Discussion
  • AI Search

Bing’s AI Performance report shows Copilot citations and grounding queries but no clicks. Are the grounding queries useful for content planning, or too sampled?

Published 1 reply

Discussion starter: an AI search question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.

Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.

1 reply

  • admin
    Admin
    Original poster

    They are useful, but for a narrower job than keyword research. Grounding queries show the search phrases Bing’s AI used when it retrieved pages of yours that ended up cited. That makes them a good qualitative map of how the retrieval system describes the topics it associates with your content. It does not make them demand data: they are sampled, they carry no search volume or clicks, and they only cover cases where you were already cited.

    So use them to diagnose and improve pages you already have, and to spot sub-questions and qualifiers you have not covered. Do not use them to size content investments, forecast traffic or rank topics by count, especially when the counts are small.

    What the report actually contains

    Microsoft launched AI Performance in Bing Webmaster Tools as a public preview in February 2026, and added a link between the Grounding Queries and Pages views in March 2026. The documented details decide how much weight the data can bear.

    Element What Microsoft says What it means for planning
    Coverage “Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations” Microsoft surfaces only. Partners are not named, so do not assume the data describes ChatGPT or other assistants.
    Grounding queries “Shows the key phrases the AI used when retrieving content that was referenced in AI-generated answers.” Machine-written retrieval queries, not user prompts, and only where your content was referenced
    Sampling “The data shown represents a sample of overall citation activity.” Sampling rate and method are not published; small counts are noisy
    Interpretation limits The data “does not indicate ranking, authority, or the role of any page within an individual answer.” A citation count is not a position or a measure of prominence
    Query-to-page mapping (March 2026) “One grounding query may be mapped to multiple pages and vice versa.” Useful for spotting the wrong page being cited and internal competition
    Clicks and traffic Not reported Any traffic or revenue view has to come from analytics

    A grounding query is the machine’s search, not the user’s question

    Microsoft has not published the full retrieval pipeline, but its definition implies a familiar chain: a user asks a conversational question, the system turns it into one or more search phrases, Bing retrieves candidate pages, and the model chooses which ones to cite. The grounding query sits in the middle of that chain.

    User prompt in Copilot or a Bing AI summary (conversational, often long)
    ↓
    AI generates one or more search phrases: the grounding queries
    ↓
    Bing retrieves candidate pages
    ↓
    Is your page selected as a cited source?
    ↓
    If yes, the grounding query may appear in your report (sampled)

    Two consequences follow. First, grounding queries are likely to be shorter and more keyword-like than the prompts behind them, and one prompt can produce several. They show how the system breaks a topic down, not how people phrase their questions. Second, the data has a survivorship problem. You see queries where you were retrieved and cited. You do not see queries where you were retrieved but passed over, or where competitors were cited instead. A report built only from your wins cannot show your biggest gaps.

    Think of grounding queries as the request slips a research assistant filled out at a library, except you only see slips for books from your shelf that made it into the final report, and only a sample of those. They reveal what the assistant was looking for when it found you. They say nothing about how many people asked the assistant, or what it was looking for when it chose someone else’s book.

    What “sampled” means for the decisions you make

    The announcement says the data is a sample, and that Microsoft “will continue to refine this metric as additional data is processed”. It does not describe the rate or method, so you cannot assume sampling is uniform across topics. Three working rules follow:

    • Aggregate before you interpret. Individual queries with single-digit counts will appear and disappear week to week. Cluster queries into topics by intent, entity or product, and read the topic totals over 28 to 90 days.
    • Compare shares, not raw counts. A topic’s share of your total citations is more stable than its raw count, and it survives any change in sampling rate that affects all topics equally.
    • Treat absence as unknown, not zero. A query missing from the list may simply not have been sampled in that window.

    Because this is a preview and Microsoft says the metric is still being refined, annotate any abrupt shift in the trend line before crediting or blaming your own work.

    Where grounding queries earn a place in content planning

    Use Value Why Pair it with
    Finding the wrong page being cited High The query-to-page view shows, for example, an old blog post cited for a pricing or compliance question Page audit, internal links, update or consolidate
    Finding sub-questions and qualifiers High Grounding queries expose the facets retrieval looks for: cost, comparisons, location, year, integrations, industry SERP review, sales and support questions
    Accuracy and brand risk checks High Frequently cited pages shape what Copilot says about you; outdated facts spread Freshness reviews; Bing recommends IndexNow for keeping content current
    Vocabulary checks Medium If a qualifier appears persistently and your page never addresses it, human readers probably miss it too Bing and Google Search Console query data
    Prioritising new topics by volume Low Sampled, no volume, and only covers existing citations Keyword tools and Search Console demand data
    Competitive gap analysis Low Shows only your own citations Manual prompt checks or third-party AI visibility tracking
    Forecasting traffic or revenue None No clicks; a citation is not a visit Analytics referral data

    Retrieval, citation and traffic are three different outcomes:

    Being retrieved, being cited and receiving a visit are separate events. The Bing report measures the middle one. Google’s generative AI reports in Search Console show AI impressions but no queries or clicks, so Bing’s grounding queries are currently one of very few first-party views of retrieval language at all. That scarcity makes them valuable. It does not make them complete.

    Keep the business context in proportion too. Third-party data suggests ChatGPT accounts for the large majority of AI referral traffic (BrightEdge data released in September 2026 put it at about 95.1% of tracked AI referrals, without detailing methodology). The Microsoft surfaces covered here are a slice of AI visibility, so avoid building a whole content roadmap around them. They are most useful as a window into retrieval behaviour that other platforms do not expose.

    Hypothetical example:

    An Australian payroll software company exports 90 days of grounding queries and clusters them. Three patterns stand out. Queries about Single Touch Payroll requirements map to a 2023 blog post rather than the current compliance guide, so the older post is shaping answers with outdated detail. A cluster of “payroll software for [industry]” queries maps to the generic features page, suggesting retrieval is looking for industry-specific fit the site never explains. And “[brand] pricing” maps to a reseller partner page instead of the pricing page. None of those findings depends on the counts being precise. Each points to a specific page fix, which is exactly the kind of decision this data supports.

    How to test whether acting on grounding queries pays off

    1. Hypothesis: improving the page that best answers a grounding-query cluster, or fixing a mismatch where the wrong page is cited, increases citations for that cluster and referral visits from Microsoft’s AI surfaces.
    2. Groups: pick two to four clusters to change and a similar number of comparable clusters to leave alone.
    3. Change: update the pages, submit them through IndexNow, and record the date.
    4. Primary metric: citation counts for the target pages and each cluster’s share of total citations, compared with the control clusters over six to eight weeks. Secondary: referral sessions from Copilot and Bing in analytics, and leads from those sessions. GA4’s AI Assistant channel may catch some of this, but Google has not published its full source list, so check where Copilot referrals actually land in your property.
    5. Confounders: model updates, Bing index changes, seasonality and the preview’s own metric refinements. This is why a control group matters more here than in most SEO tests.
    6. Decision rule: if treated clusters show a sustained rise relative to control across several weeks and more than one page, grounding-query-led updates earn a regular slot in the content calendar. If they move no more than the control, keep using the report as a diagnostic for mismatches and accuracy, and plan new content from demand data instead.

    My position: grounding queries are too sampled to plan by volume and too informative to ignore. Use them as a map of how Microsoft’s AI breaks down your topics and which of your pages it relies on, fix what that map exposes, and keep demand-led research as the basis for deciding what new content to build.

    Need help with this on your own site?

    Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.

    See SEO services

Add a reply

Replies are open to approved members. Log in or apply to join.