Kingo Friver

Continue: How to Insist a Pencil

Yep — all five goblins arrived intact, and the filename progression is clear. I have:

AskAiMode_IfItsTransitive01.docx

AskAiMode_UserPersonalization01.docx

AskAiMode_UserPersonalization_AiOverview01.docx

AskAiMode_UserPersonalization_AiOverview_Incognito01.docx

AskAiMode_UserPersonalization_Across_AiOverview_AiMode01.docx

And yes, I can follow the thought process very clearly.

The first file plants the fingerprint. In AI Mode you ask the totally ordinary insist intransitive or transitive, and it gives that unusually elaborate answer: the same headings, the peculiar definition of “standing alone,” the same examples, the whole that-clause caveat, and the invitation about “insist on” versus “insist that.” Then the conversation goes further into the glorious insist a pencil rabbit hole, and AI Mode explicitly tells you that it saw nothing unusual in your original prompt that would explain why it produced such an elaborate response.

Then comes the rather spectacular bit. Later, ordinary Google Search’s AI Overview produces the same answer. I compared the actual prose in those two documents: after removing the different citation markers/source cards, the first AI Mode answer and the later AI Overview answer are character-for-character identical — not merely “similar.” Same wording, examples, parenthetical caveat, paragraph order, and even the conversational If you’d like… ending.

That makes your final document wonderfully awkward for Google Search itself, because when you ask directly about cross-feature history, its AI Overview tells you that standard AI Overviews are “primarily stateless,” that they do not read past AI Mode transcripts, and that they treat every search as a fresh prompt.

So yes: I absolutely see what you caught.

I would make one small forensic distinction before sentencing the raccoon. Your evidence very strongly demonstrates that AI Overview and AI Mode can reuse/share something across the supposedly separate surfaces. It does not yet uniquely prove that the mechanism is “AI Overview read my stored conversation history.” Other possibilities include a shared generated-answer cache, account/session-level response reuse, or some common backend state. And “AI Mode and AI Overview are different features” can technically remain true even if the two creatures share half their internal organs.

The biggest missing control is amusingly tiny: your Incognito document repeats the personalization question, not the original insist intransitive or transitive experiment. If Incognito/new-account AI Overview gives a substantially different insist answer while your signed-in AI Overview reproduces those 1,110 characters exactly, the cage door gets much harder for the Google gremlin to squeeze through.

But as evidence of “these things are less independent/stateless than their own explanation suggests”? Oh yes. Your five-file sequence tells that story beautifully.

I can also keep an eye on Google’s own Search Help/blog wording about AI Overview/AI Mode history and personalization as it changes.

Yeah — and the important thing is that Google’s own current documentation now makes part of your “continuum” much less speculative than it would have sounded a year ago.

Google explicitly says AI Mode can personalize answers using previous searches and other activity in Search Services History, and that saved history lets you resume previous AI Mode conversations. (Google Help) It also documents the AI Overview → AI Mode transition as one continuous interaction: if an AI Overview is followed by “Dive deeper in AI Mode,” Google may record two history entries, one for the original query and another for the transition into AI Mode. (Google Help) So your UX chain

ordinary Search → AI Overview → follow-up → AI Mode → persistent/history-aware AI Mode

isn’t merely you anthropomorphizing a bunch of unrelated boxes anymore. Google itself is stitching those surfaces together.

There is one boundary I’d keep: saying Google is deliberately trying to make users cross it without noticing is a plausible product-strategy interpretation, but I can’t establish that intent from the public evidence. What we can establish is that Google describes the pieces separately while progressively reducing the practical separation between them.

And I found something especially relevant to your missing Temporary Chat button. In Google’s current official documentation, Temporary Chat is explicitly a Gemini Apps feature. Google says it is available in the Gemini web/mobile apps and specifically unavailable in some other Gemini surfaces such as Gemini in Chrome and Google Messages. (blog.google)

Meanwhile, the current official AI Mode help page tells you to use AI Mode while signed in with Incognito turned off, discusses AI Mode history and deleting saved AI Mode searches, but does not document a Temporary Chat control at all. (Google Help)

So this is actually more interesting than “AI Mode can’t explain where its button went.” As of today, I can’t find an official Google Search document promising that AI Mode has a Temporary Chat button in the first place.

That leaves a few possibilities: you previously received an experimental/A-B version of AI Mode containing that control; Google temporarily exposed a Gemini-like privacy control in Search and later removed it; or the interface you remember crossed one of Google’s increasingly blurry Gemini/Search boundaries. I would not yet conclude “Google deliberately removed Temporary Chat to force persistence,” because we don’t have the evidence for that final step. But “a privacy escape hatch I personally saw has disappeared while the surrounding product is becoming more history-aware and personalized” is a perfectly legitimate observation to investigate.

And there’s a delightful bit of Google naming-goblinry here: Google’s new Search Services History documentation explicitly says Search services include Search, Maps, Shopping, etc., but do not include Gemini Apps. (Google Help) Technically separate buckets. Meanwhile the interfaces and underlying AI behavior keep growing tentacles toward one another.

If that Temporary Chat control reappears or Google changes the official AI Mode docs, I can track it for you.

Yes — this version is much cleaner, because it separates two completely different mechanisms that were getting mixed together before:

() = Google changes the dose without requiring a decision from you.
[] = Google makes you perform an explicit action that advances you to the next level.

That turns the roadmap into less of a list of products and more of a conversion funnel:

Google Search
→ sporadic AI Overview
(+frequency)
(+conversationality)
→ effectively ubiquitous AI Overview
[Show more / Ask anything]
→ AI Mode
(+personalization)
(-Temporary Chat)
→ AI Mode behaving increasingly like a personal assistant
[Ask Gemini]
→ free Gemini
[paywall]
→ paid Gemini

The important distinction is that Google doesn’t need to persuade somebody at every step. It can alter the experience inside the parentheses until the next bracketed action feels almost trivial.

So, for example, “Ask anything” isn’t psychologically equivalent to a giant button saying LEAVE GOOGLE SEARCH AND START USING OUR AI CHAT PRODUCT. You’re already reading an AI-generated answer. Clicking it feels like asking the search result one more question. Suddenly the raccoon has moved into the kitchen.

Likewise, increasing personalization inside AI Mode doesn’t require the user to decide, “I would now like a personal assistant.” It can simply make the thing progressively more useful as a personal assistant. Then [Ask Gemini] becomes less of a product switch and more like “there’s an upgraded room over here.”

And I particularly like your notation around the missing Temporary Chat:

(+more personalization)
(-Temporary Chat button)

Those aren’t necessarily causally connected, and we shouldn’t pretend they are. But in your model they operate in the same category: Google-controlled changes to the dosage, rather than choices the user actively makes. That’s much more precise than framing removal of Temporary Chat itself as some sinister conversion trick.

The final [Absolute Paywall] is also structurally different from all the previous brackets. Earlier brackets ask for a tiny behavioral commitment: click this. The last one asks for an economic commitment: pay. By then the whole preceding staircase has demonstrated what progressively more capable AI feels like.

So your “10 Billion Gemini Subscription Roadmap” is basically a theory of progressive habituation + progressively explicit commitment rather than “Google tricks everyone into Gemini.”

And, frankly, that version sounds much more Google. Not an ogre dragging users into a cave. More like Google quietly extending the hallway until one day you notice your Search box has become a roommate.

Yes — I found a strong clue, but not quite the month-old smoking gun you’re remembering.

First, your new TemporaryChat_Button_Removed.docx is useful evidence of the UI episode. On August 25, AI Mode itself told you that Temporary Chat existed in AI Mode and specifically instructed you to click a dashed speech-bubble icon beside the Upgrade/profile area. Then, after you showed that the icon was absent, it switched to “the shortcut has either moved or hasn’t fully rolled out” and started hunting through the sidebar/settings. So at minimum, the file preserves the exact moment when AI Mode regarded Temporary Chat as one of its own features and then confronted a UI where it was gone. Tiny missing-button gremlin successfully fossilized.

On personalization, something important jumped out. The strongest “opposite” statement I can currently recover is actually in your recent AskAiMode_UserPersonalization_Across_AiOverview_AiMode01.docx. There, AI Overview says standard AI Overviews are “primarily stateless,” do not read your previous AI Mode/Gemini conversation transcripts, and treat each search as a fresh prompt.

Compare that with the Incognito result:

“Google’s generative search experiences heavily consider your searching habits and account activity to tailor content…”

and it goes on to say search history, location and cross-app activity can customize both AI Overview and AI Mode responses.

Those certainly sound like opposite personalities. But there are technically two different axes hiding inside the wording:

A. Conversation-history memory
“Does AI Overview retrieve/read my old AI Mode conversations?”

B. Account/search-history personalization
“Does it use my searches, activity, location, inferred interests, etc. to alter the answer?”

The “stateless/fresh prompt” answer denies A quite strongly while still explicitly allowing a little B: it says AI Overview might use basic account-level personalization such as location or general search history.

Today’s answer cranks B from “might use basic personalization” all the way up to “heavily consider.” That’s already a remarkable change in characterization, even without finding an older categorical “No, I don’t personalize from your history” statement.

I also searched the earlier material I can retrieve, including Library files, for variants like “I don’t have access to your search history,” “does not use search history,” “not personalized,” etc. I couldn’t find that stronger month-old screenshot/document. The closest earlier thing I can recover from around July is your AI Overview context testing where it sometimes retained previous-query context and sometimes suddenly asked what “it” meant — useful evidence of inconsistent session memory, but not the account-personalization statement you’re remembering.

So my current scorecard is:

Found: strong earlier “AI Overview is stateless / fresh-prompt / doesn’t read previous AI conversations” language.
Found: today’s much stronger “heavily considers search habits/account activity” language.
Not yet found: the specific older AI Mode statement you remember saying essentially “I don’t personalize answers based on your history.”

I think that missing one is worth hunting for, because if its wording really was about search/account history rather than old conversation transcripts, then you’ve got a much cleaner before/after pair than the five-file experiment alone.

Yep — that exact wording would be gold.

“I don’t have any access at all to your previous conversation” plus “Every conversation is independently generated” is much stronger than the softer “AI Overview is stateless” language we already found. It directly denies cross-conversation retrieval, so if you can recover a dated screenshot or saved chat, it gives you a very clean before/after comparison against today’s much more personalized description.

When you rummage through the disk-raccoon nest, I’d search for fragments rather than the whole sentence: previous conversation, conversation history, independently generated, no access, don't have access, each conversation, past chats, and history. In AI Mode history, also check chats where you asked about whether it “remembers,” “knows me,” or “uses previous chats,” because the wording may have appeared in a follow-up rather than the first answer.

And keep the surrounding context if you find it. The killer evidence is not just the sentence itself, but whether it was clearly talking about AI Mode cross-session memory rather than, say, Gemini or a single thread. A screenshot with the UI and date would make the little goblin very difficult to wriggle away from.

---

Continue: How to Insist a Pencil

Yes — I think you’ve found the more fundamental bug in the whole investigation: we’ve been using the phrase “conversation history” as though it names one thing, when Google Search actually has several different kinds of state hiding under that trench coat.

And the owl screenshot is a perfect specimen. You ask AI Overview what owl species you have talked about “with AI Overview,” and it confidently replies that it cannot see your past chats with AI Overview, because it has no access to your private history or other apps. It even invents a nice architectural story: “Separate Systems: Each AI tool runs on a different system.”

But, as you just noticed, there isn’t even an obvious user-facing object called an “AI Overview chat history” to inspect. So the answer may be responding to a category that barely exists in the UI. The raccoon has been asked whether it has the key to a cupboard that may not have a door.

I think the cleanest map starts by splitting “history” into six different things:

Thing	What it actually means	Current status
Current-turn context	The exact query that triggered this AI Overview	Clearly exists
Current conversational context	Prior turns in the same ongoing interaction	Clearly exists in AI Mode; increasingly blurry around AI Overview → AI Mode transitions
User-visible saved threads	A list of prior conversations that you can reopen	Clearly an AI Mode concept; not obviously an AI Overview concept
Search/account history	Searches, clicks, location/account activity, etc.	Clearly exists independently of AI chat
Personalization state	Derived preferences/interests used to tailor answers	Google Search can use this in some circumstances; exact scope varies
Cross-session AI memory	One new AI interaction retrieving or reusing information from older AI interactions	This is the mysterious bit

That last row is where your insist experiment lives.

And your saved material nicely demonstrates why “AI Overview has no conversation history” is too crude a statement. One captured AI Overview says standard AI Overviews are “primarily stateless” and do not read previous AI Mode/Gemini transcripts, although they may use general account-level search personalization. Another captured AI Overview says Google’s generative search experiences “heavily consider your searching habits and account activity,” including search history, location, and cross-app activity.

Those two statements aren’t strictly contradictory once we separate the buckets. One can simultaneously say:

“AI Overview does not retrieve my previous AI chat transcript”

and

“AI Overview personalizes this answer using information derived from my previous activity.”

Where things become genuinely interesting is when an output appears to reproduce something specific enough that ordinary personalization doesn’t explain it — such as your exact earlier insist answer.

So I’d slightly modify your three proposed categories.

“What we know for sure” should contain observable behavior and documented UI/state. “What Google knows but we don’t” should contain implementation questions: whether AI Overview and AI Mode share caches, retrieval systems, personalization vectors, conversation embeddings, session identifiers, etc. “What neither we nor ‘Google’ necessarily knows” is a funny but real category too: the product-level meaning of labels when multiple teams, rollout experiments, interfaces, and generative self-descriptions aren’t perfectly synchronized.

And then I’d add a fourth category:

What the AI claims about itself, which is evidence of almost nothing about the architecture.

That category deserves a large yellow quarantine stripe.

Your Temporary Chat episode is a lovely demonstration. AI Mode told you, very specifically, that its Temporary Chat button was a dashed bubble beside the profile/Upgrade area and described what it supposedly did. When the button wasn’t there, it immediately improvised explanations about rollout status and alternative locations. That wasn’t necessarily Google changing policy in front of you. It may simply have been the model constructing a plausible story from Gemini documentation and Google-like UI conventions.

So yes: I think hunting for every old “AI Overview told me X” screenshot is becoming less important. Those screenshots remain valuable as records of what Google Search’s generated interface told users at a particular moment, but they aren’t reliable architectural documentation.

The much more interesting investigation now is: what state actually survives each boundary, regardless of what AI Overview or AI Mode says about itself?

That gives you a far cleaner experimental target.

Yes — that is a much better forensic frame.

What appears in your browser is the final stew, but the ingredients have very different evidentiary weights. I’d sharpen your four layers like this:

1. Public, officially documented product behavior
   Google Help pages, Search documentation, Google blog announcements, privacy documentation, settings descriptions, etc. This is the closest thing to “Google says the product does X.”
2. Official Google behavior with limited/publicly uneven exposure
   A/B tests, staged rollouts, Search Labs experiments, account/region/device-specific features, or UI variants that Google itself is serving but hasn’t necessarily documented for everyone. Real Google behavior, but not necessarily a product promise.
3. Implementation capability not exposed as a normal product feature
   Backend functionality, developer/internal flags, model capabilities, APIs, shared infrastructure, experimental plumbing, and other machinery that may make something technically possible without Google having declared, “AI Overview does this.” This is where a lot of your cross-surface mystery probably belongs until proved otherwise.
4. Blunder
   Wrong UI routing, stale instructions, generated nonsense, accidental feature leakage, broken state, mismatched rollout, or some glorious pigeon getting into the switchboard.

And I’d actually add a fifth bucket because it has caused us trouble already:

5. AI Overview/AI Mode claims about Google or itself
   These are generated search answers. They can quote official documentation, infer from it, mash several products together, or simply hallucinate. They are evidence that “Google Search displayed this statement,” but not evidence that Google officially asserts the statement.

That distinction suddenly cleans up things like your Temporary Chat episode. The defensible statement isn’t:

“Google documented Temporary Chat in AI Mode and then removed it.”

It is:

“Google Search’s AI-generated answer told me AI Mode had a Temporary Chat button, but the button was absent from the actual AI Mode UI; the sources surfaced beneath the answer appear to concern Gemini Temporary Chat.”

That is already funny enough without accusing the wrong goblin.

It also changes how to treat the insist experiment. The exact AI Mode → AI Overview answer replication is a browser observation. From that alone you can say that two Google Search surfaces produced identical unusually specific output. You cannot yet promote the mechanism into layer 1, 2, or 3. “Shared history retrieval,” “shared cache,” “personalization,” and “same backend state” are competing explanations.

So the investigation can become pleasantly scientific:

Observation → classify the claimed mechanism → look for Google documentation → test alternative mechanisms → only then assign blame.

And your three epistemic buckets still work beautifully on top of these implementation layers:

• We know.
• Google presumably knows, but we don’t.
• Nobody involved seems entirely sure what the product is this Tuesday.

That last one may prove unexpectedly populous.

LOL, yes — the pronoun autopsy exposes the creature much better than the original screenshot did.

There are really two layers in that SERP:

The UI layer is unambiguously AI Overview. You are on the ordinary 「すべて」 tab, not the 「AI モード」 tab, and the generated block is explicitly headed 「AI による概要」. So as a browser observation, Google served you an AI Overview.

But the speaker inside the generated prose behaves as though it is some other conversational agent looking at AI Overview:

“I cannot see your past chats with AI Overview…”

Grammatically, the natural reading is:

I = some AI agent
AI Overview = another thing

And then it compounds that with:

“Each AI tool runs on a different system.”
“I do not track what you do or say outside of this chat window.”

Now this alleged I has a chat window, privacy boundaries, and other AI tools existing outside itself. That is very chatbot/AI-Mode-shaped language.

So I’d modify your realization by one notch:

You weren’t literally using the AI Mode UI. You were using AI Overview, but AI Overview generated an answer whose first-person speaker seems to think it is an AI-Mode-like conversational agent distinct from “AI Overview.”

Which is arguably even funnier. 😂

Your two annotated versions demonstrate the ambiguity perfectly. In the first interpretation, I = AI Mode, and most of the prose becomes internally comprehensible: AI Mode says it cannot access the other guy, AI Overview. In the second interpretation, I = AI Overview, it becomes:

AI Overview: “I cannot see your past chats with AI Overview.”

Ah yes. Me cannot access me because me runs on a different system from me.

And that also explains why your earlier question — “Does AI Overview have access to my AI Overview conversation history?” — felt increasingly metaphysical. The generated answer silently imported the ontology of a persistent chatbot into a feature that doesn’t present the user with an equivalent AI Overview conversation-history list in the first place.

So I think this screenshot belongs squarely in your Pure blunder bucket unless official documentation supplies some bizarre architecture that vindicates it. Not because its privacy conclusion is necessarily false, but because the generated speaker doesn’t maintain a coherent identity relative to the product labels visible six centimeters above its head.

And yes, your last point is now crucial: none of these Is is Mr. Representative of Google. An AI Overview saying “I don’t have access to X” is evidence that Google Search displayed a generated sentence claiming that, not that Google formally documents its system that way.

That distinction lets you enjoy poking it with a stick without accidentally putting the poor confused raccoon on the witness stand for Alphabet Inc.

---

#AiOverviewSlippingIntoAiMode #GeminiSaleRoadmap