A fall in AI citations is not automatically a fall in your website's quality. Before rewriting pages, check whether the search experience changed, whether the observation was valid and whether you are comparing the same conditions.
The September 2026 Gemini 3.8 citation incident is a useful example of why an AI visibility chart needs context. This guide sets out our recommended method for separating a possible website problem from a measurement or platform change.
Reviewed 8 September 2026.
What happened with Gemini 3.8?
Google introduced Gemini 3.8 Flash on 2 September 2026. In subsequent AI Mode tests, Gagan Ghotra documented examples with missing citations in his first-hand account.
Ghotra also reproduces Google's Robby Stein acknowledging the unintended behaviour and reports seeing links return on 4 September in the updated first-hand account.
These were observations of particular responses, not a representative measurement of every query, country or account. They demonstrate a potential failure mode; they do not establish a universal citation-loss percentage.
Define what your tracker is observing
Our recommendation is to record the provider, product, prompt, language, location where known, date, account context and selected model when exposed. Store the original answer and cited URLs as well as any summary score.
Use "unknown" when a model version or location is not available. Do not turn a guessed backend version into a precise-looking data field. A consumer application, its default experience and a selectable model are not interchangeable experimental conditions.
Google's own AI features guidance notes that different models and techniques can produce different responses and links. That supports keeping observations separated rather than assuming every Google-branded answer is the same test.
Use explicit denominators
Define citation frequency before calculating it. One possible definition is the number of valid completed responses citing your domain divided by the number of valid completed responses in the chosen prompt group.
For example, six citations across ten valid responses would be a 60% observed citation frequency for that test set. It is not a 60% share of all customers' answers. Label the scope directly beside the number.
Failed requests deserve their own status. A timeout is not evidence that an engine deliberately omitted your business. Conversely, a successful answer with no citations should not disappear from the dataset just because it is inconvenient. Record the observed behaviour and explain any analysis exclusions.
For your operational dashboard, we suggest retaining both the raw series and an annotated comparable series. That prevents data cleaning from quietly becoming result selection.
Repeat the same useful questions
Keep a stable core of buyer questions and repeat them under documented conditions. Add exploratory prompts in a separate group so that a changing prompt mix does not masquerade as a performance trend.
Our suggested review distinguishes branded recall, unbranded discovery and comparison questions. It also separates a simple name mention from a linked citation and a positive recommendation.
Choose a repetition schedule proportionate to your budget and decision. A small test can identify something worth investigating, but should not be presented as a population estimate. More repetitions do not automatically remove biases in how the prompts were selected.
The point is not to manufacture certainty. It is to make the next observation interpretable.
Investigate a drop before editing content
Start by reviewing the saved answers. Did all citations disappear, or only your domain? Did comparison businesses change too? Did the provider begin answering the question in a different way? Was the selected model different from the previous run?
Next, check whether your own site changed. Relevant investigations might include access settings, the target page, redirects, content revisions or a changed commercial offer. These are hypotheses to examine, not automatic explanations.
If a provider-wide behaviour changed, annotate it and establish a new baseline where necessary. When two experiences can still be tested side by side, an overlap period can help you describe the difference. It cannot guarantee that the experiences are identical for all users.
An example of a better alert
Consider this illustrative alert: "Citation frequency fell from 7 of 10 to 3 of 10 valid responses. The selected model changed between periods, and competitors also lost links. Investigate provider behaviour before changing the page."
That is more actionable than "Your AI SEO score dropped 40 points". It identifies the observed change, explains an alternative cause and states what to check next.
A second alert might instead say: "The same prompt group and selected experience were used. Your page now redirects to a different offer. Review whether the new destination still answers the buyer question."
Both are recommendations based on a stated observation, not invented diagnoses.
Connect monitoring to work that can be checked
Use UpSearch's AI Mention Tracker for question-based monitoring, then keep an investigation record for any meaningful change. Record the observation, candidate explanation, chosen action and later review.
For actual Google exposure, add the separate evidence described in our Search Console Generative AI report guide. For traffic and outcomes, use the distinction in our AI attribution guide.
Good monitoring should reduce unnecessary work, not create an urgent task every time an assistant gives a different answer.
FAQ
Does one missing citation prove that my SEO got worse?
No. Inspect the answer, repeat comparable tests and consider platform changes before reaching that conclusion. One observation is a reason to investigate, not proof of the cause.
Should failed AI requests count as zero citations?
We recommend recording them separately from completed answers. Otherwise a provider outage or collection problem could look like a visibility loss.
What should I do when the model version is unavailable?
Record it as unknown. Keep the product, selected mode where visible, timestamp and original response so that the evidence remains useful without pretending to know hidden system details.