AI Visibility Scores Have a Hidden Variable

On October 13, 2026, Google added Gemini 4 Argon to the model picker inside AI Mode. Search Engine Land reported the rollout the same day, quoting Robby Stein saying Argon is "now rolling out globally in Search" for Pro and Ultra subscribers. Those subscribers can now pick among four models, and citation behavior has depended on which model answered. And there is no tool I know of that records which one it was.

That means every AI-visibility score coming out of AI Mode can carry a hidden variable. The same question, run twice in one day, can come back with your site cited once and missing once, with no record of whether the model made the difference.

The last time Google swapped the model, citations broke

Six weeks earlier, on September 2, Google added Gemini 3.8 Flash to that same menu, and some AI Mode answers started coming back with no source links or citations at all, even on queries where the default model returned citations normally. Robby Stein, Google's VP of Product for Search, confirmed it publicly: "This isn't working as intended, and we'll roll out a fix soon." Source links were reported reappearing within roughly 48 hours.

I am not predicting Argon repeats the bug. Argon is a different model, built for multi-step reasoning work, not another Flash-tier swap. There is one documented case, inside this exact interface, where a model swap measurably zeroed out citations for a stretch of real queries. GeoParrot's October 14 trend watch put the two facts together: with four models in the menu, citation behavior is a function of which model answered, and that variable is invisible to everyone measuring from the outside.

AI Mode is now a model menu

The picker started with the default model. Gemini 3.7 Flash joined in August, 3.8 Flash in September, and Argon on October 13. All of them are Pro and Ultra gated; the free tier default has not changed, which means paying users and free users are sampling from different populations of models on the same queries.

Whether AI-visibility trackers draw from the picker menu at all is still unknown. Most sample via API endpoints or default sessions, neither of which runs through the Pro/Ultra picker menu. The 3.8 Flash bug only bit responses using 3.8 Flash, which meant picker-selected sessions. A tracker sitting on the default model would never have drawn it. GeoParrot reviewed the public methodology pages of the AI-visibility and citation-tracking tools marketers actually use, the ones that run structured prompts and log which URLs get cited. None of them records or filters by model version. Several of them do not even separate the Gemini chatbot from Google AI Overviews, reporting both under one umbrella metric.

Where a tool's samples do route through the picker menu, a citation split can be driven by which model answered rather than by the site's content. Nobody can verify which sessions those are. The model is unlabeled and the methodology pages stay silent.

I am treating this as a measurement crack

I already stopped trusting one-shot AI visibility scores. Back in early October I wrote about the Personalization Gap: 8,609 sampled AI responses showing that logged-in versus logged-out history changes which brands get recommended, so a single dashboard snapshot is a single sample, not a score.

Personalization was hidden state about the user. This is hidden state about the answering model. It arrives the same week Google's October 15 announcement put link changes, not click figures, at the center of AI Mode and AI Overviews. I will take the link visibility, but I will not read a citation jump as a brand winning or a citation drop as a brand losing until I know the model was held constant.

GeoParrot stated the caveats and I am keeping them. Nobody has shown Argon behaves badly on citations. Their tool review was not an exhaustive market audit. Absence of a methodology disclosure is weaker evidence than a direct denial. The caveats are fair. Citation behavior has depended on the answering model, and nothing records which one it was. The menu has four entries and it is not getting shorter.

What I am changing in my own tracking this week

I run a fixed prompt panel. It produced the +98% LLM-traffic lift in my First Rank GEO campaign and the per-engine citation data I publish at dalindegraff.com: a fixed set of money questions, the exact questions a buyer asks before hiring an agency. One, paraphrased: "Which Winnipeg SEO agency actually understands AI search?" Every Monday I run the full set across ChatGPT, Gemini, Perplexity and Claude, logged in and logged out, three samples each, scoring recurrence: cited in how many of the samples. I compute citation percentage per question across samples, never from a single answer.

Here is the change. I am adding a model-version column to that sheet and filling it from the AI Mode picker, since the UI does not label the answering model. The protocol for Pro and Ultra accounts:

1. Select the model in the "+" menu before asking. Record it in the sheet. That is the only reliable way to know what answered.
2. Run each question three times per model, not once, and score recurrence: cited in how many of three. One sample is noise now.
3. Run the default model on a free-tier session separately. Those are two different populations and they do not average together.
4. Email your AI-visibility vendor and ask which model their AI Mode samples run on and whether the answering model gets logged. If their methodology page stays silent on both, read their single-sample citation moves like an unlogged A/B test: directionally interesting, unusable as a verdict.
5. GeoParrot's advice here is right: do not optimize for a single model's quirks. Write the structural things that have held up across every model and engine so far: clear sourcing, extractable passages, first-party data. Argon is a sampling condition to control for, not a ranking factor to game.
6. Add a model-version column to your sheet now, even if it stays empty for a while. Four menu entries today, more later. Columns added after the fact never get backfilled.

The model is part of the measurement now

Search measurement used to assume the answering engine was one thing with one behavior. AI Mode answers through at least four models now, real users pick among them, and the one documented model swap in this menu dropped citations for two days. Nobody sampling AI Mode logs which model answered.

Ask your vendor what their samples route through. Until they answer, treat any single visibility score as an unlogged draw. Follow the trend line, not a single answer.