Fundamentals

Google has three AI answers, not one. They do not run on the same permissions.

By Arnav Mukherjee, founder of TofuBofu · August 17, 2026

A draft of ours once said that Google's two AI surfaces "behave alike". We caught it in review before it shipped, because our own database said one of them names a brand roughly ten times as often as the other.

That sentence was written by someone who knew this category well. It is still the single most common mistake made about Google's AI products, and it is made by people who would never confuse ChatGPT with Perplexity. One logo, one company, one search box, therefore one thing. The confusion is not niche either: search for AI Mode and Google's own related searches come back with "AI Mode vs Gemini vs AI Overview", "Is Google AI Mode Gemini" and "Should I use Gemini or Google AI Mode?"

There are three surfaces. They draw on one index, under two different permission regimes, and Google has documented every part of that in public. What follows is sourced from Google's own pages, because the SEO industry's summaries of this have been wrong often enough that we no longer read them. Every quote below was read at its source page in August 2026, and where Google has said the same thing twice I have given you both.

Three products, in Google's words

AI Overviews is a feature inside the ordinary results page. You did not ask for it, it appeared above the blue links, and Google's help page will not tell you why it appeared for that query and not the next one. Its only public statement on triggering is that you will see one "when our systems determine that generative AI can be especially helpful", with a search where you want to understand information from a range of sources given as the example.

AI Mode is a destination you choose. Google calls it "Google's most powerful AI search experience", where "you can ask anything and get an AI-powered response, with the ability to go deeper through follow-up questions and helpful links to the web". Crucially, Google ties it back to the index: "AI Mode relies on Google Search's deep understanding of web information, meaning that responses are supported by high quality web content to improve factuality." Google also documents a failure mode worth knowing about, because it is the only case where the surface hands the reader straight back to the web: "In some cases, AI Mode will provide a set of web links if there's not high enough confidence in the quality or helpfulness of an AI response."

The Gemini app is not a search product, and Google does not describe it as one. Its help pages sell it as a place to "Get help with writing, brainstorming, learning, and more", and on Android as the thing that replaces Google Assistant: "when Gemini is your primary mobile assistant, it can help you with some of the same tasks as Google Assistant." Read the Gemini app documentation end to end and the vocabulary is assistant vocabulary. The vocabulary on the AI Mode page is search vocabulary. Google is telling you what these things are for, and the two answers are different.

Sources behave differently too. Google says of the Gemini app: "Not all responses include related links or sources. If you don't have the Sources button below a response, Gemini Apps didn't provide any links for that particular response." A surface that may return no links at all is a surface where being cited is not guaranteed to be a thing that can happen on any given answer. That is a structural ceiling on the channel, set by Google, and no amount of content work moves it.

The line that settles the argument

If you want one piece of evidence that these surfaces are governed separately rather than being three coats of paint on one product, it is in Google's crawler documentation, and it is a single sentence.

Google-Extended is described as "a standalone product token that web publishers can use to manage whether content Google crawls from their sites may be used for training future generations of Gemini models that power Gemini Apps and Vertex AI API for Gemini and for grounding (providing content from the Google Search index to the model at prompt time to improve factuality and relevancy) in Gemini Apps and Grounding with Google Search on Vertex AI". Then Google states the consequence: "Google-Extended does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search."

Read that twice, and read the parenthesis, because the parenthesis is the part everyone skips. Google defines grounding as "providing content from the Google Search index to the model at prompt time". The Gemini app is not fed by some separate Gemini crawl of the web. It reaches into the same Search index that AI Overviews and AI Mode draw on. There is one crawler and one index here, not two.

Which makes the guarantee sharper, not weaker. Google is not saying these systems read different data. Google is saying that one line in a text file changes what may be done with your data on one set of products, and does nothing to your presence on another set, from the same index. That is a permission boundary, and Google confirms it from the other side too. Its crawling overview repeats the promise in different words, "Utilizing Google-Extended doesn't affect a site's inclusion in Search, nor do we use Google-Extended as a ranking signal in Search", and Search Central's site owner guidance scopes the token out of the Search surfaces explicitly, telling you to read about Google-Extended in order to limit training and grounding "in some of Google's other systems".

Other systems. Google's own documentation for AI Overviews and AI Mode points you at Google-Extended as the control for things that are not AI Overviews and AI Mode. If the three were one product, that sentence could not be written.

Two of them share a front door. That does not make them the same room.

AI Overviews and AI Mode sit under the same permission, and Google's requirements for both are identical and refreshingly boring: "To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements."

Then Google closes the door on the entire cottage industry that grew up around this, in one sentence: "You don't need to create new machine readable files, AI text files, or markup to appear in these features." No AI Overviews schema type. No AI Mode markup. No feed to submit. Google adds that "you can apply the same foundational SEO best practices for AI features as you do for Google Search overall", and points you at the ordinary preview controls, nosnippet, data-nosnippet, max-snippet and noindex, if you want less of your page used. If a vendor is selling you technical eligibility for these two surfaces, they are selling you something Google says does not exist.

Here is where I have to correct a distinction that gets drawn constantly, including in the first version of this post. It is tempting to say that AI Overviews answers the question as typed while AI Mode fans it out into sub-questions. Google does not support that. Search Central says both surfaces "may use a" fan-out technique, describing it as issuing multiple related searches across subtopics and data sources in order to develop a response. Query fan-out is not the thing that separates the two. It is the thing they have in common, and it is the single most important mechanical fact for anyone writing pages, because it means the page that wins is often not the page that would have ranked for the phrase the buyer typed. It is the page that happens to answer sub-question four.

What actually separates them is investment and control. When Gemini 3 reached Search on 18 November 2025, Google gave AI Pro and Ultra subscribers the ability to select it "by selecting 'Thinking' from the model drop-down menu in AI Mode". A model picker, in AI Mode, and nowhere else. The automatic routing Google announced alongside it covers both surfaces, "Search will intelligently route your most challenging questions in AI Mode and AI Overviews to this frontier model", but only one of the two is a place where a user can reach in and choose.

And AI Mode is no longer the experiment it was when Google introduced it in Labs on 5 March 2025. At I/O on 19 May 2026 Google reported that "Just one year after its debut, AI Mode has surpassed one billion monthly users, with queries more than doubling every quarter since launch", and made Gemini 3.5 Flash "the new default model in AI Mode for everyone globally". A surface with a billion monthly users, doubling quarterly, running Google's current default model, is not a tab you can defer thinking about until next year.

What we measured: a 10x gap between two Google surfaces

Here is the part no secondary summary can give you, because it requires running the scans.

Across 46 completed reports covering 34 brands, all of them companies who came to us already suspecting they were missing from AI answers, we counted how often each engine named the scanned brand when it answered a buying question. Buying question means exactly one thing here: the intents our analyzer calls category_bofu, longtail_tool and decision. It deliberately excludes the probe that contains the company's own name, because every engine answers that one with the company between 82 and 100 percent of the time. Including it inflates every rate by roughly 2.6 times. We know the multiplier precisely, because we published the inflated set once, on eight engine guides and in eleven places elsewhere, and had to go back and correct all of it.

On that basis, measured on 12 August 2026 within that self-selected sample: Perplexity 7.4 percent, Google AI Mode 4.0 percent, Microsoft Copilot 3.1 percent, ChatGPT 1.7 percent, Claude 1.2 percent, Gemini 0.4 percent. One footnote on the ChatGPT row before anyone builds an argument on it: that figure measures a probe we have since retired, because until August 2026 we asked ChatGPT without web search and it answered from training weights. We have not restated it, because no post-cutover corpus exists yet and inventing one would be worse than leaving a labelled number alone.

Two of those six rows are Google, on the same questions, over the same corpus, in the same period. AI Mode named the brand roughly ten times as often as Gemini did. If the two were one system with one logo on it, that number would be impossible.

The best argument against my own number

If you are going to attack that 10x, attack it here, because this is where it is weakest and I would rather hand you the ammunition than have you find it later.

The raw counts are small. AI Mode named a brand in 12 of the 299 buying cells it answered. Gemini managed 2 out of 483. Two. A rate built on two events is exactly the kind of number that moves if one engine has a bad afternoon, and anyone who quotes 0.4 percent as though it were a precision instrument is overselling it.

Worse for me, two of the faults behind that number were ours, not Google's. Our Gemini model spent most of its token budget on internal reasoning and truncated answers mid-sentence at our production limit, so a brand named late in an answer was never counted. And our brand-matching guard was one-directional: it could push a mention to absent, but it could not rescue one, until we made it symmetric. Both are fixed. Neither was fixed in time for this corpus, which is why we hold 0.4 percent as a floor rather than a measurement, and why you will not find a corrected Gemini figure anywhere on this site. We have not re-measured it. Publishing a number we did not observe is the one thing we will not do, even when the missing number would flatter the argument.

So does the gap survive? Run the interval arithmetic on our own counts and it does. A 95 percent Wilson interval on 12 of 299 runs from 2.3 to 6.9 percent. On 2 of 483 it runs from 0.1 to 1.5 percent. The two do not overlap, and they are not close to overlapping: the most generous reading of Gemini still sits below the least generous reading of AI Mode. Small numerators widen an interval, they do not move one interval on top of another.

Then note where our two faults land. Truncation and a one-way matching guard both operate on the text of an answer we received. Neither can conjure a vendor recommendation into an answer that never made one, and neither touches AI Mode's column at all. They compress the distance. They do not invent it. My position, stated as a judgement rather than a measurement, is that the true Gemini rate is higher than 0.4 percent and the ordering is not in doubt.

One asymmetry I will concede outright, because I would raise it if I were reading this. The denominators differ, 299 against 483, because AI Mode only became a tracked engine in our product in August 2026 while Gemini has been in every scan for far longer. Same corpus of brands and questions, shorter window on one column. That is a real limitation of the comparison and it is why the honest headline is "roughly ten times", not "9.7 times", and why we will re-run this when AI Mode's window is as long as Gemini's.

The measurement hole Google's own tools leave

The split runs through Google's analytics too, and it leaves a hole you cannot fill with Google products.

Search Console's generative AI performance report covers AI Overviews and AI Mode, and per Google's documentation it "includes data from the Web search type in the Performance report (Search results)". Two surfaces, one combined bucket, with no dimension that splits them. The dimensions on offer are pages, countries, dates and devices, so there is no query breakdown at all, and Google notes that the usual limits including the 1,000 row cap still apply. In that report you can learn that Google's AI showed your page, and you cannot learn which of the two showed it. The question-level data is not lost, it simply is not there: it sits in the ordinary Performance report instead, which we have written up separately. Two surfaces, split across two reports, neither of which will tell you which surface you are looking at.

The Gemini app is not there at all, and never will be, because it is not Search. There is no Gemini property, no Gemini impressions column, nothing. So the assistant sitting on hundreds of millions of phones is invisible to the tool most marketing teams treat as ground truth, which is the entire reason products like ours query the engines directly instead of waiting for a dashboard that structurally cannot exist.

What to actually do

Stop saying "Google" as if it names one destination. In a reporting line it now means at least three things with different rules, and a single blended Google number hides the one movement worth knowing about.

For AI Overviews and AI Mode, the work is indexing and content, exactly as Google says, and there is nothing to buy. Since both surfaces fan a query out, write for the sub-questions: a page that answers one question well is competing against pages that answer the eight narrower searches the system actually issued. Comparison pages, explicit selection criteria and plain-sentence answers to narrow questions do better here than a broad pillar, because they match sub-questions rather than headlines.

For the Gemini app, go and read your own robots.txt today, and look for Google-Extended. That check takes thirty seconds and almost nobody does it, because the line usually went in during an AI-blocking decision made by somebody who has since left, and nothing has forced a review since. If you want to be recommended by an assistant that ships as the default on Android, that line is doing the opposite of what you want. And Google's guarantee cuts both ways: taking it out costs you nothing in Search, on Google's own written word, twice.

Then measure the surfaces separately and hold the questions fixed between runs. A number that moves because you changed the question is not a trend, it is a new test. And an average across surfaces that behave ten times differently from each other is not a summary, it is a way of not looking.

This is the layer SEO does not cover. Your Google rank is the floor and it still matters, because all three of these surfaces reach into the index you rank in. AEO is the layer on top, and you can win it on questions where your organic position is nowhere. We have watched Google's AI cite us for queries our own site did not rank on page one for. Same company, same index, different game.