Optimizing for Google AI Mode: a dozen searches you never see

It is not AI Overviews with more words

Google's help documentation describes AI Mode as expanding what AI Overviews can do, with more advanced reasoning and ways of interacting, and then gives the mechanical difference in one sentence: it divides your question into subtopics and searches for each one simultaneously. Users reach it at google.com/ai, through the AI Mode button on the results page, or from the Google app home screen. The same page puts availability at more than 150 countries and territories and more than 90 languages.

The line that matters commercially sits in Google's developer documentation, not the consumer help pages: AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary. Read that as written. A citation in an AI Overview is not evidence of a citation in AI Mode, and a report that merges the two is telling you less than it looks like.

Query fan-out changes the unit of competition

Google defines the retrieval technique in its own optimization guide as a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query. Dounia Berrada, a Search Senior Engineering Director at Google, gave the scale of it plainly in a company post on 5 March 2026: AI Mode is basically doing a dozen searches for you in the time it takes to do one.

Sit with what that does to a keyword strategy. The query your rank tracker watches is not the query that retrieved you. You are pulled in by one branch of a decomposition you never see, on a phrasing nobody typed. And you cannot reconstruct it afterwards, because Google's generative AI performance report gives impressions without query rows, which is one of the reasons measuring AI visibility needs four instruments rather than one dashboard.

The useful reframe is to stop asking what term this page targets and start asking what a model would break this question into. If a buyer asks which option suits a small team, the fan-out is going to want pricing, limits, setup effort, migration, support and a comparison. A page that answers the headline and gestures at the rest is a page that wins one branch out of six.

Google's own answer is that there is no AI Mode lever

This is the part most guides on this topic quietly skip. Google states the eligibility rule in one sentence: to be eligible to be shown in generative AI features on Google Search, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. That is the whole gate. There is no second index and no separate qualification.

Then it goes further. From the same guide, last updated 10 July 2026: you do not need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search including its generative AI capabilities, because Google Search itself does not use them. That sentence retires llms.txt as a Google tactic. We publish one and we keep a guide to it, because some other engines refer to it in their own documentation, but anybody selling it to you as the way into AI Mode is selling something Google has explicitly said it does not read.

Same treatment for markup: structured data is not required for generative AI search, and there is no special schema.org markup you need to add. Schema still earns its place for rich results and for the machines that do parse it, and it is cheap to maintain, so we keep recommending it. It is simply not the switch, and it should not be priced like one.

The switches that do exist all point down

Google names the controls that limit what appears: nosnippet, data-nosnippet, max-snippet and noindex. Because snippet eligibility is the stated gate for generative AI features, those are the settings with a documented effect on whether you show up in AI Mode, and every one of them subtracts. Set nosnippet and you leave the surface. The answer still gets written, from whoever is left.

Which makes the crawl and permission layer the first thing to audit, not the last. If a firewall rule, a stale robots line or an inherited max-snippet is quietly holding you out, no amount of content work reaches the surface. That layer is covered in the guide to AI crawlers, and it is where our audits find the cheapest wins.

Preferred sources: a real lever, held by your readers

One direct control does exist, and it belongs to the searcher rather than to you. Google's preferred sources documentation, last updated 27 May 2026, says that when a user selects your site as a preferred source your content is more likely to appear in Top Stories, and lists AI Mode and AI Overviews among the surfaces it affects wherever they are available.

The practical detail is worth knowing. Only domains and subdomains qualify, so a subdirectory like /blog is not eligible. Site owners can point readers at it directly with a deeplink in the form https://google.com/preferences/source?q= followed by the site URL, and Google publishes button assets in 15 languages for putting the ask on your own pages.

Be honest about who this suits. For a publisher with a mailing list and returning readers, one line in a newsletter is a genuine AI Mode lever. For a brand with no audience to ask, it is a button nobody presses, and the effort belongs elsewhere.

What actually moves

• Cover the subtopic space, not the head term. One page per real buying question, and inside it a section per sub-question, phrased the way people ask rather than the way you name your features.
• Write passages that survive being lifted alone. The answer in the opening line of the section, the qualification after. A section that only makes sense in sequence is a section the model will not use.
• Be unambiguous about who you are. Fan-out retrieval rewards clean entity signals, because the model has to decide you are the right company before it decides you are the right answer. That is the argument in entity SEO.
• Give it more than text. Google notes that generative responses can bring in relevant images and video, which means more opportunities for a site to appear beyond web page links.

Google's own summary of the whole exercise is worth quoting because it is the opposite of a growth hack: creating content that people find unique, compelling and useful will likely influence your website's presence in generative AI search in the long run more than any other suggestions.

The trap that comes with fan-out

Once you understand that one question becomes many, the tempting move is to publish a page for every branch. Google saw that coming. On 8 May 2026 it clarified that its spam policies also apply to generative AI responses in Google Search, and named scaled content abuse specifically. A hundred thin pages built to catch every limb of a decomposition is the exact behaviour that policy describes. Depth on the questions you can genuinely answer beats breadth you cannot defend.

A 30-day plan

Week one, confirm you are eligible at all. Indexed, snippet allowed, no inherited max-snippet, crawler access clean at the firewall. Week two, take your five highest value buyer questions and write out the decomposition for each, six to ten sub-questions per question, then mark which ones your site currently answers properly. Week three, close the gaps inside existing pages before creating new ones, and restructure so each sub-answer opens its section. Week four, sample. Run the same prompts repeatedly in AI Mode, log who gets cited, and keep the prompt set frozen so next month's run is comparable. That last habit is the one most teams skip and the only one that turns any of this into evidence.

Sources

Everything above is anchored to Google's own documentation. Start with Google's guide to optimizing for generative AI features, updated 10 July 2026, then AI features and your website for the snippet controls, preferred sources for the reader-side lever, and Google's help page on AI Mode in Google Search for how people actually reach it. The fan-out quotation comes from a Google company post dated 5 March 2026.

Common questions