← 返回资讯
2026-11-21约 29 分钟阅读责任主体:芯芸达

One Question, Two Visible AI Answers: What Belongs in a Monitoring Record?

Record two visible AI answers without selecting the favorable result or silently changing the ordinary-answer denominator.

A B2B marketer submits one supplier-selection question. The browser displays two responses and asks which is better. The company appears in the left response and is absent from the right. A screenshot of the left side would make an attractive result for the weekly report. It would also leave the reader unable to reconstruct what happened.

The encounter produced one submitted question, two visible answer objects and a comparison interface. Those are different counting units. A monitoring policy needs to say which unit enters its ordinary visibility rate before anyone knows which response favors the company.

For an ordinary-answer study, a defensible default is to preserve both responses, classify the encounter separately and exclude it from the ordinary-answer denominator. A study designed to measure preference comparisons can analyze it under its own rules. Neither approach requires discarding the material or choosing the more favorable response.

Identify what the interface actually presented

Several events can leave more than one answer on screen. A platform may present a comparison. A user may request another draft, regenerate an answer or edit the original question. The collector may have opened two conversations. The browser may simply be showing the same answer twice.

The visible interface and action history determine which description is supported. Two panels do not establish which model generated either one, how they were allocated or why the platform displayed them. Record an unknown model as unknown.

Google documents ways to regenerate a Gemini response and switch between response versions. That computer help page establishes a user-facing workflow. It does not establish that every two-answer encounter is the same experiment, or that response versions are independent observations from a target population.

Research also uses paired outputs for a different purpose. The MT-Bench and Chatbot Arena paper studies model evaluation and preference judgments. A comparison-oriented evaluation should not be silently substituted for a study of what a buyer receives in an ordinary consumer session.

The collection record should therefore describe what was visible rather than diagnose a hidden experiment.

Visible eventRecord separatelyAvoid assuming
Two answers with a preference promptBoth bodies, panel labels, prompt wording, selection stateModel identity, random allocation or population frequency
A collector requests another draftOriginal output, requested action and later draftTwo fresh ordinary sessions
A question is editedBoth question versions and their outputsThe answer still belongs to the frozen question
The same response appears in two viewsResponse identity and display locationsTwo generated answers

Work through the denominator before reviewing the brand

Consider a fictional monitoring day with ten planned question–surface cells. Eight return complete ordinary answers. One returns a two-panel comparison. One reaches a usage limit without producing an answer. The brand receives a qualified recommendation in three of the eight ordinary answers; it appears in one comparison panel.

Under the separate-comparison policy, the ordinary recommendation result is 3/8. Collection coverage for complete ordinary answers is 8/10 planned cells. The comparison encounter is reported separately as one encounter with two retained outputs. The limit remains a collection outcome, not a brand-negative answer.

Picking the favorable panel and reporting 4/9 changes the inclusion rule after seeing the result. Treating both panels as ordinary answers produces another rate with a different dependence structure. Reporting 4/10 while counting the unsatisfied usage limit as an ordinary negative introduces a further assumption.

All those fractions can be calculated. Arithmetic does not make their meanings interchangeable.

These numbers are invented to explain the policy. They are not Semvi measurements, a customer result or an estimate of how often any platform displays comparisons.

Store the encounter above its answer objects

A useful data model has a parent encounter and child outputs. The parent contains the exact submitted question, collection time, consumer surface, account context, visible mode and interface classification. Each child holds one complete answer, its visible label, captured sources and capture status.

Keep the selection state explicit: no choice made, left selected, right selected or unknown. A preference click is a collector action. It can change the interface and should not be added merely to make collection easier.

Also distinguish generation from capture. If the right answer appears complete but the collector copied only its first paragraph, the record contains an incomplete capture of that output. It does not establish that the platform generated only a paragraph.

An encounter-level exclusion needs a reason that survives later review. “Brand absent” is not a methodological reason to exclude an otherwise eligible ordinary answer. “Nonstandard comparison interface, excluded under the predeclared ordinary-answer rule” is a rule that can be applied to positive and negative outcomes alike.

Preserve both answers without flattening their sources

A source panel can belong to one response or change when the active response changes. Capture its association instead of combining all visible links into a single apparent bibliography.

If the interface provides source cards without readable destinations, keep the visible titles and mark the URLs unavailable. A favicon is not a reliable source URL. A link that can be opened later is evidence of that destination at the later time; it is not a substitute for the original answer text.

For each panel, ask whether the entire recommendation passage was retained, whether citations were actually attached to that panel, and whether the cited page supports the specific recommendation reason. A brand named in a citation title can still be absent from the answer's recommendation.

The source review should produce separate judgments for the two outputs. One may contain an accurate shortlist while the other misclassifies the company. That is useful evidence about the encounter even when neither output enters the ordinary rate.

Write the policy while outcomes are hidden

A short protocol can settle most disputes:

  1. Define an ordinary eligible answer and the planned collection cell.
  2. Define how comparison interfaces, requested drafts and edited questions are classified.
  3. State whether a collector may select a preference. For observational collection, leaving it unselected often preserves the most useful state.
  4. Retain all completed visible outputs and document incomplete captures.
  5. Put comparison encounters and missing answers in separate reporting fields.
  6. Allow only a predeclared, later recovery in a fresh ordinary session; preserve the original event.

A recovery is a new observation on its actual date. It should not overwrite the comparison or backfill an earlier date with an answer obtained later. Repeatedly opening sessions until the ordinary interface returns would also make the sampling process depend on platform behavior in an uncontrolled way.

If an active limit or experiment prevents the planned collection, stop that cell and preserve its disposition. A monitoring report can explain a gap without disguising it as a negative recommendation.

Accept an export using a deliberately awkward fixture

Before relying on an answer archive, test it with an authorized demonstration fixture containing one encounter, two outputs and different source lists. This fixture can be synthetic, provided it is clearly labeled and kept out of customer-effect statistics.

The receiving analyst should be able to reconstruct that one question was submitted, neither preference was selected, two answer objects were captured and the encounter was excluded from a named ordinary-rate calculation. Sorting or joining the exported files should not collapse the two bodies or duplicate the parent as two submissions.

Then add an incomplete output. The archive should preserve the difference between a missing answer, a partial capture and a complete answer that does not recommend the brand. Those distinctions matter more than whether a dashboard displays an attractive overall percentage.

Semvi by Xindar's product scope describes AI-visibility monitoring. Product-team confirmation of full-answer and citation-URL exports is separate from acceptance of this particular two-output workflow. Buyers who require it should verify the actual interface, fields and export using their agreed test. This article does not claim that Semvi has passed that fixture or identifies a platform's internal experiment.

Reporting language that keeps the result usable

A useful line in a report would read: “Eight complete ordinary answers were retained from ten planned cells. One comparison encounter produced two preserved outputs and was excluded from the ordinary rate. One cell produced no answer because of a visible usage limit. Three of eight ordinary answers contained a verified recommendation.”

The statement tells the next analyst what was counted and what remains available. It does not imply ten negative-or-positive answers when only eight ordinary answers exist.

Xindar's published AI-visibility field record provides an example of separating ordinary valid records from excluded events. Its figures describe the dated version of that first-party record. They are not the current total of an ongoing project or independent proof of a monitoring product's capabilities. The historical URL is not the record's current collection cutoff.

Questions teams ask during review

Can we count the selected response if the platform keeps it?

Only under a policy that defines the selection process and the quantity being estimated. A collector-selected response is not automatically equivalent to an unselected ordinary answer. Preserve how the choice was made and avoid changing the policy after seeing the brand.

Does an excluded comparison mean the monitoring failed?

It means the encounter did not meet the ordinary-answer inclusion rule. Collection may still have succeeded in preserving a useful nonstandard encounter. Method eligibility, capture completeness and brand outcome are separate judgments.

Should the two responses be averaged?

An average answers a defined comparison question only if the units and weighting are justified. It does not recover the ordinary response that a buyer might have received in a different interface. Keep the paired evidence and its purpose visible.

For the next monitoring review, choose one retained comparison encounter and ask a colleague to reconstruct it from the exported records alone. If they cannot tell which question, panel, source list and inclusion rule belong together, repair that record before interpreting the rate.