on forums and review sites. Model-written prompts drift into a register that does not match messy human phrasing, so generated expansion only fills gaps.
Stratify by intent. Buyer questions span discovery, comparison, constraint-based, problem-led, and validation. Sample across those buckets rather than over-weighting the easy 'best X' form that dominates whenever prompts are generated.
Weight by frequency. Sample the pool in proportion to real demand, so a common phrasing counts for more than a rare one. That is what turns a list of prompts into a representative sample.
Validate before use. Check that the set's intent mix matches the demand mix, and that the prompts read like real questions when sampled against the phrasing sources. This is Basano verifying its own instrument before it measures with it.
Document the construction. Record the buckets, the weights, and the sources. That record is the validity disclosure the honest account and the report's coverage note both rest on.
The honesty surface
Every share reports with its interval.
The category is full of tools that run one prompt once and hand back a confident rank in AI search. That number is noise dressed as signal, and a buyer who acts on it is acting on a coin flip.
Lanthe reports each share with its interval, names the surface it measured, and discloses how the prompt set was built. It will not convert a sampled, drifting signal into a precision it does not have. A surface can be read through an API or through the consumer product, and the two can diverge; the report names which one it read, and never presents an API result as the consumer answer.
When a surface blocks capture or a run errors out, that shows up in the report as exactly what it is. A blocked read is never counted as a brand's absence.
What this cannot tell you
This is a standing section, not a line of small print. A measurement like this has real limits, and naming them is part of the method rather than a hedge against it.
Precision is not accuracy. The interval reflects the sample size. The accuracy reflects whether the prompt set matched your buyers, which is why the construction is disclosed and worth reading first.
It is a snapshot. Citation results drift from run to run, by account, by location, and by date. Every measurement is timestamped, with its variance stated.
It measured one surface. The consumer product and the API can disagree. The report names which was measured, and a result on one does not transfer to the other.
It cannot promise a fix will land. A model's internal weighting is not observable, so the reason a cited source wins a slot is a hypothesis, marked tested or untested, never a guarantee.
Blocked and failed reads are shown. A capture stopped at a bot wall, or an API that errored out, appears in the report as what it is. It is never quietly folded into a brand's absence.
The method is open
The rigor behind Lanthe is publishable, and it is published. Seed from demand, anchor in real questions, stratify by intent, weight by frequency, validate before use: that is the method, and anyone can run it. The template on GitHub is that measurement discipline, runnable against any brand and category.
Powered by Basano
Basano proves.
Lanthe is not a separate engine. It is Basano pointed at a different standard: presence in AI answers instead of correctness in a build. Krine decides, Tholo builds, Basano proves. Lanthe is one of the ways Basano proves.