Google AI Mode: what it is and how a small site gets cited

AI Mode is a tab in Google Search. It writes an answer in plain text and links to sources. It opened to all US users in May 2025, then more countries during 2025.

It uses Google’s index and Gemini models. It issues several sub-queries at once, a query fan-out, to cover the angles of your question. It then cites pages Google’s ranking systems already surface for those sub-queries.

There is no separate crawler for AI Mode. Googlebot is the crawler. There is no submission form. There is no opt-out beyond snippet controls that already apply in Search.

AI Mode is not Gemini. Gemini is an assistant. AI Mode is part of Search that writes an answer and points to the web pages it pulled from.

AI Mode cites pages that Search already trusts for the sub-queries it fans out. That means you win by being the best specific answer to parts of the task, not by trying to rank for the head term alone.

• Pages with clear structure: headings that match questions, short paragraphs, lists and tables that can be quoted
• Consistent names for your product and people, so the model can attribute correctly
• Numbers and comparisons it can lift, like a short feature table or a price range if you publish one
• A visible date and an author line, to show freshness and ownership
• Fast pages with content in HTML, not hidden behind scripts, since many answering crawlers do not execute JavaScript

Generative engine optimisation, GEO, is the craft here. The aim is to be named and cited by AI systems. That includes AI Mode and AI Overviews, ChatGPT with search, Perplexity, Claude, Gemini and Microsoft Copilot. There is no submit button. You earn citations with answers the systems can quote and by being on the lists and comparisons they already cite.

AI Mode issues several related searches to break down the task. You win citations by covering those angles on one page with scannable blocks the model can use.

• Aim to answer the main job-to-be-done in the first 2 to 3 paragraphs
• Follow with sections that map to likely sub-queries: pricing, setup steps, integrations, limits, examples, comparisons
• Use question-style H2s and H3s that match how people search

Example. Your page is yourproduct.com/guides/getting-started. The query is “set up invoicing for a yoga studio”. Fan-out sub-queries might be “invoice template yoga studio”, “recurring class billing”, “Stripe integration for invoices”, “cancel policy on invoices”. Add sections with those exact headings. Open each with one to three crisp sentences and a short list or a table.

What a fixed page looks like. The top answers the question in 60 to 120 words. The middle has three H2s like “Recurring billing for classes”, “Stripe setup in 3 steps”, “Template with cancel policy”. Each block opens with the answer, then a short list. The bottom links to a comparison and a pricing note if relevant. The same product name is used throughout.

• Put the direct answer first. One paragraph, then a bulleted list if it helps
• Name the product the same way everywhere: “Acme Invoicing” on /, /pricing and docs
• Use question headings: “How to export to CSV”, not “Export”
• Add a small table for comparisons: three to five rows, clear labels
• Publish an updated date and an author or team name
• Keep numbers concrete: counts, durations, limits, file sizes
• Keep critical content in HTML on load. Do not rely on client-side rendering for the main text
• Compress images and keep LCP under about 2.5 seconds, CLS under about 0.1, INP under about 200 ms

Worked rewrite. Before: “Our platform is flexible and powerful.” After: “You can import 5,000 contacts from a CSV under 10 MB. The import keeps tags. It finishes in under two minutes for most files.” The second block is quotable. It is specific and scannable. It contains figures the model can lift with attribution.

Worked outline. For /pricing, add “Is there a free plan?” with a one sentence answer, “What happens at the limit?” with a short table of caps, and “How to change plans” with three numbered steps. These blocks match real sub-queries and can earn links from AI Mode as side cites in the answer.

AI Mode does not have its own opt-out. The only controls are the standard Search snippet settings. You can shorten or block snippets on a page or section with robots meta tags or the data-nosnippet attribute. This will also affect classic results and featured snippets.

• To reduce what can be quoted, use max-snippet with a low value on sensitive pages
• To block a block-level quote like a price table, use data-nosnippet on that element
• Avoid noarchive and nositelinkssearchbox here, they are not about snippets
• Do not use Google-Extended for this. It controls Gemini training and grounding, not Search or AI Mode

Google AI Mode relies on Googlebot. Keep it allowed. Beyond Google, other assistants also cite sites. Let the answering crawlers index you or they will not name you. This is a precondition for GEO as of 2026.

• Answering and search crawlers to allow: Googlebot, Bingbot, OAI-SearchBot, Claude-SearchBot, PerplexityBot, Applebot
• On-demand fetchers to expect in logs: ChatGPT-User, Claude-User, Perplexity-User
• Training toggles you may choose to block while staying visible in answers: GPTBot, Google-Extended, Applebot-Extended, CCBot, meta-externalagent, meta-externalfetcher, Amazonbot, Bytespider, cohere-ai, DuckAssistBot

Blocking a training crawler does not remove what it already collected. Robots.txt is honoured voluntarily. Blocking Googlebot, Bingbot, OAI-SearchBot, Claude-SearchBot or PerplexityBot will make you invisible to those assistants and to AI Mode or Copilot that rely on Search.

Keep critical content server-rendered. ChatGPT’s crawler, as one example, does not execute JavaScript. If your pricing table renders only after scripts load, these systems will not read it and will not quote it.

Clicks from AI Mode are reported with the rest of web search in Search Console. There is no separate filter. The same is true for AI Overviews. Expect impressions and clicks to move without a neat label.

1. 1

   Segment by page and query clusters

   Use the Performance report to group pages by intent, like “how to” guides, comparisons and pricing. Track queries that match your question-style headings. Watch for rises in impressions without matching position changes.

2. 2

   Read referrers in analytics

   Assistant clicks arrive as different referrers. chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com are separate. Google’s AI surfaces arrive as google.com like any other search click.

3. 3

   Ask the assistants on a schedule

   Build a list of buyer questions. Once a week, ask ChatGPT search, Perplexity, Gemini and Claude. Log who is cited and linked. Check whether your page appears and what block is quoted.

4. 4

   Track with a tool if you prefer

   Products like Profound, Peec AI, Otterly, Scrunch and the AI toolkits of Semrush and Ahrefs automate the asking. DataForSEO also offers LLM mentions data.

What a pattern looks like. Your /guides/getting-started page gains impressions for “how to set up recurring invoices” without a big change in average position. Referral traffic from google.com ticks up and you see a spike in page entrance rate. When you ask AI Mode the same task, your “Recurring billing for classes” block is cited under the answer. That is your link path.

AI answers often cite list and comparison pages. If your product appears on those, you gain indirect citations. This is the most reliable route when you are new or your domain is small.

• Find the lists that appear in AI Mode and AI Overviews for your space: “best invoice apps for studios”, “Stripe alternatives for small gyms”
• Pitch inclusion with a factual blurb and figures they can quote: limits, integrations, unique features
• Publish your own comparison pages and alternatives pages. Use neutral, verifiable claims and a small table
• Use consistent naming so you are recognised across sites and assistants

Worked example. Your “/comparisons/stripe-vs-acme” page opens with a three-row table: fees, payouts, and refunds timing. It has a clear date and sources. AI Mode can quote the table for sub-queries like “Stripe payout timing compared” and attribute you cleanly.

1. 1

   List buyer tasks and questions

   Pull from support tickets and sales calls. Turn each task into one page or a tight section with a question heading.

2. 2

   Draft the answer-first block

   Open with 60 to 120 words that solve the task. Follow with a short list or table. Add a date and author.

3. 3

   Add sub-question sections

   Use three to five H2s that match likely fan-out queries. Keep each to one screen, with direct language and figures.

4. 4

   Check crawl and speed

   Fetch with Google’s URL inspection. Make sure your HTML contains the content on load. Keep LCP under about 2.5 seconds.

5. 5

   Ship and measure

   Watch Search Console for the page’s queries over 28 days. Log citations from assistants weekly. Tweak headings and blocks, not just keywords.

What a fixed page looks like. /pricing adds “What happens at the limit?” with a three-row table, marks up the table in HTML, and shows an updated date. A month later, the page earns impressions for “pricing limit policy” and gets cited in AI Mode under the pricing part of an answer.