There is a strange habit, in our industry, of ignoring Microsoft. People look at the search-engine market share, see Bing at single digits, and conclude that whatever happens inside Microsoft's index is a rounding error.
This was always a slightly lazy way to think about it, but in 2026 it has become genuinely misleading. Because the index that powers Bing also powers, in whole or in significant part, ChatGPT Search, DuckDuckGo, Yahoo, Ecosia, and Brave. When you treat all of those surfaces as one ecosystem — which is, in the end, what they are from a citation point of view — the Microsoft index suddenly accounts for something far larger than its standalone market share would suggest. Some analysts now place the combined downstream reach at twenty percent or more of all AI-mediated search interactions, though I would urge anyone reading a number like that to remember it is an estimate and to verify it against their own analytics before banking on it.
Either way. The point is that Bing Copilot citations matter, and almost nobody is tracking them properly. The dashboards do not exist yet, or if they exist they are expensive and unreliable. So we have to do it the old way — by hand, with discipline, on a rhythm that gives us something useful to compare against.
Let me walk you through the method I have been using with the small businesses we work with. It takes about ninety minutes the first time and forty-five minutes every quarter thereafter, and it produces a number you can defend.
The premise behind the method
Before we get into the mechanics, I want to be honest about one thing. There is no single objective measurement of "Copilot citation share". The number does not exist as a public metric. Microsoft does not publish it. No third-party API exposes it in a clean form.
What we can measure is something more modest but, frankly, more useful — the share of citations for the questions we care about that include our domain. If we pick the right questions, and we ask them consistently, and we record what we find, we get a directional metric that moves up and down for understandable reasons. That is enough. We are not trying to publish a paper. We are trying to make decisions.
The whole method rests, therefore, on the question list. Choose the wrong questions and the number will move randomly. Choose the right ones and it will move in lockstep with the business outcomes you actually care about.
Building the question list
The list should contain between fifteen and thirty questions, no more. Fewer than fifteen and you will not have enough signal to see trends. More than thirty and the manual work becomes a quarterly burden you will stop doing after the second pass.
Three kinds of questions belong on the list, and the mix matters.
The first kind is commercial intent. Questions a customer asks when they are close to spending money. For a plumber in Pordenone, this would be things like who is the best emergency plumber in Pordenone, or plumbers open on Sunday near me. These are the queries where a citation is most likely to translate into a phone call within forty-eight hours. They should make up roughly half the list.
The second kind is evaluative intent. Questions that come earlier in the decision cycle, where the customer is weighing options. Is it better to repair an old boiler or replace it? — that kind of thing. These citations do not produce immediate revenue but they shape which businesses get considered. A quarter of the list.
The third kind is category-defining intent. Broader questions where being cited establishes you as an authority in a field. How does underfloor heating actually work? The remaining quarter of the list.
Write the questions in the natural language a customer would use, not in the keyword shorthand an SEO would use. The reasoning models that sit behind Copilot — yes, Copilot is now Gemini-flavoured in its prompting style even when it runs on GPT-class models underneath — respond to natural questions. Type plumber Pordenone Sunday and you get one set of citations. Type are any plumbers in Pordenone open on Sunday and you get a different, and usually richer, set.
The recording template
Open a spreadsheet. Create columns for the date, the question, the position of your citation in the source list (or "not cited"), the position of your nearest local competitor, and a short note on the model's answer quality.
That last column matters more than people realise. A citation in an answer that says Bottega Rossi is highly recommended for emergency plumbing is worth roughly twenty times a citation in an answer that says several plumbers operate in the area, including X, Y, and Bottega Rossi. The first is a recommendation. The second is a list. The position number alone does not capture this.
I score the answer quality on a simple three-point scale. Three points if you are the named recommendation. Two if you appear in a short, qualified list of two or three names. One if you appear in a longer list or in an undifferentiated mention. Zero if you are not cited at all. This gives you a weighted citation share you can roll up at the end.
Running the queries
Here is where people most often go wrong. They run the queries from their normal browser, on their normal laptop, while signed into all their normal accounts. The personalisation layer in Copilot is more aggressive than people assume. You will see citations the rest of the world is not seeing, and vice versa.
The fix is mundane but important. Use a clean profile — a separate browser profile in Chrome or Edge with no signed-in account, no extensions, and ideally a location set to the city your business actually serves rather than wherever you happen to be sitting. Run the queries from this profile, in a single session, on a single day. The session matters because Copilot will sometimes adjust later answers based on earlier ones.
The truth is, even with all this hygiene, you are still measuring an estimate. But it is a consistent estimate, and consistency is what makes the quarterly comparison meaningful.
What the first pass tells you
When you do this for the first time, you will see one of three patterns, more or less.
The first pattern is narrow citation. You are cited for a small number of queries — usually the ones where you have built strong topical authority over many years — and absent from everything else. This is the most common starting point and, frankly, the easiest one to improve. The fixes lie in publishing genuinely useful pages on the evaluative and category-defining questions you are not yet appearing in.
The second pattern is brand citation only. You appear for queries that contain your own brand name and almost nowhere else. This means Copilot knows you exist but does not yet associate you with any category. The fix is harder. You need third-party mentions — coverage, citations from trusted local sites, structured presence on platforms like Reddit and Wikipedia — that bind your brand to the categories you serve.
The third pattern, which is rare but lovely when it appears, is broad shallow citation. You appear in many queries but always in long lists, never as the named recommendation. This means you have built reasonable authority but not differentiation. The fix is positioning: pick two or three things you are genuinely the best at, and write deeper, more confident content on those topics.
The rhythm of the quarterly review
Every three months, repeat the exercise. Same questions, same clean profile, same scoring system. Do not be tempted to expand the question list mid-stream — the comparability is the whole point. If you must add questions, do so at the start of a new quarter, and keep the original set running alongside the new ones for at least two cycles so you have a baseline.
The number you compute — total weighted citation points divided by maximum possible points — is your Copilot citation share. It is not a number anyone else can verify, and it is not perfectly precise. But it moves. And when it moves, you will almost always be able to point at something specific that caused the move. A new page you published. A press mention you earned. A competitor who started taking AI search seriously. A core update at Microsoft that reshuffled the index.
This is, in the end, what a useful metric does. It does not have to be exact. It has to be responsive to your actions. The Copilot citation share, tracked this way, is.
A small note on automation
People ask, naturally, whether all of this can be automated. The answer, today, is mostly no — Copilot's terms of service prohibit programmatic querying, and even if you accept that risk, the answers vary enough that any automated capture would need extensive validation. I would not recommend it.
What you can automate, and should, is the prompt to do the manual run. Put a recurring event in your calendar for the first Monday of every quarter. Block ninety minutes. Open the spreadsheet. Run the queries. Update the column. Compute the share. Note one observation.
The discipline of doing it consistently is worth more than any tool that promises to do it for you.
The honest closing thought
I will say one last thing, because it would be dishonest not to. The Copilot citation share you track this way is a leading indicator. It tells you about the opportunity surface for Microsoft-index AI traffic. It does not tell you whether anyone actually clicked through to your site, or called, or bought something.
For the actual outcomes, you need the boring complement — your GA4 referral data, your phone-call tracking, your booking records. The Copilot citation share is where you look first. The conversions are where you look second. Both are necessary. Neither is sufficient on its own.
And that, in the end, is the honest truth about every AI-search metric we have available in 2026. They are partial. They are manual. They are imperfect. But they are also — and this matters more than people think — the only tools we have right now that can answer the question your boss is going to ask, sooner or later, about why the phone is or is not ringing.
Quarter by quarter, the numbers will tell you. Just commit to looking.