How to Tell Whether an AI Mode Model Change Moved Your Citations

本文目前只有英文版。站点其余部分已汉化，这篇内容尚未翻译。

Short answer: you can't tell from a before-and-after pair of readings. To attribute a change in your Google AI Mode citations to a model change, you need a fixed set of questions asked repeatedly before and after the change date, a measure of how much your citations vary when nothing has changed, a record of your own page edits, and a comparison surface that the model change shouldn't touch. If the shift is bigger than normal variation, lines up with the rollout date, and doesn't show up in the comparison, a model change is a reasonable explanation. Short of that, it's a guess.

We'll say it plainly: we don't yet have enough readings to say whether any recent Gemini model change moved AI Mode citations. This post is about how to find out, not about what happened.

Why this question keeps coming up

Google's model updates are arriving fast. Its Gemini 3.7 Flash announcement is dated 2026-08-13. Its Gemini 3.8 Flash announcement, dated 2026-09-02, calls 3.8 the third Flash release in six weeks and says it's available to Google AI Pro and Ultra subscribers across the Gemini app, AI Mode in Google Search and Gemini in Google Sheets. (Both read 2026-09-24.)

Two details from those sources shape the measurement problem:

• The 3.7 Flash announcement doesn't mention AI Mode at all. Its arrival in AI Mode was reported separately. Search Engine Journal's report (read 2026-09-24) attributes it to a post on X by Google's VP of Product for Search and describes it as a model-menu choice for AI Pro and Ultra subscribers, in English. The same report notes Google didn't say whether 3.7 Flash became AI Mode's default or joined its automatic model routing.
• Availability to subscribers is not the same as the default for everyone. A model you can choose in a menu may never answer a logged-out user's question. Most people asking AI Mode about your category won't have picked a model at all.

So "Google put a new model in AI Mode" might change what most people see, or it might change almost nothing for them. The announcements don't say which.

What you can and can't observe

It helps to be precise about what a measurement records.

You can observe: for a given question, asked at a given time from a given location, whether the AI Mode answer cites your domain, which URLs it cites, and in what order.

You usually can't observe: which model wrote that answer. The AI Mode answers we collect through our data provider don't name a model, and a logged-out browser doesn't show one either. You also can't see the retrieval step: which pages were considered and left out.

That gap is the whole difficulty. You're trying to attribute a change to a cause you can't see directly, so you have to rule out the causes you can see.

A method in six steps

1. Fix your questions before the change, and keep them fixed

Pick the questions that matter to you, the ones a buyer in your category would ask, and write them down with the exact wording. Don't edit them afterwards. A reworded question is a new question, and comparing it with the old one measures your edit, not Google's. Our prompt tracking guide covers choosing them.

Fix the location too. AI Mode answers can differ by location, so a comparison across two different locations isn't a comparison of two dates.

2. Measure your normal variation

Ask each question several times across a couple of weeks before the date you're investigating. Count how often you're cited, per question. That spread is your noise floor. AI answers vary between runs even when nothing has changed, and the only way to know how much is to measure it.

If a question cites you in, say, half of its runs before the change, then one "not cited" after the change tells you nothing.

3. Log your own changes

Keep a dated list of every edit to the pages you expect to be cited: new sections, rewrites, removed pages, redirects, schema changes. Do the same, where you can see it, for the pages cited in your place. If you rewrote your comparison page the same week a model shipped, you have two candidate causes and no way to separate them after the fact.

4. Pick a comparison surface

Google runs more than one AI surface. AI Overviews, the summary block on a normal results page, is a separate product from AI Mode, and we track the two as separate series for exactly this reason. Run the same fixed questions on AI Overviews alongside AI Mode.

The logic is a simple difference-in-differences:

• If AI Mode moved and AI Overviews didn't, something specific to AI Mode changed. A model change is one candidate.
• If both moved together, the cause is more likely something they share, such as Google's index, a ranking update, or a change to your own pages, rather than one AI Mode model.
• If neither moved beyond normal variation, there's nothing to explain.

This isn't a perfect control, since the two surfaces don't share everything. But it's the cheapest way to separate "AI Mode changed" from "Google changed."

5. Look at source lists, not just the yes/no

A citation verdict is one bit per answer. The source list carries much more. For each question, compare the set of cited domains before and after:

• Did the whole list change, or just your slot?
• Did a particular kind of page gain ground (forum threads, documentation, comparison pages)?
• Are the new sources ones that also moved in AI Overviews?

A model change that shifts what kind of source AI Mode prefers often shows up in the lists before it shows up clearly in your own yes/no count.

6. Line the dates up last

Only after the steps above should you check whether the shift starts on the rollout date. A date match is necessary but weak evidence on its own: with updates shipping this often, some model change will be near almost any date you pick.

Write down what you'd conclude before you look. For example: "if the citation rate for these questions falls by more than the pre-period range and AI Overviews doesn't, I'll treat it as an AI Mode change." Deciding afterwards makes it too easy to find a story in noise.

What this looks like with our own data

Our paid sweeps record Google AI Mode as its own series, separate from AI Overviews and from Gemini, and each reading carries its date and, where one was set, its region. That's the raw material the method needs. What we don't have yet is enough readings on each side of the recent Flash changes, across enough sites, to say whether they moved anything. When we do, we'll publish the numbers with their dates. We won't infer a result from the rollout calendar.

If you want a free starting point, our free AI visibility checker gives a dated per-engine reading of whether AI engines cite your domain. It doesn't cover AI Mode. For how we record and compare readings across engines, see how TriRank measures AI citations.

FAQ

Which Gemini model does Google AI Mode use?

It depends on who's asking and what they choose. Google's 2026-09-02 post says Gemini 3.8 Flash is available to Google AI Pro and Ultra subscribers in AI Mode. Coverage of the earlier 3.7 Flash rollout describes it as a choice in AI Mode's model menu for those subscribers, and says Google didn't state whether it became the default or joined automatic routing. The AI Mode answers we collect don't name the model that wrote them.

My AI Mode citations dropped the week a new model shipped. Was it the model?

Maybe, but the date alone can't tell you. Rule out the other causes first: ordinary run-to-run variation, changes to your own pages, changes to the pages cited instead of you, and a shift across all of Google's AI surfaces. If the drop is larger than your normal variation, starts on the rollout date, and doesn't appear in a comparison surface like AI Overviews, a model change becomes a reasonable explanation. It's still not proof.

How many readings do I need before I can say anything?

Enough to know your normal variation before the change. As a working minimum, run the same fixed questions several times over at least a couple of weeks before the date you care about, and the same number after. One reading on each side can only tell you that two samples differed, which AI answers do even when nothing changed.