METHODOLOGY
What repeated AI answers reveal about brand visibility
Explore 91 recorded answers to see how brand appearances vary and why useful visibility checks need more than a single response.
A brand can appear in several answers to the same shopping question and disappear from another. Repetition helps show that variation, but the design of the test determines what the observations can tell you.
What the historical example shows
This historical chart from Lantern’s July 2026 repeated-question experiment shows one brand and one question in an OpenAI example. Each column represents a collection morning. Across 91 displayed observations, the brand appears in 73 answers and is absent from 18.
The observations include ranks from 1 to 18. One column contains ranks 2 and 18, showing how far the reported position can vary within a collection morning. This describes the recorded example; it does not establish that the prompt was the only influence on the answer or that all underlying conditions stayed fixed.
Dots show recorded ranks and crosses show completed answers that omitted the brand. Eighteen columns contain five observations each; column 10 contains one. The four missing observations in that column have unknown outcomes.
The observations behind the chart
The table lists the observations shown in the chart. Dates appear only where the chart provides a label.
Column | Original date label | Ranks when present | Recorded absences | Observations shown |
|---|---|---|---|---|
1 | Jul 4 | 1, 2, 7, 1, 3 | 0 | 5 |
2 | Not labeled | 2, 1, 1, 2, 1 | 0 | 5 |
3 | Not labeled | 2, 1 | 3 | 5 |
4 | Not labeled | 6, 1, 1, 1, 1 | 0 | 5 |
5 | Jul 9 | 1, 1, 11 | 2 | 5 |
6 | Not labeled | 1, 1, 3, 1, 1 | 0 | 5 |
7 | Not labeled | 3, 4, 1, 2 | 1 | 5 |
8 | Not labeled | 5, 15, 5 | 2 | 5 |
9 | Jul 14 | 9, 2, 13 | 2 | 5 |
10 | Not labeled | 1 | 0 | 1 |
11 | Not labeled | 12, 7, 2, 9, 7 | 0 | 5 |
12 | Not labeled | 13, 7, 2, 18, 2 | 0 | 5 |
13 | Jul 18 | 12, 12, 7, 2 | 1 | 5 |
14 | Not labeled | 14, 12, 3, 2, 7 | 0 | 5 |
15 | Not labeled | 9, 2, 1 | 2 | 5 |
16 | Not labeled | 13, 10, 12, 14 | 1 | 5 |
17 | Jul 22 | 10, 11, 16, 10, 7 | 0 | 5 |
18 | Not labeled | 16, 14, 10 | 2 | 5 |
19 | Not labeled | 9, 10, 1 | 2 | 5 |
What this example cannot establish
The chart does not disclose the exact prompt, model identifier, search settings or full collection log. It is an anonymous historical illustration, not a reproducible provider benchmark. It cannot establish how ChatGPT's consumer interface, another model, a different query or a different month will behave.
It also cannot tell us which part of a change came from generation, retrieved information, product availability or another condition. A repeated prompt controls the wording. It does not control every input to the assistant.
Answers collected close together can share conditions, such as a source being unavailable throughout a session. Spread repeat checks across time as well as within a sitting so one temporary condition does not define your view of the brand.
A practical sampling plan
Choose a fixed set of buyer questions and write down the decision the test should inform. Keep exploratory questions separate from the set used for trend comparisons.
Record the assistant, model when exposed, search mode, locale, date, time and conversation setup. Save the full answer and cited links so claims can be checked.
Collect some repeated observations within sessions and some across days. This separates variation within a sitting from variation over time without assuming a universal best interval.
Distinguish answers that omit the brand from requests that fail or produce no answer. Keep gaps in the available evidence visible.
Compare equivalent prompts and settings. If either changes, report the changed cohort separately before making a trend claim.
Choose the next sample size based on the decision and the observed variability. A small diagnostic sample can reveal a problem; a reliable estimate of a small change generally requires more evidence.
For routine monitoring, choose a schedule the team can sustain and review results more closely around launches, product changes and unexplained shifts.
For a shorter explanation of interpreting repeat checks, read Why one AI answer is never enough.
Common questions
How often should a brand check its AI visibility?
There is no universal interval established by this example. Choose a schedule around the decision, collect observations across time, and adjust when the data or business context justify it.
Are five questions in one sitting five independent observations?
Not necessarily. They share some conditions, and a count of five alone does not establish independence. Keep session and date information so the analysis can account for that relationship.
Does the missing data in column 10 count as four omissions?
No. The chart shows only one observation for that column and does not establish what happened to the other expected records. Missing observations and completed answers without the brand are different states.
Can this chart establish the best provider for brand discovery?
No. It is one historical, anonymous example from one provider. A comparison would need comparable prompts, modes, periods, sample sizes and a documented method across the providers being compared.