AI Mode in Google Search: what it is and how to rank in it.

1. What AI Mode in Google Search is.

AI Mode is a tab. That is the least glamorous and most useful way to describe it. Alongside the usual All, Images, and News tabs, Google puts a conversational surface that takes a question, writes a prose answer, attaches a set of source links, and keeps the thread open for follow-ups. It shipped as a Labs experiment in March 2025, went general in the US that May, and expanded from there.

Its scale stopped being theoretical some time ago. In the Search announcements at I/O 2026, Google said AI Mode had surpassed one billion monthly users just one year after its debut, with queries more than doubling every quarter since launch. Whatever you think of the interface, a surface at that size is not a pilot you can wait out.

What makes it different from a chatbot is that it is wired directly into the search index rather than sitting beside it. What makes it different from classic search is that the answer is written rather than listed, so the competitive unit is a citation inside a paragraph instead of a position in a list. That shift is the whole discipline covered in our generative engine optimization guide, and AI Mode is currently the largest single instance of it.

One point of vocabulary before anything else, because it causes more confusion than any other part of this topic. AI Mode is the tab. AI Overviews is the generated block that appears above the blue links on the standard results page. They are two different surfaces with two different behaviours, and section two is about why treating them as one thing produces bad decisions.

2. AI Mode vs AI Overviews: the split that matters.

Same company, related technology, different surfaces. AI Overviews is passive: you run a normal search and a generated block may or may not appear at the top of the page, with the ranked results still underneath it. AI Mode is chosen: the user opens a tab and gets a conversation with no ranked list at all. The first is something that happens to your SERP. The second is a place your SERP no longer exists.

The interesting part is that they do not cite the same pages. Ahrefs ran 540,000 query pairs through both surfaces and found only 13.7% of cited URLs overlapping, rising to just 16.3% among the top three citations. Word-level overlap between the two answers was 16%, and they opened with the same sentence 2.51% of the time. Semantically they agreed: mean similarity across the pairs was 86%. The two surfaces say roughly the same thing and credit almost entirely different people for it.

Two operational consequences follow. First, a citation win in one surface is not evidence of a win in the other, so any report that blends them into a single AI visibility number is averaging two unrelated measurements. Second, AI Mode is the more citation-dense of the two: only 3% of its responses carried no citation at all, against 11% of AI Overviews, and it named 3.3 entities per answer against 1.3. More slots, more brands named per slot, and a user who is further into a question. That is a better surface to be optimizing for, not a worse one.

The framing you actually want is one workstream and several outputs. Direct-answer surfaces are the subject of our answer engine optimization guide, the discipline-wide view sits in AI search optimization, and if you are still untangling the acronyms themselves, the GEO and SEO comparison does that job. AI Mode is one destination those pages feed.

3. How AI Mode picks the sources it cites.

Google is unusually direct about the mechanism. Its documentation on AI features and your website describes query fan-out, issuing multiple related searches across subtopics and data sources to develop a response, and says the technique lets it surface a wider and more diverse set of links than a single search would. That one sentence explains most of the behaviour people find surprising.

Walk it through with a real question. A user asks which CMS to pick for a content-heavy site. AI Mode does not run that string against the index and read the top ten. It generates its own set of sub-questions: what content-heavy means in practice, how the main platforms handle editorial workflow, what each costs at volume, what migration involves, what breaks at scale. Each sub-question retrieves separately. Then the model assembles the answer from whichever passages won each of those individual retrievals, and cites them.

So the page competing for the visible query is not the page being retrieved. The page answering sub-question four is. This is why a modest site gets cited next to a household name in the same answer, and why the strongest predictor of appearing is coverage across a topic rather than authority on one string. Ranking still helps, because it makes you a candidate in more of those background searches: Ahrefs found 38% of AI Overview citations coming from pages in the top 10, down from 76% a year earlier. Two thirds now come from outside it.

None of this suspends the entry requirement. AI Mode retrieves from the same index as ranked search, so crawlable HTML, indexation, internal links, and page quality still decide whether you are eligible at all. That is the same pre-flight our SEO automation pipeline runs before it writes a word, and it is unglamorous for a reason: it is the part that disqualifies people.

4. What AI Mode means for publishers.

The honest summary: fewer clicks per query, better clicks per visit, and a reporting layer that has not caught up with either. The Pew Research Center instrumented the browsing of 900 US adults and found that users clicked a traditional search result on 8% of visits where an AI summary appeared, against 15% of visits where none did. AI Mode goes further than a summary, because there is no ranked list beneath it to fall back to.

The visits that do survive are different in kind. Somebody arriving from a cited link has already read a synthesis, formed a shortlist, and clicked specifically to verify one claim or see one product. That is a late-stage visit wearing the clothes of a top-of-funnel one. Teams that judge the channel on raw sessions read a working program as a failing one and cut it at exactly the wrong moment.

There is a structural upside worth naming. Fan-out rewards breadth over authority in a way the ten blue links never did, which is the first meaningful opening a small site has had in years. If your competitor has one comprehensive pillar page and you have nine focused pages that each answer a real sub-question properly, fan-out favours you on more of the underlying retrievals. That is a clustering problem rather than a writing problem, and it is what building topical clusters at scale is for.

The cost side is where most publishers get this wrong. Covering a question set is more pages, and more pages written to the same evidentiary standard, which is exactly the trade-off that turns into thin content when a team tries to hit it by hand. Our fact-check pass exists because the standard has to survive volume, not despite it.

5. How to rank in Google AI Mode.

Start with what Google says not to do, because it deletes most of the market's advice. The same documentation states plainly that there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations necessary. No AI-specific schema, no machine-readable manifest, no separate writing style for machines. If a vendor's core mechanic is a file the engine has said it ignores, that is not a strategy, it is a product looking for a problem.

What is left is short and mostly editorial. Be retrievable first: server-rendered HTML, indexed, internally linked, not thin. Then map the fan-out for your target question by writing down the six to ten sub-questions a reasonable system would generate from it. Those sub-questions are the actual retrieval targets, and most teams have never written them down, which is why their pages answer the head term beautifully and the sub-questions not at all.

Then answer each one in a passage that survives extraction. Two to four sentences, complete on their own, no pronouns pointing at the paragraph above, the claim before the explanation rather than after it. Attach one specific figure and one named source per claim, because a number with attribution can be lifted whole and an adjective cannot. Name entities in full and use the same name every time so the retrieval layer can resolve what the passage is about instead of inferring it.

Then cover the set rather than the string. One page can carry several sub-answers under clear H2s, and a cluster carries the rest. This is the same logic behind LLM SEO and the reason the GEO agent encodes these constraints in the draft prompt and enforces them at the quality gate, rather than leaving them to whoever remembers on the day. The direct-answer half of the same job runs through the AEO agent.

6. What you can actually measure.

Less than the dashboards imply. There is no impression report for AI Mode, so every product in the category works the same way: it runs a fixed set of prompts on a schedule, records which sources each response cited, and reports your share over time. That is sampling, not measurement. AI Mode responses vary between runs of the same prompt, so a single reading tells you nothing and only monthly trend lines across twenty or more prompts are worth acting on.

Two traps follow. Prompt selection is the entire experiment, and a set written by your own marketing team will quietly favour the vocabulary your marketing team already uses, producing a flattering number that no buyer would have generated. Cross-tool comparison is close to meaningless, because two products sampling different prompts at different frequencies will report different shares for the same brand in the same week. We went through which of them earn their price in our roundups of the GEO tools worth paying for and the broader field of AI visibility platforms, and which surfaces are worth sampling in the first place in our list of the AI search engines people actually use.

Three signals you already own beat most of what a tracker adds. Search Console impressions on answer-shaped queries persist even when the click does not, so they still tell you whether you are being surfaced for the sub-questions. Referral traffic from assistant hosts is low in volume and unusually late-stage in intent. And branded search volume is the honest lagging indicator that being quoted is doing something commercially. The industry-level figures behind all of this live in our AI search statistics reference rather than being restated on every page.

A worked example: one page, fanned out.

Take a fictional payroll product, Tallow, with a page targeting how much payroll software costs for a small business. Below is the fan-out written out, then the same passage twice: the version most pages ship, and the version AI Mode can use. The company and its numbers are invented. The shape of the difference is the point.

Four things changed and each maps to a sub-question rather than to the head term. The answer arrives in the first sentence and is complete without the paragraph around it, so it can be extracted whole. Real figures replace a promise of figures further down. The source and date of those figures are attached, which is what makes the number quotable rather than merely specific. And the second paragraph answers the add-on and scaling sub-questions directly, so the page is a candidate for three retrievals instead of one.

Note what did not change: the URL, the title, the schema, the internal links, the overall word count. This is a top-of-page rewrite applied to the twenty pages that carry real commercial intent, and for most teams it is the entire AI Mode project for the quarter. Everything a tracker reports afterwards is downstream of whether that edit happened.

Common mistakes with AI Mode.

1. Treating AI Mode and AI Overviews as one surface. They overlap on 13.7% of cited URLs. A blended AI visibility score averages two unrelated measurements and hides which surface is actually working, which is the one thing you needed the report for.
2. Optimising the head term and ignoring the fan-out. Retrieval runs against sub-questions the system generated, not the string you targeted. A page that answers the headline question perfectly and none of the follow-ups competes for one retrieval out of eight.
3. Buying a file the engine ignores. Google states there are no additional requirements and no special optimizations for AI Mode. Every product sold on AI-specific markup or a machine-readable manifest is charging for a mechanism the documentation disowns.
4. Burying the answer. The page contains the answer, in paragraph nine, phrased as a continuation of paragraph eight. It cannot be cut out, so it never is. Retrieval takes passages, and a passage that only makes sense in sequence is not a passage.
5. Judging the channel on sessions. Citations send fewer visitors and send them later in the buying process. Measured on raw traffic alone, a program that is winning reads as one that is losing, and gets cancelled at the wrong moment.
6. Buying the monitor before fixing the pages. A citation dashboard is an instrument, not a treatment. With no liftable answers and no sourced claims on the site, it will faithfully report zero for two quarters and invoice you monthly for the privilege.
7. Abandoning classic SEO to chase the new tab. AI Mode retrieves from the same index as ranked search. Let technical health and indexation slide and you lose both surfaces at once, which is the most expensive version of this mistake.