I Measured My Copilot Citation Drop Against Its Best Five Weeks

On September 27 I exported Bing Webmaster Tools’ AI Performance report for my own site and wrote one line in my audit notes: Copilot citations were “decaying ~40% fortnight over fortnight.” The window behind that line was a day off. Fixed, the two fortnights averaged 32.2 and 17.7 citations a day, a 45% fall.

The window was the small mistake. The comparison was the big one: the fortnight I measured against sat inside the busiest five weeks in the report’s 90-day view.

Quick Take

I called a 45% fall in my Copilot citations. Against June and July it was about 16%, and it didn’t start where I thought. The first week of my “decay” fortnight sat right at the early-summer level. The second week was well under it, and a later pull says the fall carried into October. So there’s a real decline in here, smaller and later than my note said. Before I call a citation drop now, I find the last stretch that wasn’t a high, deal with the report’s zero days, and write the comparison down before I look.

What the AI Performance report counts

Microsoft put AI Performance into Bing Webmaster Tools as a public preview in February 2026. It counts how often your pages appear as sources in Microsoft Copilot, Bing’s AI summaries and some partner surfaces. You get total citations, the average number of your pages cited per day, the pages cited, and the grounding queries, which are the searches the AI ran behind the scenes to find sources. It has no click or traffic figures.

The tables are a sample, and the report says so on the page. My 30-day per-page table came back “No results found” while the 30-day total read 699, so I treat every figure in it as a floor.

And it has zero days I can’t explain. My 90-day series reads exactly zero on July 27, August 1 through 6, and August 22 and 23. Nothing on my site was down on those days. I cannot tell whether Copilot stopped citing me or the report stopped counting, and the export doesn’t say. In a 14-day window, two zero days cut the average by a seventh on their own.

Where my comparison went wrong

Three windows from the 90-day series:

Window	Days	Citations per day
June 28 to July 26	29	21.1
August 7 to September 11, excluding two zero days	34	37.8
September 12 to 25	14	17.7

The middle row is the outlier. It holds the report’s nine highest days, including two at 74. It began on August 7, four days after I shipped a new cluster of answer pages, one of which became my second most-cited page, and it peaked during a round of indexing and internal-link fixes on August 11 to 13. That launch landed in the middle of a six-day run of zeros, so I can’t tie the high to it, or to the fixes, or to anything else I did. My note compared the bottom row against the tail of that run, so what it measured was the end of a high.

I picked the June and July window after seeing the data too, so I checked whether the answer depends on it. It does, a little. Against late June and early July (17.5 a day) the fortnight is flat. Against the second half of July (24.9) it’s 29% down. Split the fortnight in half and the picture gets clearer: September 12 to 18 averaged 21.7, right at normal, and September 19 to 25 averaged 13.7, which is 22 to 45% under every June or July window I tried. My note got the size of the fall wrong and the start of it wrong. It didn’t get the direction wrong.

I also tried the obvious control, Bing’s ordinary search numbers, to see whether the whole site was losing visibility or only the AI layer. My Search Performance report showed 396 impressions and 6 clicks over roughly the same three months. At about four impressions a day, it’s too thin to tell me anything either way. My own working notes leaned on that same series to rule out an indexing problem, and I’m withdrawing that conclusion here.

Before I call a citation drop

I check in this order now, and I’ll use it on the next client report that shows a fall. If someone hands you an AI visibility report with a two-week change in it, these are the questions to ask them.

1. Find the last window that wasn’t a high. Plot the full daily series and pick a stretch with no spike in it, at least four weeks long. Then try a second reasonable window and see whether the answer changes.
2. Mark the zero days and leave them out of both sides. A run of zeros can be a reporting gap, and an average that includes them drops for reasons that have nothing to do with the site.
3. Open every cited URL and confirm it still resolves. I did this on October 19 for the 40 pages the report had cited. 33 still build as pages, and the other 7 redirect to live pages that absorbed them. Two of the 33 are blog posts I noindexed on October 11, which held 16 of the 1,515 sampled page citations between them. That’s mine to own, but it isn’t the fall.
4. Write the comparison down before the next pull. Which windows, which metric, and what size of change counts as real. I write that down before I open the next export.

October, and the edits I stacked on it

A later pull, taken October 18, put my October daily rate below the June and July baseline. That fits the slide that started around September 19. I have not re-measured it for this post, so I’m not printing the figure, and I don’t know the cause.

I also made it harder to find out. Between October 3 and October 19 I changed the four pages Copilot cited most 16 times, across 9 commits. Those four held 964 of the 1,515 sampled page citations, about 64%. Some of those edits were fixes Copilot’s own grounding queries pointed at, and one put my name into the paragraph each page leads with. The last was a one-line price update. With that many edits stacked on four pages, I won’t be able to point at the one that moved the number, whichever way it moves.

So apart from price changes, I have stopped editing those four pages until the 30-day re-test of my AI search audit on October 28. Changing the thing you’re measuring while you measure it is the same mistake as picking the baseline after you’ve seen the number.

What I did not claim

• That the decline isn’t real. It is. It started later and ran smaller than my note said, and the October pull says it kept going.
• A cause for the August high or the fall after it. Both overlap with work on my site, and the zero days sit in the way of tying either to a change.
• That the counts are complete. Every figure here comes from a report whose tables Microsoft labels a sample, and some of its zero days may be missing data rather than real zeros.
• Anything about other sites. This is one site’s report. I’d run the same checks on a client’s, but I would not expect these numbers.

Keep going

If this hit, the next two pieces in the same universe:

• I ran six AI search audits. Crawler access was almost never the problem. The buyer-question side of the same measurement, and why a citation count isn’t the same as getting named.
• Migrating a site AI assistants already cite. What to protect before a rebuild touches the pages engines quote.

Citation counts measure how often an engine leaned on your pages. Whether it named you when a buyer asked who to hire is a separate number. If you run a practice or a law firm, the free AI answer snapshot is a first look at that one: ten real buyer questions put to three engines, with a ranked fix list. It’s a single run, so treat it as a starting read rather than a trend. The full audit in paid work locks its question panel before the first run, and every re-test reuses the same file.