How the Citelight Score is built.

The formula, the weights, the sample, and the things we refuse to compute. Published because a score you can't inspect is a number, not a measurement.

Version 3·In effect since 24 August 2026·Last revised 3 September 2026
The short version

Mention rate, citation position, cross-platform reach and citation quality, weighted, out of a hundred.

Mention rate
How often you are named at all, across every question we ask about your market.
30
Citation position
Where you appear when you do. First in an answer is worth more than eighth.
25
Cross-platform reach
How many of the scored platforms cite you, rather than how loudly one does.
25
Citation quality
What kind of page carried you. A ranking article and your own site are not equivalent.
20

Weights have been fixed since 24 August 2026. Any change registers as an instrument boundary, described below.

The one worth showing in full

Cross-platform reach, and its denominator.

Reach

Four platforms are scored: ChatGPT, Perplexity, Gemini and Google Search. Reach is the share of them that cited you in a given week.

The denominator is not always four, and that is deliberate. A platform that failed on our side is removed from it, because a failure of ours is not an absence of yours. A platform we chose not to run that week stays in, because a cadence decision is ours to answer for and you should be measured against the full set.

Those are different facts and the score treats them differently. Your reading states which platforms answered and which did not, every week.

AI Overview is measured and reported. It is not scored. Google withholds it on roughly a quarter of queries, and a denominator that moves with someone else's product decision is not a denominator. You will see it in your reading, marked as reported rather than scored.

The sample

What we ask, and how much of it.

Questions are generated per agent from your market, your neighborhoods, your specialties and your buyer or seller focus. They are the questions a buyer types, not keywords.

Your saved profileQuestions per week
No specialties saved0
One specialty7
Two specialties14
Three specialties16

Plus eight shared questions about the team brand on a Beacon subscription. Coverage is per agent and is never divided among a team: a nine-person team asks nine times as many questions, not the same number split nine ways.

Every question runs on every scored platform, every week, except Google Search, which runs in the first ISO week of each month. Your reading states which platforms answered and which did not.

One exception, stated because it is currently true

Agents measured as part of a team are read on ChatGPT, Perplexity and Gemini. Google Search does not currently run for them, so their reach is computed over three platforms rather than four.

This is a limit of our pipeline, not of their visibility, and we are removing it. Until we do, a team member's reading says three where a solo reading says four.

When the instrument changes

A boundary, not a quiet correction.

We change what we ask and how we read it. Query families get reworded, a parser gets fixed, a platform changes its model.

When that happens we register a boundary at that week, and any week-over-week comparison that crosses it is suppressed rather than shown. Your score still appears. The arrow beside it does not, and the reading says why.

The reason, plainly

A delta across an instrument change measures us, not you. Showing it would credit an agent for a parser fix or blame them for a reworded question, and neither is a thing they did.

Two boundaries are currently registered: week 32 and week 37 of 2026.

What we refuse to compute

Four numbers we could show and don't.

Week-one delta

A first reading has nothing to compare against, so the change reads as blank rather than zero. One point is a position, not a trend.

Per-dimension deltas

A dimension-level change needs a baseline captured under the same instrument. Ours has moved three times in five weeks, so we publish the composite delta and no others.

Predicted scores

No recommendation says what a number will become. We report what moved and what didn't, including the weeks when nothing did.

A zero that means unmeasured

Where a platform didn't answer, the reading says so. It never renders as a nought, because absence and zero are different facts and only one of them is about you.

The corpus

Where the comparison comes from.

Your rank is against the agents the assistants actually name in your market, not a list we assembled. Every name an answer produces is recorded, and names that resolve to the same person are held apart unless a rule merges them, so a team brand and the person behind it are not silently combined.

Broward County, our longest-running market
Citations recorded-
Readings-
Weeks measured-
Agents discovered-
Named in three or more separate readings-

Two caveats, and what they make the numbers. The per-reading record disagrees with the running counter on 131 of 1,079 discovered names, and 252 entries in that record point at readings we can no longer find, either deleted or never persisted. A reading the record does not carry cannot be counted, so discovered and corroborated are floors rather than exact counts. A floor we can defend beats a number we cannot.

Figures update weekly from the live corpus.

Our own test accounts are removed before any figure is computed. They accounted for 68% of raw citation rows, and excluding them reorders every ranking. We mention it because a number computed over the wrong population looks exactly like a number computed over the right one.

Sources

We record the page, not just the mention.

Every answer arrives with the sources the platform used. We keep all of them, and we keep where in the answer each one sits, so a citation can be tied to the sentence that names you rather than to the response as a whole.

PlatformHow the source is anchored
ChatGPTCharacter offsets into the answer text
PerplexityInline markers indexing the citation list
GeminiGrounding supports mapping text segments to chunks
Google SearchOrganic position and the result's own snippet
One consequence worth stating

Perplexity retrieves considerably more sources than it references: 13,293 fetched against 5,551 actually used in answers.

We count what was referenced. A page in the retrieval set was read and not mentioned, and reporting it as a citation would count a search as a recommendation.

Limits

What this doesn't tell you.

It is observation. We measure what the assistants say and what they read to say it. Where a recommendation can be verified, a later reading checks whether it moved, and we say which reading did the checking.

Assistants vary. The same question can return different names on different days. We run the same set weekly and report the trend rather than a single reading, which is why a first week has no arrow beside it.

Coverage depends on your profile. An agent who has saved one specialty is measured on seven questions and one who has saved three is measured on sixteen. Your reading states which, every week.

And a name we could not find is not a name that is absent. Where our matching is uncertain, the reading says we could not confirm rather than reporting a zero.

Citelight LLC · Score model version 3 · In effect from week 34 of 2026
This page is revised whenever the model changes, and the revision is dated.
Questions about the method: hello@citelight.ai