A brand can receive a perfect consistency score while appearing in none of the answers being monitored. The result makes sense when the statistic measures how often presence changes between observations. A series of uninterrupted absences has complete consistency.
Before using an AI visibility stability score in a purchasing decision or client report, read it beside the mention count, observation history and reporting window. Then inspect the answers behind those numbers. These records establish what the score measures and whether the observed pattern supports the business conclusion.
Xindar publishes this guide and provides Semvi, its AI visibility monitoring product. The numerical histories below are fictional teaching examples. The reporting procedure is an editorial proposal, with supplier definitions attributed where used.
Establish which kind of stability the dashboard measures
Several questions can sit behind a label such as stability. Does the target brand keep appearing? Does the entire list of named suppliers remain similar? Does the answer continue giving accurate reasons for selecting a provider? Each requires a different assessment.
Traqer's public documentation provides a concrete example of the first two. Its brand-presence score measures changes between mentioned and absent states. Its whole-answer brand-set measure uses Jaccard similarity between successive lists. The documentation explicitly explains that an always-absent brand also receives a presence-stability score of 100. These are the provider's published definitions. Traqer stability documentation
The supplier's reporting page also separates per-model mention percentages, raw counts and topic-level coverage. Those measures provide additional context for a stability statistic. Traqer reporting definitions
A buyer should request the formula and the underlying observations before interpreting the label. Ask which surface, question version, brand aliases, dates and completed answers entered the calculation. A score with those details attached can support a specific decision.
Also establish whether the displayed history concerns mentions or citations. A company name in an answer and a URL in a source list are separate events. The target entity needs its own verification, particularly when several businesses share a short name.
Calculate presence and changes separately
For a complete sequence, assign `1` to an answer that mentions the verified target brand and `0` to a complete answer that omits it. Let `n` be the number of answers and `F` the number of changes between adjacent answers.
The following calculation expresses the switching-based definition described in the supplier documentation.
`Mention proportion = number of 1s / n`
`Stability score = 100 × (1 − F / (n − 1))`
The second expression requires at least two observations. A sequence with 12 observations has 11 adjacent comparisons. Consecutive pairs such as `00` and `11` contribute zero changes. Pairs such as `01` and `10` each contribute one.
Five fictional histories show how the two measures behave. Each row represents one exact question on one specified consumer surface under a documented, unchanged collection setup. Twelve observations make the arithmetic easy to inspect.
| Complete history | Mentions | Changes | Stability |
|---|---|---|---|
| `000000000000` | 0/12, 0% | 0/11 | 100 |
| `111111111111` | 12/12, 100% | 0/11 | 100 |
| `111111000000` | 6/12, 50% | 1/11 | 90.9 |
| `000000111111` | 6/12, 50% | 1/11 | 90.9 |
| `010101010101` | 6/12, 50% | 11/11 | 0 |
The first two histories have the same stability and opposite presence. The middle two have identical mention proportions and stability, while the order of their appearances describes opposite recent directions. The alternating history has the same overall mention proportion as both middle rows and a very different pattern.
A client report can preserve all these facts. Include the complete sequence or an equivalent dated chart alongside the summary. The figures then support a conversation about sustained inclusion, persistent absence, a recent loss or a recent gain.
Read the reporting window before the headline
Consider the fictional `111111000000` sequence again. Across all 12 answers, mention proportion is 50% and stability is approximately 90.9. Across the final six answers, mention proportion is 0%. Across the first six, it is 100%.
All three proportions describe their stated selections correctly. The operational question is which selection answers the client's decision. A historical overview can describe the full record. A review of a recent disappearance needs the dated recent segment as well.
Traqer's current help distinguishes a single selected run from historical views and offers narrower recent-run selections. Its separate Trends documentation describes another rule for flagging changes between recent and earlier runs. A dashboard can therefore expose several views of the same retained history. Record the selected view with the result. Historical-view definitions, Trends documentation
Keep the actual start and end dates. Twelve weekly observations cover a different elapsed period from 12 daily observations. A selection called recent may also contain a different number of observations after collection failures. Calendar period, scheduled opportunities and completed answers deserve their own fields.
Choose the comparison window before reviewing the outcome when evaluating a change. Preserve earlier windows when a business need justifies a new one. This gives a reviewer enough information to reconstruct why the reported result changed.
Check a changed row count against the selected dates
On November 14, 2026, Xindar checked one question and platform in its legacy monitoring application's owner-labelled demonstration/seed dataset. The default date filter covered October 15 through November 14. A normal CSV export contained 25 metadata rows, each matched to the visible table fields. The previous day's saved export contained 26 rows. A complete-row comparison found all 25 current rows in that earlier file.
Expanding the date selection exposed the older record that accounted for the difference.
| Selected view | Reported rows | Actual inspection |
|---|---|---|
| October 15–November 14 | 25 | All 25 metadata rows matched the exported file |
| August 16–November 14 | 85 | The first 50 rows were read, including the older record |
| Restored October 15–November 14 | 25 | The same original 25 rows returned |
The wider filter label initially appeared while the table still displayed the earlier 25 rows. The review waited for the wider-window records to appear before accepting that state. A changed heading alone would have left the comparison incomplete.
This demonstration establishes a bounded date-filter and metadata-export workflow. The CSV's ten columns covered fields such as question, platform, mention status and rank. Full answer text, citation URLs and labelled time zones require separate evidence. For a client report, preserve the settled selection, exported fields and inspected row count with the comparison.
Give missing observations their own state
A login barrier, rate limit or incomplete capture leaves the answer unknown. Use a missing state for that event and retain its time and collection status. Reserve `0` for a complete, reviewed answer in which the target brand is absent.
Suppose three scheduled observations produce the following fictional record.
| Scheduled observation | Recorded state | Interpretation |
|---|---|---|
| First | `1` | Complete answer, target mentioned |
| Second | Missing | Collection failed |
| Third | `1` | Complete answer, target mentioned |
The retained complete answers have a mention proportion of 2/2. Completion is 2/3. The intervening answer remains unknown.
For a procedure that assesses changes between adjacent scheduled observations, neither of the two scheduled pairs is fully observed. The first pair includes the failed collection, and so does the second. Report eligible adjacent pairs as zero and the corresponding stability statistic as unavailable.
Another procedure might explicitly compare successive successful captures across gaps. Such a procedure should retain the gap and elapsed time in its definition. Agree which procedure the report uses before comparing scores across products or periods.
The adjacent-scheduled-pair treatment here is a proposed review rule. A supplier's actual handling of failed runs requires its own documentation and demonstration. Ask to see an example containing a preserved failure, alongside the displayed denominator and exported records.
A stable target can sit inside a changing shortlist
Now consider two fictional answers. The first names suppliers A and B. The second names A and C. Supplier A appears in both, giving it uninterrupted presence across that pair.
The two complete brand sets have one shared member and three distinct members overall. Their Jaccard similarity is therefore:
`|{A, B} ∩ {A, C}| / |{A, B} ∪ {A, C}| = 1/3`
The target's presence and the full shortlist's membership describe different aspects of the answer. This distinction matters when a buyer wants to understand which competitors rotate around a consistently named brand.
Normalize entity identities before constructing those sets. A trading name and its verified parent reference may refer to one business, while a similarly named company may be unrelated. Preserve the text and the identity decision so that the normalization can be reviewed.
List similarity also leaves recommendation reasons to inspect. A stable name could be accompanied by a correct capability in one answer and an unsupported capability in another. Review the individual reason against current evidence and record the resulting recommendation assessment separately.
Keep the question's version in the series
A series needs a clear observation unit. An agency can generate several related questions around one buying topic, then observe each question repeatedly. Those are two dimensions of the study.
Traqer's prompt documentation describes generating the topic text plus four related questions, with editing available to the user. That is a useful illustration of why a topic label alone cannot identify the question submitted in every run. Prompt-generation and editing documentation
Retain an identifier and exact wording for each version. Record when the wording, surface, model or relevant account conditions change. A series assembled from different questions can reflect a change in the instrument as well as a change in answers.
For client reporting, keep an unchanged-question comparison alongside an expanded exploration panel. The exploration panel can investigate new needs. The unchanged-question comparison supplies continuity for the decision already under review.
Changes to the active question selection also affect historical percentages. The related prompt-lifecycle guide covers the retention and reporting consequences of editing, archiving and deleting questions. The stability calculation here addresses the ordered answer sequence within a declared selection.
Put an inspectable record behind the client report
An agency reviewing AI visibility can request a compact handover containing the following items.
| Item | What the reviewer should be able to establish |
|---|---|
| Series identity | Exact question version, target entity and selected surface |
| Collection context | Timestamps, mode, visible model label and documented environment |
| Completion ledger | Scheduled observations, completed answers and retained failures |
| Presence history | The dated sequence used to calculate mentions and changes |
| Metric definition | Formula, window, rounding and treatment of missing observations |
| Answer evidence | Complete answer text and available visible source URLs |
| Recommendation review | Entity decision, individual reasons and supporting evidence |
A commercial agreement can assign who reviews exceptions and how a corrected classification affects a delivered report. Keep the earlier report and correction record available under the agreed access policy.
Apply the same review to Semvi by Xindar. Specify the required question history, collection surface, export and interpretation in the proposed scope, then inspect the supplied records. The Semvi product reference explains the published monitoring and evidence boundaries.
For an export manufacturer, connect that review to a concrete purchasing question. If a monitor repeatedly names the wrong legal entity or describes outsourced work as in-house production, assign the correction to an approved source and responsible reviewer. A consistent description still needs to be accurate and relevant to the buyer's task.
Questions to ask before using the score
What does a stability score of 100 establish?
Under the switching-based definition used here, the observed binary state never changed within the declared sequence. Read the mention count to establish whether that state was presence or absence, then inspect the dates and evidence.
Can a high stability score accompany a recent decline?
The fictional six-present, six-absent sequence has one change in 11 comparisons and a stability score of approximately 90.9. Its recent segment contains six absences. The full sequence and recent window make that situation visible.
How should a report handle one completed answer?
Report the single answer and its assessment. An adjacent-change statistic has zero comparisons in that case, so its denominator is zero and the statistic is unavailable.
Does a repeated mention establish a repeated recommendation?
Evaluate each answer's actual recommendation, entity and supporting reasons. Presence history supplies one part of that review. Preserve unsupported or incorrect reasons alongside the original answer and the assessment.
What should a manager receive with the headline number?
The question version, collection surface, reporting window, completion counts, ordered history and metric definition. Include links to the permitted underlying records and a named owner for unresolved observations.
Sources and scope
Public source review began on November 13, 2026. The stability and reporting pages were checked again on November 14. Traqer's stability, reporting, prompt and trends pages supply attributed definitions and examples of documented product behavior. Its account operation and collection quality require a separate product demonstration. The fictional sequences, missing-pair rule, Jaccard example and procurement checklist are this article's worked teaching material. The November 14 date-filter case comes from Xindar's retained screen and CSV records for an owner-labelled demonstration/seed dataset. Its scope covers the stated interface workflow and metadata fields.
