Get a free AI visibility report
.png)
Copilot citations are decided by Bing's retrieval layer, not by your position in a ranked list. This guide covers how grounding queries work, how Copilot selects sources, and what to change to get cited.
Teams usually notice the problem in a sales call. A buyer says they asked Copilot which vendors to look at, and your name was not in the answer. Nothing in Bing Webmaster Tools looked wrong that week. Your pages are indexed, your rankings are fine, and yet the answer assembled itself out of four other domains.
That gap has a specific mechanism behind it, and the mechanism is knowable. Microsoft has published more about how its AI answers select sources than any other platform operator. It even ships a first-party citation dashboard. This guide works through that documentation and turns it into a sequence you can act on.
Primary lens: citation supply chain. The useful question is not whether your site is optimized. It is which domains Bing already trusts for your category's grounding queries, and what it takes to join them.
Before changing anything, get the pipeline right in your head. Three mechanics matter, and each one has a practical consequence.
Copilot does not maintain a separate web index. It retrieves from Bing, so everything you know about Bing indexing applies directly to Copilot citations. In its guidance on AI search answers, Microsoft describes Copilot and Microsoft Start as powered by Bing's search index, handling billions of queries each month.
The practical consequence is blunt. If a page is not in Bing's index, it cannot be cited. Teams that treated Bing as a rounding error on organic traffic now have a fresh reason to care. That reason has nothing to do with Bing's blue-link market share.
This is also the clearest case of model divergence in the category. Other assistants build their own retrieval paths. A page cited constantly on one platform can be absent from another. Copilot is the one surface where a traditional search index sits directly underneath the answer.
This is the detail most teams get wrong. When someone asks Copilot a conversational question, the system does not search Bing for that sentence. It generates its own retrieval phrases first.
Microsoft calls these grounding queries. In its AI Performance announcement, it defines them as the key phrases the AI used when retrieving cited content. A single user question can produce several of them, each pulling a different set of candidate pages.
The consequence is that your keyword research and your grounding query data will not match. Users ask long, messy, conversational questions. The retrieval layer rewrites them into something more structured before it searches. Optimizing for the user's phrasing while ignoring the retrieval phrasing leaves you targeting the wrong string.
Microsoft has been unusually direct about this. Assistants like Copilot break content down through a process Microsoft calls parsing. The output is smaller, structured pieces that get evaluated for authority and relevance. Those pieces are then assembled into an answer, often drawing on several sources at once.
Microsoft frames the shift plainly. Ranking still happens in AI search, but it is less about ordering whole pages. It is about which pieces of content earn a place in the final answer.
Read that twice, because it changes what a page is for. You are no longer competing to be the best page. You are competing to own one extractable piece of the answer, alongside three or four other domains.
Once you accept chunk-level assembly, the strategic picture reorganizes itself. An answer is not a prize awarded to one winner. It is a small supply chain, and you are trying to become one of its suppliers.
A typical Copilot answer draws on several domains. Some will be review platforms, some industry publications, some competitors. One might be you. Spending the entire budget on owned pages optimizes a single link in a chain you do not control.
The alternative is to work backwards. Identify the answers you want to appear in. List the domains currently supplying them, then treat that list as a target set. Our guide to tracking competitors in AI search results covers the competitive side of that exercise.
Source mention rate is the metric for this. It measures how often a given domain gets cited across a prompt set. That reveals which third-party properties the retrieval layer trusts on your topic. Cognizo reports it per prompt and per platform, so the Copilot picture stays separate from the others.
Patterns emerge quickly. In most B2B categories, three to six domains supply a disproportionate share of citations. They are rarely the vendors themselves. Getting accurate, current information onto those domains moves Copilot citations faster than another owned blog post will.
Prioritize placements by two things at once: how often a domain gets cited in your category, and how easy it is to get accurate information onto it. A review profile you already own scores high on both. A trade publication that cites nobody in your space scores low on the first, whatever its domain authority looks like.
Microsoft shipped something no other platform operator has. In February 2026, Bing Webmaster Tools added an AI Performance report in public preview. It shows where and how often your content is cited across Copilot, AI-generated summaries in Bing, and select partner integrations.
The report covers total citations, average cited pages, grounding queries, page-level citation activity, and visibility trends over time. Access requires nothing more than a verified site. If you have Bing Webmaster Tools set up, the data is already waiting.
Indexing is a prerequisite, not a strategy. Work through it first so the later steps have something to act on.
Verify the site in Bing Webmaster Tools if you have not, then check index coverage against what you expect. Pay particular attention to pages that matter commercially: pricing, comparisons, integration documentation, and detailed use-case pages.
Microsoft notes that Bing respects content owner preferences expressed through robots.txt and other supported control mechanisms. Blanket crawler blocks written for a different purpose will quietly remove you from the candidate pool.
Freshness carries real weight in retrieval. Microsoft points site owners to IndexNow specifically in the context of AI answers. Faster discovery of changes helps AI systems reference the current version of a page.
For a brand with pricing or product details that move, this matters more than it sounds. A stale cached version lets Copilot cite accurate-looking information that is nine months old. The buyer has no way to tell.
Microsoft's own guidance lists the patterns that reduce eligibility, and they are worth taking literally. Long undifferentiated walls of text blur ideas together. Content hidden inside tabs or expandable menus may never render for the system reading the page. Core information locked in PDFs loses the structural signals that HTML provides. Key details presented only inside images add extraction complexity and reduce accuracy.
Each of those is a common design pattern on B2B sites. Comparison data in an accordion, pricing in a PDF, specifications in a graphic. All are normal choices that quietly cost Copilot citations. An llms.txt file helps machine readers navigate what you do publish, and our guide to llms.txt covers the format.
This is the step almost nobody runs, and it is the one with the most upside. Grounding query data is genuinely new information, not a proxy for something you already had.
Be precise about the limits. Microsoft states that total citations reflect how often content is referenced, without indicating placement within a specific answer. Page-level activity reflects citation frequency rather than page importance, ranking, or placement.
Search Engine Land made the same point at launch. The metrics do not indicate ranking, prominence, or how a page contributed to a specific answer. There is no click data either. Treat the report as a frequency signal. Do not let anyone read it as a traffic number.
Grounding queries are also a sample rather than a complete list, which Microsoft says it will keep refining. Directional is still useful when the alternative was nothing.
In March 2026, Microsoft connected the two views. Query-to-page mapping went live. Click a grounding query to see the pages cited for it, or click a page to see the queries driving its citations. The mapping is many-to-many.
That upgrade is what makes the report actionable. A list of queries and a separate list of pages is trivia. The connection between them is a work queue.
Run the following pass once a month. Export your grounding queries, group them by topic, and mark which ones already map to a page of yours. The unmapped ones are your gap list. They are phrased the way the retrieval layer phrases things, not the way your keyword tool does.
Then widen the aperture. The report only shows queries where you were already cited once. It is structurally blind to topics where you have zero presence. Cognizo's Prompt Volumes module fills that blind spot. Built on billions of real-world signals, it shows what buyers ask across the whole category.
Now the content work. The target is not a better page overall. The target is a page made of clean, liftable pieces.
Microsoft treats these three as interpretation signals for purpose and scope. It recommends that the H1 closely reflect the page title while setting clear expectations. Consistency between them improves the confidence signals a system has when classifying your page.
This is cheap to fix and frequently broken. One title in the tab, a different H1, a meta description written by a third person. That gives the retrieval layer three conflicting stories about one page.
Microsoft compares headings to chapter titles that define clear content slices. A vague heading gives the system no boundary to cut on. A specific question or claim as a heading marks exactly where one idea starts and stops.
Write headings as the question a buyer would ask, not as an internal label. "Pricing" is a label. "What does the platform cost for a 20 person team" is a boundary.
Microsoft is explicit that direct questions with clear answers can be lifted word for word into AI responses. Lists and tables break details into clean, reusable segments. Comparison tables and numbered steps are called out specifically for how-to and feature-comparison queries.
Use them where they are honest. A table comparing four real options earns its place. A table invented to hit a format target adds noise and gets ignored. Schema markup in JSON-LD reinforces the structure. Labelling content as product, review, or FAQ means the machine reading it does not have to infer.
Microsoft's snippet checklist names four things: concise one or two sentence answers, structured formatting, strong headings, and self-contained phrasing. Cognizo's Content Optimization module applies exactly this where your visibility data shows a gap. It generates briefs, outlines, drafts, and FAQ blocks rather than leaving the rewrite to guesswork.
One more detail from the same guidance, and it is a satisfying one. Microsoft advises caution with em dashes, since overuse can confuse sentence structure for machines. A period or semicolon is clearer. It also warns against decorative symbols and unanchored claims like "next-gen" that give a system nothing to classify.
Owned pages are the part you control and the smaller part of the answer. This step is where most of the remaining upside sits.
Review platforms, industry publications, comparison articles, and community threads all feed the candidate pool. They carry independence your own pages cannot. In practice the earned side of citation share outweighs the owned side across every major platform.
Approach it as a placement program rather than a link-building one. What gets cited is a sentence containing a specific, verifiable fact. Give writers real figures instead of adjectives. A publication that prints a number creates something quotable. One that prints "streamlines operations" creates nothing.
Microsoft advises reducing ambiguity across formats. Align text, images, and video so they represent the same entities and products consistently. Extend that instruction past your own site.
Suppose your homepage, your review profiles, and last year's press coverage describe your category three different ways. The retrieval layer now has three competing descriptions. Positioning accuracy problems usually trace back to that inconsistency. They are worse than absence, since the buyer reads a confident wrong answer and moves on.
For location-based questions, Microsoft points businesses to Bing Places for Business. Keeping address, hours, and contact details current keeps them eligible for AI answers. Multi-location brands should treat that as a recurring data hygiene task rather than a one-time setup.
The AI Performance report answers one question well: how often did Copilot cite me. It leaves several others open, and those are the ones executives ask about.
Cognizo's framework covers six numbers, and each answers something the Bing report cannot.
MetricWhat it addsVisibility ScorePercentage of tracked prompts where your brand is mentioned, including answers that never cite you with a linkShare of voiceYour proportion of mentions relative to named competitorsCitation shareYour proportion of cited sources, split into owned and earnedSource mention rateWhich third-party domains the platform trusts on your topicSentimentWhether Copilot describes you positively, negatively, or neutrallyPositioning accuracyWhether your category and capabilities are described correctly
The distinction that matters most is mention versus citation. Copilot can recommend your brand without linking to your domain at all. The Bing report will never see that. Visibility Score catches it. Our guide to checking whether your brand appears in AI search walks through the manual version of the check.
Copilot data alone is a partial view. A brand can be strong on Copilot and absent on ChatGPT or Google AI Overviews, because the retrieval paths differ. Tracking each platform separately is the only way to see that. Cognizo's Answer Engine Insights module breaks results down by model, topic, prompt, and region.
Cadence matters too. Citation activity shifts as the index refreshes and coverage changes. A monthly look tells you something moved without telling you why. Daily tracking against a fixed prompt set is what connects a change you shipped to a change in the answer.
Two numbers travel well internally. Citation count from the Bing report shows Microsoft's own view of how often your pages get used. Visibility Score shows how often your brand appears at all, including answers that never link to you.
Present them together and the picture holds up. Present citation count alone and someone will eventually ask what it converted, at which point the honest answer is that the report does not say. Setting that expectation early is easier than retracting a number later. Our guide to improving AI search visibility covers the wider reporting frame.
Most Copilot answers carry no link, and the AI Performance report has no click data. The realistic path runs from mention to branded search to direct visit, which no UTM parameter will ever capture. Add a "how did you hear about us" field with an explicit AI option. Watch branded search alongside your citation counts.
Cognizo's AI Traffic Analytics closes part of the gap from the other side. It tracks crawler behavior from major bots alongside referral sessions and conversions per platform.
Five patterns account for most of it.
Ignoring Bing because of its organic share. Bing's index is the retrieval layer for Copilot. Its blue-link market share is beside the point.
Optimizing for user phrasing only. Grounding queries are generated by the retrieval system, not typed by the user. Read them and write to them.
Publishing walls of undifferentiated text. Chunk-level selection needs boundaries. No headings means nothing to cut on.
Hiding answers in tabs, PDFs, and images. All three lose the structural signals that make content extractable.
Reading the Bing report as a traffic metric. It reports citation frequency. It says nothing about clicks, ranking, or prominence, and presenting it otherwise will damage your credibility internally.
The AI Performance report is free, first-party, and worth checking. It is also limited to Microsoft surfaces, blind to prompts where you have no presence, and silent on how you are described. Cognizo covers the rest.
Autopilot is the content agent with auto-publishing, and it ships on every plan. Agents handle market research, prompt planning, content production, and publishing. That turns Copilot citations from a quarterly audit into a running system. A gap appears, a brief gets generated, content ships, and visibility gets re-measured against the same prompt set.
There is a small piece of relevant history here too. One of Cognizo's cofounders worked on the Microsoft Copilot team from 2019 to 2025. A pleasant coincidence for a product that now tracks it.
Answer Engine Insights monitors Copilot alongside other supported platforms. It covers visibility, share of voice, citation share, source mention rate, sentiment, positioning accuracy, and competitive benchmarking.
Content Optimization converts that data into prioritized recommendations, automated briefs and drafts, and technical audits. Its toolbox spans PR, affiliate, social, and owned media.
Prompt Volumes shows what buyers actually ask, built on billions of real-world signals. It covers the topics your Bing report cannot see.
AI Traffic Analytics ties crawler behavior to referral sessions and conversions per platform.
ChatGPT Ads pairs organic visibility with paid placement and surfaces competitor ad copy.
Cognizo also captures rendered answers through UI scraping. The data reflects what a user actually sees, not API sampling alone. Our Copilot rank tracker comparison goes through that methodology question in detail. The explainer on how Copilot rank tracking works covers the mechanics.
Growth is $499 per month. It includes 8 AI-optimized articles, Autopilot, 150 tracked prompts, and 5 selectable platforms, Copilot among them. Pro is $999 per month with 20 articles and 350 tracked prompts. Enterprise is custom and covers all 10 supported engines. It adds tailored content volume, multiple brands and workspaces, a dedicated AI search strategist, and SSO, SLA, and enterprise controls.
Every plan includes daily tracking, unlimited seats, regions, languages, and competitors, plus MCP and API access. Programmatic access on entry tiers is unusual in this category. It matters for teams that want citation and visibility data inside their own warehouse, dashboards, or agent workflows.
A grounding query is the retrieval phrase Copilot's system generates when it searches for content to support an answer. It is not the question the user typed. A single conversational prompt can produce several grounding queries, each pulling a different set of candidate pages from Bing's index. Microsoft surfaces a sample of these in the AI Performance report. That is the closest available view of how the retrieval layer interprets a topic. Writing to grounding query phrasing is usually more effective than writing to the user's original wording.
You need to be indexed in Bing, which is not the same thing. Copilot retrieves from Bing's index, so a page absent from that index cannot be cited under any circumstances. Beyond indexing, the relationship with classic ranking is loose. Selection happens at the chunk level, evaluating individual pieces of content for authority and relevance rather than ordering whole pages. A page ranking modestly can supply the cited chunk if that chunk answers the grounding query cleanly.
The Bing report is first-party, free, and limited to Microsoft surfaces. It tells you how often your pages were cited and which grounding queries drove it. There is no click data and no indication of ranking. It cannot show prompts where you have zero presence, answers that mention you without linking, how you are described, or other platforms. Those gaps are what a dedicated visibility platform covers, and the two work better together than either does alone.
Indexing and freshness fixes move fastest, sometimes within days once IndexNow or a corrected robots.txt lets Bing re-discover the pages. Structural content rewrites typically register within two to six weeks, depending on re-crawl frequency. Earned coverage runs slowest, usually one to two quarters, since it depends on publication cycles you do not control. Sequence the work accordingly and judge the slower streams on leading indicators rather than on citation counts alone.
Yes, and this is a significant blind spot. Copilot can recommend a brand in an answer while citing review sites, publications, or competitors for the supporting links. Your Bing dashboard records nothing in that case, because no page of yours was used as a source. Visibility Score, which measures the percentage of tracked prompts where your brand is mentioned, catches these answers. A brand can be well recommended and barely cited, or heavily cited and poorly described. Only tracking both reveals which you are.
Microsoft's guidance names four specifically. Long undifferentiated walls of text give the system no boundaries to parse on. Content hidden in tabs or expandable menus may not render at all. Core information published only as PDFs loses the structural signals HTML provides. Key details placed only inside images add extraction complexity and reduce accuracy. Each pattern is common on B2B sites, particularly for comparison tables, pricing, and specifications. Those pages carry the most commercial weight in AI answers.
It helps, though it is not a switch. Schema in JSON-LD labels content as a product, review, FAQ, or event. That turns prose into structured data a machine can interpret without inference. That reduces ambiguity, which matters most in crowded categories where several companies have similar names or overlapping descriptions. Treat schema as a supporting signal that reinforces good structure rather than a substitute for it. Clear headings, self-contained answers, and accurate entity descriptions do more of the work.
Continuously, and from two places. The Bing AI Performance report is worth a monthly export for grounding query analysis. That data updates on Microsoft's cadence. Prompt-level tracking should run daily against a fixed set. Only that connects a change you shipped to a movement in the answer. Monthly-only checking tells you something shifted without telling you why. That is the difference between a report and a decision.
.svg)
.svg)
