Auditing the Instrument: What Bing's Report Does Not Show

Before I advise a client on AI-answer visibility, I establish how much of the report can be trusted. This is that audit, run on my own property so the raw instrument can be published alongside the result.

Instrument: Bing Webmaster Tools, AI Performance, property citability.dev, citation sources "Microsoft Copilots and Partners". Three-month period beginning 19 May 2026, read 19 August 2026. "Citation" means an answer engine linking a source inside a generated answer, not an academic citation.

Four things the report will not tell you unless you check

Finding 1: The three-month total is eleven days.

The headline reads 27.3K citations over three months. The page carrying most of it has data on 7 August to 17 August 2026 and nowhere else, at 2.0K to 2.5K per day. Eleven flat days after a long stretch of near-zero. Quoting the three-month figure as a trend describes growth that did not happen.

Finding 2: The query list is truncated, not filtered.

The top-cited page reports 21 grounding queries. The site reports 26. I checked whether the page list contains anything the site list omits, because that would mean the site view is filtered. It does not. The page list is a strict subset.

But the listed queries do not account for the total. On a page small enough that every count is an exact integer, and rounding therefore cannot hide anything, the report shows 122 citations and two queries totalling 120. Two citations attach to no listed query.

So the list is honest at the top and incomplete at the bottom. Low-count queries are dropped below a display threshold. Site-wide that unlisted tail is 680 to 1,373 citations, roughly 2.5 to 5 percent.

Finding 3: The last row hides behind a page break.

Tables default to 25 rows. The site had 26 grounding queries. The 26th is only reachable through the next-page control, and the footer that says so is below the fold.

Finding 4: Bing says twice, on screen, that this is a sample.

"The data shown below represents a sample of overall activity." "Intent and topic labels are AI generated and can filter potentially unsafe queries." A third notice appears once you filter: "Chart totals and table results may differ when filters are applied."

An export is a sample until the vendor's own documentation says census. This one says sample.

Why this is the first thing I do

Each of these four turns a confident number into a wrong recommendation. A step read as a trend produces a content plan built on a spike. A truncated list read as complete produces a keyword strategy missing its tail. A sample read as a census produces a forecast with no error bars.

None of this makes the report useless. It makes it a floor. Floors are worth a great deal, provided nobody sells them as ceilings.

Method: every figure above names its instrument, its window, and the date it was read. Every count came from the live report rather than an export. Anyone with access to the same property can reproduce all of it.

Control page for the exact-integer check: https://citability.dev/blog/what-is-ai-citation-rate