Direct answer: Official AI citation data can establish that a publisher’s URL was displayed as a source on a supported AI surface, how often that display occurred, which pages were involved, and—in Bing’s public preview—a sample of retrieval phrases. It does not establish source rank, authority, answer accuracy, recommendation strength, causal influence, traffic, or conversion. Every report should preserve the platform, field definition, time range, aggregation rule, and denominator.
AI visibility measurement changed when platforms began exposing first-party fields. The difficult part did not disappear. It moved from “Can we observe anything?” to “What exactly did the platform count?”
A metric name can invite an interpretation broader than its definition. “Total citations” sounds like authority. “Average cited pages” sounds like reach. “Grounding queries” sound like user prompts. The official documentation supports narrower conclusions.
This article builds a metric dictionary from Microsoft and Google’s current public documentation and shows how to keep observation, inference, and business outcome separate.
Start with the event, not the dashboard label
The basic observable event is a source display: a URL from the measured site appears as a cited or referenced source in an AI-generated answer on a supported surface.
That event has several attributes:
- platform and included product surface;
- observation time;
- site and URL;
- generated answer or answer class, if available;
- retrieval or grounding phrase, if available;
- position or presentation, only if the product exposes it;
- click or visit, measured by a separate system;
- downstream action, measured after the visit.
If a field does not expose one of those attributes, analysts should not reconstruct it from the field name. A count of displays is not a rank. A source URL is not proof that the answer used the source correctly. A phrase used for retrieval is not necessarily the text the user entered.
Bing AI Performance: the four documented fields
Microsoft introduced AI Performance in Bing Webmaster Tools as a public preview on February 10, 2026. The announcement says the view covers Microsoft Copilot, AI-generated summaries in Bing, and selected partner integrations. Because it is a preview and aggregates supported surfaces, the product boundary belongs in every report.
| Official field | Microsoft’s documented counting idea | Defensible conclusion | Conclusion the field does not support |
|---|---|---|---|
| Total Citations | Number of citations displayed as sources during the selected period | The site was referenced this many times on covered surfaces | The site ranked first, was trusted, or caused the answer |
| Average Cited Pages | Average daily number of unique site pages displayed as sources over the period | This many distinct pages were cited per day on average | The pages had equal exposure or authority |
| Grounding queries | Key phrases used by the AI when retrieving cited content; shown as a sample | These phrases were associated with sampled retrieval activity | These are all user prompts or a complete query log |
| Page-level citation activity | Citation counts for specific URLs | These URLs were displayed more or less often in the selected period | The most-cited URL was the most important or highest ranked |
Microsoft explicitly says Total Citations does not indicate placement or presentation within an answer. It says Average Cited Pages does not indicate ranking, authority, or a page’s role in an individual answer. It says page-level counts reflect frequency rather than importance, ranking, or placement. It also labels grounding queries as a sample of overall activity.
Those sentences are the measurement boundary. A publisher can report “127 citations on supported Microsoft AI experiences in August.” It should not translate that into “127 top rankings” or “127 recommendations.”
Averages need a denominator and distribution
Average Cited Pages is a daily average across the selected range. A value of 10 could mean ten pages cited every day. It could also reflect a short spike followed by long silence. The average alone does not reveal concentration.
Pair it with:
- number of days in the selected range;
- median cited pages per day, if export data permits;
- active days with at least one cited page;
- top-page share of citations;
- new versus previously cited URLs;
- distribution by topic or page type.
These are analyst-derived summaries. Label them as such. Do not attribute them to Microsoft unless the interface or export provides the fields needed to calculate them.
A rise in unique cited pages can mean broader coverage. It can also mean the system rotated among many pages while total citation frequency remained flat. Read Total Citations and Average Cited Pages together, but keep their units distinct.
Grounding phrases are retrieval evidence, not a prompt transcript
Microsoft describes grounding queries as key phrases the AI used when retrieving content that was referenced. This is unusually useful because it exposes part of the retrieval layer. It can reveal that a page is found for a specific technical attribute, comparison, problem, or named entity.
It remains a sample. It is not necessarily:
- the user’s original wording;
- every subquery issued during query fan-out;
- a count of unique users;
- a search-volume estimate;
- proof that the phrase appeared in the final answer;
- proof that the cited page supplied the central claim.
Treat a grounding phrase as a retrieval association. Use it to inspect whether the cited page contains a clear, current, and properly qualified answer. Do not convert sampled phrases into demand forecasts without another data source.
Google has a dedicated AI impressions report alongside Web performance
Google's older AI-feature guidance places AI-feature traffic within overall Web search performance. That remains useful context, but it no longer describes the complete reporting surface. Google's August 31, 2026 update announced the worldwide rollout of a dedicated report.
The current Generative AI performance report documentation describes impressions of links from a property in AI Overviews and AI Mode. It offers page, country, device, and date breakdowns. It does not expose a prompt transcript, full answer, brand recommendation, or separate AI Overview-versus-AI Mode breakdown. Its impression unit should not be relabeled as Bing's citation unit.
Keep three boundaries in the metric dictionary:
- The AI report measures link exposure. It cannot establish that the answer recommended the linked company or used its evidence correctly.
- The Web report retains its own metrics. Clicks, impressions, CTR, and position describe a broader search result set. A conversational-looking query is not an AI-surface label.
- Aggregation and dates matter. Property totals can differ from page sums; report dates use Pacific Time. Do not add overlapping Web and AI totals as though they were unique users.
Save the selected filter, visible data extent, dimensions, and unavailable fields. A missing AI click or query field is unavailable, not zero. Link exposure still needs separate visit analytics and a declared attribution rule before discussing leads or sales.
Citation, mention, recommendation, and visit are separate events
Use a simple event ladder.
| Stage | Observable event | Example metric | Main limitation |
|---|---|---|---|
| Mention | Brand or entity appears in answer text | mention rate across frozen prompts | may have no source or positive judgment |
| Citation | Inspectable source URL is displayed | total citations, source-selection rate | display does not prove support or influence |
| Recommendation | Answer explicitly presents the entity as an option | recommendation rate or ordinal position | wording and list structure require coding |
| Visit | User reaches the site | referred sessions or clicks | attribution rules and untracked paths matter |
| Outcome | User completes a defined action | qualified lead, signup, order | affected by offer, page, sales process, and many other causes |
A citation can occur without a brand mention. A mention can occur without a citation. A recommendation can cite a third-party review rather than the recommended brand’s site. A user can see a recommendation and later visit directly. Each event needs its own source and denominator.
What official citation counts cannot establish
They do not measure citation support
A source can be displayed beside a sentence it does not support. Review claim-source pairs separately using a documented support rubric.
They do not measure textual influence
A displayed source may contribute the central fact, a minor qualification, or nothing discernible. Source contribution requires answer and source analysis, and causal contribution requires an intervention.
They do not measure authority
Frequency may correlate with broad topical coverage, freshness, or retrieval availability. The official Bing definitions explicitly reject interpreting page counts as authority.
They do not measure rank
Unless a platform exposes a defined order or placement field, do not assign a rank from source-panel appearance. Interfaces can group, collapse, or reorder sources.
They do not measure conversions
Conversion occurs downstream. Connect citations to referrals and referrals to outcomes only under a declared attribution rule. Preserve unattributed and direct return visits rather than forcing credit.
Build a metric dictionary before building a dashboard
For every field, record:
- Official name and definition. Save the current documentation URL and access date.
- Platform boundary. Name the products, surfaces, regions, and accounts included.
- Unit. Citation displays, unique pages per day, sampled phrases, impressions, clicks, sessions, or conversions.
- Aggregation. Sum, daily average, percentage, rolling window, or sampled list.
- Dimensions. Date, page, phrase, market, device, or other available breakdowns.
- Missing-value rule. Distinguish zero, unavailable, suppressed, delayed, and outside scope.
- Permitted conclusion. Write one sentence the metric can establish.
- Forbidden conclusion. Write the tempting overclaim it cannot establish.
- Version. Record documentation and export versions because preview fields can change.
This dictionary should travel with exports. A chart detached from its counting rules becomes difficult to audit after the platform changes.
A defensible monthly reporting layout
Use four panels rather than one blended score.
Panel 1: First-party platform observations. Report Bing citation fields, Google Generative AI link impressions, and Google Web performance separately under their documented counting rules. Do not normalize unlike units into one index.
Panel 2: Controlled prompt observations. Track mention, citation presence, recommendation, and support across a frozen prompt panel. Mark these as your own sample, not platform census data.
Panel 3: Site behavior. Report referred sessions, engaged sessions, and defined conversions with attribution settings and consent limitations.
Panel 4: Evidence health. Track whether priority facts are current, whether cited pages remain accessible, and whether important claims retain sources and qualifications.
Xindar’s public measurement specification follows the same separation principle: mention, recommendation position, citation presence, citation support, qualification coverage, and risk disclosure are individual dimensions. It does not approve a composite score without documented weights, missing-value handling, confidence intervals, and a minimum sample size.
Frequently asked questions
Is Total Citations the number of AI answers containing my site?
Do not assume so. Microsoft defines it as the total number of citations displayed as sources during the selected period. Multiple citations and aggregation across surfaces affect interpretation.
Are Bing grounding queries the user’s original searches?
Microsoft describes them as key phrases used by the AI for retrieval and says the display is a sample. Report them as grounding phrases, not verbatim user prompts.
Can Search Console isolate AI Overview or AI Mode citations?
Search Console has a dedicated Generative AI report for link impressions across AI Overviews and AI Mode, alongside overall Web performance. The documented AI report does not separate those two surfaces or provide full answers, prompt text, or recommendation judgments. Keep its impression count separate from Bing's defined citation count.
Can I combine all platforms into share of voice?
Only after defining a common event, prompt frame, weighting rule, and missing-data policy. First-party fields with different units should remain separate.
Source and method note
The Google section was revised on November 2, 2026 against the current AI report definition, Web performance definitions, and rollout announcement. The prior September 9 version omitted the dedicated AI report; its statement about Web aggregation was incomplete as a description of available reporting.
Microsoft's February 2026 public-preview announcement was rechecked on November 2. Its four field definitions above remain source-bound. Xindar's measurement specification is retained as a first-party method reference, not an independent endorsement. No private account export was analyzed for this article, and no ranking, authority, traffic uplift, or conversion result is claimed.