Reading an AI Citation Spike: 26,000 Citations, Zero Clicks, Eleven Days

One of our pages logged 26,283 AI citations in Bing Webmaster Tools and sent zero visitors. We spent a day trying to explain why the citations did not convert. The explanation turned out to be a different question, and the thing that answered it was a chart row we had not read.

This post is about that chart row, and about the four checks we should have run first. Every figure below names its instrument and its window so you can test it against your own data.

To be clear on terms: this is about AI answer citations, meaning the source references an answer engine displays when it composes a response. It is not about academic or bibliometric citation counts.

The number, and then the shape#

Instrument: Bing Webmaster Tools, AI Performance report (public preview). Window: 19 May to 18 August 2026. Read: 19 August 2026.

• Total Citations: 26,283
• Distinct grounding queries reported: 26
• Referrals from Microsoft properties over the same window: 0

The zero is what sent us looking. A zero-click rate is normal for AI answers. Zero out of 26,283 is not a rate, it is an absence. Even a CTR of 0.05 percent predicts about thirteen visits.

We built several explanations for that absence. All of them assumed we were explaining a process with a rate. Then we read the daily series.

19 May to 25 Jul     0 to 26 per day, median about 2
26 Jul to 6 Aug      0 0 0 0 0 0 0 0 0 0 6 0
7 Aug to 17 Aug      2.3K 2.4K 2.6K 2.7K 2.2K 2.3K 2.7K 2.3K 2.3K 2.4K 2.7K

Eleven days carry essentially the entire total. Eleven multiplied by roughly 2,400 is 26,400, against a reported total of 26,283. Before those eleven days sit twelve consecutive days of zero.

Why flat is the tell#

Look at the variance in that last row. The eleven days range from 2.2K to 2.7K, a spread of under plus or minus ten percent, with no weekend dip anywhere in it.

Human demand does not behave like that. Search volume has weekly seasonality, it has news shocks, it ramps. A metric that sits at zero for twelve days, jumps to 2,300 overnight, then holds a flat line for eleven days is the signature of a switch, not a curve.

There are at least three things that produce that shape, and our data does not yet separate them:

1. A platform reporting change. Something began being counted on 7 August that was not counted before. The twelve preceding zeros are the strongest hint here, because they are zeros rather than low numbers, which is what a pipeline gap looks like.
2. A scheduled automated process. Something started issuing the same small query family daily. That would explain the flat plateau, the absent seasonality, the tiny query set, and the zero referrals in one move, without requiring any audience at all.
3. A genuine change in retrieval behaviour. Bing changed how it grounds this query family and our page became a standing candidate. Possible, but it still does not explain the flatness.

The honest position is that we know the shape and not the cause.

The four checks we should have run first#

Check the instrument before trusting a negative. Our first move was site:citability.dev on Bing to see whether the page was indexed. It returned ten fedex.com tracking URLs. The operator is broken for this domain, so it could not support any conclusion at all. Exact-phrase search worked, and showed the page at position one. If you are about to write "we are not indexed," run a query whose answer you already know first.

Read the vendor's sampling note. Microsoft's February 2026 announcement of the AI Performance report states that the grounding-query data represents a sample of overall citation activity. We had read the 26 reported queries as a complete census and built an argument on it. An export is a sample until the documentation says otherwise. Worth separating from truncation: we queried the underlying report endpoint at a page size of 500 and it returned a total record count of 26, so 26 is everything Bing exposes, and Bing calls what it exposes a sample.

Separate retrieval from exposure. A citation records that a system referenced the page. It does not record that a person saw the reference, that the reference was prominent, or that the answer left the reader with any reason to click. Treat "was it retrieved" and "did a human see it" as two different columns.

The part that surprised us#

Bing's report also labels each grounding query with an intent and a topic cluster. By intent, about 18 percent of the citations are commercial or comparison, which is the only slice a business should care about. The topic labels are where it gets strange.

Eighty-four percent of these citations are filed under Research Tools and Databases. Three tenths of one percent under Search Engines and SEO. We build an AI answer visibility product. Bing has categorised this activity as academic bibliometrics.

The tempting story is that "citation rate" carries a large, old, academic meaning, that the retrieval process resolved it to that meaning, and that the page is therefore being shown to people who want Scopus and got us. That would explain the zero clicks without any measurement anomaly at all.

We are not making that claim. Microsoft documents Topics as a way to group grounding queries into thematic clusters for analysis, and warns that the labels can be broad for niche domains during preview. So the 84 percent may describe real retrieval routing, or it may describe a classifier mislabelling a young category. Those are different worlds and this data does not choose between them.

What we are not saying#

We are not saying the page earned 26,000 citations. We are saying Bing recorded roughly 26,000 citations for this domain inside an eleven-day block beginning 7 August, at a flat daily rate, after twelve days of zero. Every clause in that sentence is checkable. The causal sentence is not available yet.

We are also not treating raw citation volume as a goal any more. A page with 26,000 citations, zero clicks and an 84 percent wrong-topic classification is worse for a business than a page with 2,000 citations in the right cluster and a hundred qualified visits. Raw volume ranks the first page higher. A useful measurement product should not.

What to do with your own spike#

If a dashboard hands you a large AI citation number, five steps in order:

1. Pull the daily series. Before anything else. If it is flat, you are looking at a process, not an audience.
2. Check for a zero gap immediately before it. History does not usually go silent right before real demand arrives.
3. Find the vendor's sampling language and decide whether you are holding a census or a sample.
4. Run a control query on any tool you are about to use to prove a negative.
5. Check whether the citations are commercially adjacent at all. Intent and topic labels are free in Bing's report and they answer this faster than any analysis you can run yourself.

Then, and only then, verify by hand. Run your top grounding queries in the actual answer engine and record whether your page appears, where it appears, what the answer is about, and who else is cited. That is the step that tells you whether there is a human-facing surface behind the number.

The open question#

We are rechecking the daily chart on 21 August. A plateau still running at about 2,400 a day points to a standing process. An abrupt stop points to a finite window. Any backfill of the twelve zero days would be close to proof of a reporting artifact, because history does not change when the cause is demand.

We will update this post in place with whichever it turns out to be, including if it makes us wrong again. We have already retracted twice in the course of this analysis. The record of what we got wrong is more useful to you than a cleaner story would have been.