Copilot Citations Dropped in Late July. Here Is What Actually Changed.

Copilot citations dropped sharply for many sites in late July 2026 because Microsoft changed which models sit underneath Copilot, not because those sites did anything wrong. The change was published in a changelog and a tenant message, never announced. Sites that had been cited steadily for months stopped being cited within a day, with no warning and nothing to fix in the usual sense.

We watched it happen on our own properties as it happened, because we monitor citation data directly rather than sampling prompts. This is what the event looked like from the inside.

Why Copilot Citations Changed

Copilot citations changed because the models grounding Copilot's answers changed. Beginning 24 July 2026, OpenAI-operated models in the GPT-5.6 family became active by default inside Microsoft 365 Copilot tenants unless an administrator went in and changed the setting. On 29 July, Microsoft confirmed that generally available models for Copilot Business and Enterprise would switch to on by default, governed by a single organisation-wide opt-out.

Read that second part slowly, because it matters more than the first. The default is now on. New models arrive in the tenant automatically. Nobody has to approve anything, and nobody sends you a note about it.

What It Looked Like From the Inside

What it looked like was not a decline. It was a cliff.

Pages that had been cited consistently for months went nearly silent inside a single day. No content had changed on them. No rankings had moved. Nothing had been republished, redirected or edited. The same pages that were being pulled into answers on one day were not being pulled the next.

The detail that ruled out coincidence was that it happened on more than one domain at once, in unrelated markets, on separate infrastructure, on the same date. A single site falling over has a dozen possible explanations. Several unrelated sites falling over together on the same day has one.

Why Almost Nobody Noticed

Almost nobody noticed because the standard way of measuring AI visibility cannot see an event like this.

Prompt tracking works by choosing a question, asking an assistant that question on a schedule, and recording whether your brand appears in the answer. It samples the output. When the number moves, it cannot tell you whether the model changed, the grounding changed, a competitor published something, or the same prompt simply produced a different answer that morning. It reports a score with no mechanism behind it.

Citation data is the other side of the transaction. It records which of your pages were actually pulled into answers and when. When that line drops on the same day across separate properties, you are not looking at a score moving. You are looking at an event, with a date attached.

What We Did About It

We treated the date as the finding and worked backwards from it.

The first step was going to Copilot directly and probing how its behaviour had changed, comparing what it returned against what it had been returning before. That told us what the new configuration was actually rewarding. From there we made a specific set of changes to the page that had been carrying the citations, and left the rest of the property alone.

The citations came back.

We deliberately did not apply the same changes to a second property that had dropped on the same date. That one is still flat. Two sites, same event, one treated and one untouched, with the outcome you would expect from each. That is as close to a controlled test as anybody gets outside a lab.

We are not going to publish the specific changes. What matters here is not the recipe, it is that the problem was diagnosable at all, on the day, from first-party data, by someone watching the right side of the transaction.

The Public Record Caught Up Afterwards

The confirmation arrived later. Microsoft's own changelog and the tenant message centre entry describe exactly the rollout window we had already identified from the data. We did not read about the change and then go looking for its effects. We saw the effects, diagnosed the cause, fixed the outcome, and then watched the paperwork appear and agree with us.

That order matters. Anything can be explained after the fact. Being early is the only thing that proves you were reading the right instrument.

Why Copilot Citations Will Keep Moving

What this means is that citation volatility is now structural rather than occasional.

The model underneath an assistant can be swapped without notice. That is not a criticism of Microsoft, it is their stated policy, published in their own changelog: generally available models turn on by default so administrators do not have to manage them one at a time. It is a reasonable decision for IT teams and an uncomfortable one for anyone whose visibility depends on how a given model grounds its answers.

Three things follow from it.

Your citations can fall for reasons entirely outside your site. Nothing on the page has to change for the outcome to change, which means the usual diagnostic instincts, checking rankings, checking crawl errors, checking what a competitor published, will all come back clean while the problem is real.

You will not be told. There is no notification for this. The information exists in changelogs and tenant messages that nobody outside IT reads, and it surfaces in the marketing world weeks later, if at all.

And a monthly report cannot catch it. By the time a monthly cycle closes, the event is three weeks old, the cause is much harder to isolate, and the recovery window has been sitting open the entire time.

The Practical Version

If your AI citations dropped in the last few weeks and you have been auditing your own site looking for the mistake, you may not have made one. Check the date it happened. If it lands in that window, and particularly if it happened across more than one property you own, you are looking at this rather than at anything you did.

Finding that out requires watching citations rather than sampling prompts, which is the entire distinction between measuring an outcome and measuring a rumour of an outcome.

If you want to know what your own citation data has been doing, call 561-581-3232.