DeepSmith

Nov 26 · AEO & AI Visibility

18 min read

How to Track Microsoft Copilot Visibility and Rankings

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
A dark monochrome cover reading Track Copilot Visibility, surrounded by a chat bubble, connected source-link cards and faint bar-chart fragments.

If you want to know whether Microsoft Copilot mentions your brand when a buyer asks about your category, the only reliable way is to ask the questions yourself and write down what comes back. This guide walks you through copilot visibility tracking from start to finish: choosing the prompts, running them in a controlled way, recording whether your brand is mentioned and cited, noting where it shows up inside the answer, and repeating the check on a schedule so you can read trends instead of single results. It's written for marketing leads who need a process they can hand to one person and run every week.

You'll need a Copilot account you can sign in to, a spreadsheet, and about an hour for the first round. The steps below are in the order you'd do them.

Step 1: Decide which Copilot you are testing

Before you type a single prompt, decide which Copilot experience the test represents. Copilot for a personal Microsoft account and Copilot inside a company's Microsoft 365 setup are not the same thing, and Microsoft's privacy documentation for individuals covers personal accounts only. If your buyers are general web users, test the public Copilot experience with a personal account and say so in your notes. Don't mix that with a work or school account in the same data set.

Write down the details of the setup so anyone could repeat it: account type, the Copilot experience you used, language, your region if you know it, device and browser, and the date and time of each run. Then keep that setup the same from week to week. If you also watch other engines, a list of Google AI Overviews tracking tools shows what the same job looks like there.

Done when: you can describe the test environment in two or three lines, and a teammate could set up the same one.

Common mistake: switching accounts or devices halfway through and then treating the results as one series. If something has to change, write down the date it changed.

Step 2: Build a small prompt set from real buyer questions

Start with the questions a potential buyer would actually ask Copilot, and sort them by intent. A few should be category discovery ("what tools help with X"), a few should be comparisons, a few should be evaluation questions, and a few should be close to a decision. Use natural wording the way a buyer would type it.

Try not to put your brand name in every prompt. A prompt that names your brand tends to cue a mention, and that hides the thing you're trying to find out, which is whether Copilot brings you up on its own.

A good starting size is 10 to 20 prompts. That number is a practical suggestion and not a Microsoft standard or a statistical rule, so adjust it as you learn which questions give you useful evidence. Give each prompt an ID, save the exact wording, and note the date. Microsoft describes a prompt as an instruction or a question, and it notes that the same prompt can give different responses when it's run more than once, so the exact text is worth keeping.

If you'd rather not build the list from a blank page, DeepSmith's AI Visibility module has Discover Prompts, which suggests a starter set from your product, persona, and buyer-stage context. You define the questions you care about, and the platform runs them on a schedule and keeps the answers.

Done when: every prompt maps to a real buyer question and has an ID, exact text, and an intent label.

Common mistake: rewording a prompt between reporting periods without noting it. A reworded prompt is a different test, so the before and after can't be compared cleanly.

Step 3: Standardize the mode and the settings

Copilot has several response modes, including Quick response, Think Deeper, Smart, Study and learn, and Search. For a check on web-based visibility, use Search. Microsoft describes it as the mode that returns up-to-date web answers with citations. Pick the mode before you submit the prompt, because it stays on for follow-up questions until you turn it off. Some modes may need you to sign in. Keep the modes separate in your data, since a Think Deeper answer and a Search answer aren't the same kind of evidence.

Use a fresh conversation for each independent run. Copilot builds its answers from your prompt and the recent conversation, so a follow-up inside an old chat isn't a clean test. This is also why one of the most helpful habits in copilot visibility tracking is to start a new chat every time.

If you're testing with a personal account, check the personalization controls under Settings > Personalization. Microsoft lists Saved memories, One shared experience, and Web search there. Turning off saved memories stops Copilot from saving new ones, but it doesn't erase the ones already stored. To clear those, use Manage > Delete all memories. Whatever you choose, note whether web search was on and keep the configuration the same for every run. These settings help you stay consistent, but they don't promise that every personal or contextual effect is gone, so don't describe them that way in a report. The same goes for location. Note your region if you know it, but don't assume you've controlled for it.

Done when: your log names the account type, mode, key settings, and the fact that each run started in a new conversation.

Pro tip: put the settings in a header row on your sheet and copy it into every new week's tab. It's easy to forget one toggle and much harder to find out later which week it changed.

Step 4: Run the prompts and save the evidence

Submit the exact prompt in Search mode, and save everything that comes back: the full response, the visible citations and their URLs, the time of the test, the mode, the prompt version, and any interface details you'd need to repeat it. Save the response even when your brand isn't in it. An absence is a real result, and it's often the most useful one.

For web-grounded text answers, Microsoft's consumer Copilot documentation says hyperlinked citations appear below the answer. Open them and record the actual URL, not just the name of the source that's displayed. Copilot can answer in different formats, so stick to a text-based method when you judge mentions and citations, and the notes you keep stay comparable.

Copilot also lets you share a whole conversation or a single response. A shared response includes the prompt that produced it, and anyone with the link can see what's in it, although they need to sign in to view it. Shared items are snapshots, a link can be stopped, and deleting the original conversation also stops its sharing links. That makes sharing handy for passing evidence to a colleague, but not a good permanent archive. Keep your own record as well, whether that's screenshots, copied text, or a document per run.

Done when: each run has the exact prompt, the complete answer, the cited URLs, a timestamp, the mode, and a saved copy of the evidence.

Common mistake: only noting "yes" or "no" for the brand and throwing away the answer text. Later, when a number changes and someone asks why, you'll want to read what Copilot actually said.

Step 5: Code mentions, citations, and position as separate fields

This is the step where most tracking sheets go wrong, so it's worth slowing down. For every response, fill in separate fields instead of one "visible" flag.

FieldWhat to record
Brand mentionedYes or no, based on whether the response names your brand
Brand citedYes or no, based on whether at least one linked source is one of your pages
Own cited URLsEvery one of your URLs linked in the response
Competitors mentionedNames of competitors in the response
Competitor cited URLsCompetitor pages linked, when visible
Explicit list rankThe brand's place in the list, only when Copilot gives an ordered list
First-mention locationOpening, middle, or end, when there's no ordered list
Citation locationWhere your source link appears in the response

The reason to separate them is that an answer can name your brand without linking to any of your pages, and it can link to one of your pages without naming you. Those two cases mean different things for your content, so a single flag hides the difference. DeepSmith keeps the two apart too, reporting mention rate and citation rate separately.

Position needs the most care. If Copilot gives an explicitly ordered list of brands or products, you can record your ordinal place in it. If the answer is ordinary prose, record where your brand first appears (opening, middle, or end) and call it first-mention location, not rank. If your brand isn't in the answer, write "not applicable" for position. Don't enter a zero, and don't treat a missing mention as last place. And keep citation order out of it. The order of source links is not a ranking of your brand, and it shouldn't be reported as one.

It also helps to be honest about what this measure is. Answer-level position is something you score by hand using a rule you define, and it is not a Microsoft metric or an industry standard. If you want more background on why placement inside an answer is hard to pin down, does position matter in AI answers goes through it, and the same caution applies here. That's what makes copilot rank tracking a little different from classic search rank tracking: you're recording an observation about one answer, not a place in an index.

For rates, use a clear denominator, which is the number of valid, completed responses for the same prompt set and period. If your brand is mentioned in 2 of 3 valid runs, the observed mention rate for that sample is 2 out of 3, or 66.7%. If one of those three responses links to one of your pages, the citation rate is 1 out of 3, or 33.3%. These are only examples of how to do the arithmetic, not targets. If you want an average explicit rank, calculate it only across responses that had an ordered list, say how many that was, and keep prose answers and no-mention runs in their own groups.

Done when: every response has its own row with every field filled in, or marked "not applicable" on purpose.

Common mistake: counting a mention as a citation, or turning the first source link into a rank. Both blur two different things into one number.

Step 6: Repeat on a fixed schedule

One check tells you very little. Microsoft says plainly that the same prompt can produce different responses, because language models vary their output, and that recent conversation history and changing features also play a part. So you need several runs before a pattern means anything.

A reasonable starting design is to run each prompt three times, each in its own fresh conversation, during one collection window, and to repeat that window every week. A weekly check is a sensible choice for a lean team, but it's an editorial suggestion and not something Microsoft prescribes. Three runs is also a practical start and not a validated sample size, so add more runs if your results swing enough to change a decision.

Log every run, not only the final status of each prompt, and show the numerator and denominator next to each rate, for example "2 mentions in 3 responses." That lets anyone reading your report see how much evidence is behind a percentage. Compare like with like: the same prompt wording, mode, account setup, and time window. If one of those changes, mark it in the record so it doesn't quietly turn into a fake trend.

Done when: you have a calendar slot for the check, a fixed number of runs per prompt, and a sheet with one row per run.

Pro tip: always show the run count beside a rate. A 100% result from a single answer is not the same as steady visibility across several tests, and the second is what you actually want to report.

Once you have a few weeks of data, group the rows by prompt and by period. Look for sustained movement in mention rate, citation rate, which of your pages get cited, which competitors appear, and your answer position when you can measure it. Separate a real change in rate from a change in your prompt mix or run count, since adding five new prompts can move an average without anything changing in Copilot. If you're deciding how much a day-to-day swing should worry you, are AI visibility scores reliable is a good companion read. For naming the numbers in a report, the AI visibility metrics and KPIs guide lists the common ones.

Microsoft also gives site owners a second view. In February 2026 it announced AI Performance in Bing Webmaster Tools, in public preview. It covers Microsoft Copilot, AI-generated summaries in Bing, and some partner integrations, and it reports on citation activity:

  • Total Citations: how often your site was shown as a source in AI-generated answers over the period you pick.
  • Average Cited Pages: the average daily number of unique pages from your site shown as sources.
  • Grounding queries: phrases the AI used when it retrieved content for answers. These are grouped or representative phrases, and they are not necessarily the full question a person asked.
  • Page-level citation activity: citation counts for specific URLs.

It also has a timeline, a comparison to an earlier period, and CSV and Excel exports, with ranges such as 7 days, 30 days, 3 months, and a custom range within the available history.

This report is useful for spotting which URLs and which themes Copilot is drawing on, and you can use it to choose prompts worth investigating. It isn't a rank tracker, though. Microsoft says the data is sampled, can be refined as more is processed, and is meant for trend analysis and not exact accounting. A citation count isn't traffic, clicks, engagement, page importance, or rank. Average Cited Pages doesn't tell you where a page appeared within any single answer, and a comparison shows what changed, not why. Don't treat the grounding-query rows as your prompt list either. They're a different thing. That's exactly why your own prompt log is worth keeping alongside it: it can tell you what Copilot said to the questions you chose, while the report tells you how widely your pages are being cited. For tools built around this kind of data, see these Bing and Copilot citation tracking tools.

DeepSmith's AI Visibility overview gives you mention rate, citation rate, share of voice, changes over time, a per-platform breakdown, competitor comparisons, and the sources and pages being cited. Microsoft Copilot tracking is included on Enterprise and Custom plans, and it's not part of the lower tiers. DeepSmith doesn't assign a numeric within-answer rank, so if you use the position method from Step 5, it's a manual score you add on top. The API and MCP access are read-only, which means they read workspace data and can't change it. When the manual sheet gets heavy, AI visibility tools that run prompts on a schedule are one way to track microsoft copilot answers without retyping every prompt.

Done when: you can say, for each prompt group, whether mention rate and citation rate are moving, and you've compared that to the Bing report without calling either one a rank.

Common mistake: claiming a rise in citations means more visits or more trust. A citation means your page was shown as a source, and nothing more.

Step 8: Turn a repeatable finding into a content test

The point of all this is to decide what to do next. Look for a prompt where your brand is repeatedly absent, mentioned but never cited, cited on a page you didn't expect, or consistently below a competitor in an ordered answer. Then open the exact Copilot answers and the cited URLs before you change anything. A pattern across a handful of runs is a reason to look closer, and it isn't proof of why Copilot answered the way it did.

Keep your baseline prompt set as it is. After you change a page, rerun the same prompts with the same mode, settings, and number of repeats, and write the publication or revision date next to the new runs. Be careful about what you claim afterward. Unless you can tell the content change apart from ordinary answer variation and anything else that changed, don't say it caused a lift.

When a gap points to a piece of content you haven't written, that's where DeepSmith's Opportunity Agents come in. They read your AI visibility data and suggest actions like earning a citation for a tracked prompt, turning a mention into a citation, or responding to a competitor's citations, and each idea carries the data point behind it, so you can explain why it's on the backlog. If you want a longer walk through that handoff, turning AI visibility gaps into a content backlog covers it. To get a sense of what's actually involved in earning citations here, how to get cited in Microsoft Copilot is the next read.

Done when: you have one or two prompts with a written hypothesis, a page to change, and a rerun date.

What to do next

Save your prompt set and this week's results as your baseline, book the next check on the calendar, and run the same prompts again in the same way. After three or four weeks you'll have enough to see which prompts are steady and which jump around. If you want to understand why Copilot deserves its own check and isn't just a Bing result, why Copilot needs separate tracking explains the reasoning, and if you'd like to compare tools that do this for you, the best Microsoft Copilot visibility tools roundup is a good place to start.

If you'd like scheduled prompt tracking, captured answers, and separate mention and citation trends without keeping a sheet by hand, you can start a free 7-day trial of DeepSmith and see what your own prompts return. Copilot tracking is part of the Enterprise and Custom plans, so ask about it if you're on a smaller tier.

Frequently asked questions

How do I track Microsoft Copilot visibility?

Create a fixed set of buyer-style prompts, run each one in the same Copilot experience and Search mode, and save each full answer along with its cited URLs. Record mentions and citations as separate fields, then repeat the runs on a consistent schedule and summarize the rates with their run counts.

Can I track a brand mention separately from a citation?

Yes, and you should. Record whether Copilot names your brand and whether it links to one of your pages as two separate fields. A response can do one without the other, and each tells you something different about your content.

How do I track Copilot rankings inside an answer?

Record a position only when Copilot gives an explicitly ordered list. For ordinary prose, record the first-mention location instead and label it that way. Microsoft's publisher analytics don't report your brand's rank or a page's position inside an individual answer, so this is a manual score.

How often should I test the same Copilot prompts?

Pick a fixed schedule and run each prompt several times in separate fresh conversations. Weekly testing with a few runs per prompt is a reasonable starting point, but the cadence and the number of runs are suggestions and not Microsoft standards. Add more runs when your answers vary a lot.