Google does not publish separate weights for AI Overviews. Its AI features retrieve pages with the same core Search ranking systems that rank ordinary results. So the factors that count most are the ones that already rank pages: relevance, original helpful content, links and spam signals. Then the passage that answers the question best gets quoted.
Is SEO still relevant for generative AI search?
Yes. Google says it plainly: SEO best practices stay relevant because generative AI features on Google Search are rooted in core Search ranking and quality systems.
The mechanism is retrieval-augmented generation. Google describes it as a technique that improves the quality, accuracy and freshness of AI responses by relying on core Search ranking systems to retrieve relevant, up-to-date pages from the Search index. Read that as two steps. First Search finds candidate pages. Then the model writes an answer from them and cites some.
That has a plain consequence. A page that is not in the index, or that ranks nowhere near the question, has no way into the answer. No AI-specific trick changes that.
The model on top does change. Gemini 3 has been the default model for AI Overviews since January 2026, and Gemini 3.5 Flash became the default for AI Mode at Google I/O 2026. The retrieval underneath is still Search.
Query fan-out: how you get cited without ranking first
One question rarely triggers one search. Google's guide describes query fan-out as a set of concurrent, related queries the model generates to fetch more results for the user's question. The full definition is in what is query fan-out.
So the pool of candidate pages is wider than the top ten for the words the person typed. A page can be pulled in because it ranks for one of the related questions. That fits what one study found: 47% of AI Overview citations came from pages ranking below position 5.
For anyone writing pages, this is the most useful fact here. Each section that cleanly answers one concrete sub-question is a separate chance to be retrieved. A page that only covers its topic in general terms matches none of those sub-questions well.
The core ranking systems that do the work
Since retrieval runs on Search, Google's ranking systems are the real ranking factors for AI answers. SEO.com lists more than 15 of them, including BERT, freshness systems, link analysis, neural matching, original content systems, passage ranking, reliable information systems, the reviews system, site diversity and spam detection.
A few of them matter more than the rest when the output is a short answer with citations.
Links still count. Google uses PageRank and its other link analysis systems to understand how pages relate to each other and what the pages they link to are for. Links from pages in your field tell Search which pages are worth retrieving for a topic.
Passage ranking is on the list. AI answers quote a few sentences, not whole pages, so a section that answers well on its own is the unit that gets picked.
Original work is rewarded. The reviews system rewards insightful analysis and original research written by experts or enthusiasts who know the topic well.
Spam rules still apply. The policies that reach AI responses include keyword stuffing, cloaking and link spam.
AI-written text is not the problem in itself. Google does not penalize AI-generated content as part of its helpful content system. Volume without value is. Google warns against creating separate content for every variation of a search mainly to manipulate rankings or AI responses, which breaks its scaled content abuse policy.
Write something a summary cannot replace
Google's guide is specific about what it prefers. Content with a unique point of view, such as a first-hand review based on personal experience, is prioritized over summaries of existing content.
This is the factor most sites miss. If your page restates what the top five results already say, the model has that from them. What it needs is a source for something only you know: numbers from your own product, a test you ran, a mistake you made and how you fixed it, a price you checked last week.
Ways to add it to a page you already have:
- Show the result of doing the thing, with the real output.
- Name the edge cases you hit and what fixed them.
- Date your facts and say where each came from.
- Cut the paragraphs any competitor could have written.
Deduplication and site diversity are both among Google's core systems. Our reading of that: when five pages say the same thing, an answer needs only one of them. A distinct angle is how you avoid being the copy that gets dropped.
Build a clear technical structure
AI answers can only cite what Search can fetch, read and index. That puts the unglamorous basics inside the ranking story:
- Every page you want cited is indexable: no noindex tag, not blocked in robots.txt, and listed in your sitemap.
- The main text is in the HTML your server sends, not drawn only by JavaScript after the page loads.
- One topic per page, with headings that read like the questions people ask.
- Structured data that matches what the page actually says.
Format may help on top of that. One study reports that pages combining text, images, video and structured data show 156% higher selection rates than text-only pages. Treat it as a correlation. Sites that publish all of those formats tend to do the other basics well too.
What third-party studies claim, and how far to trust them
Several SEO firms have tried to measure AI Overview selection from the outside. Their numbers are a direction, not Google's formula, and Google has confirmed none of them. The largest public set is from Wellows, which says it analysed 15,847 AI Overview results.
Two cautions before you act on these. "Semantic completeness" and "E-E-A-T signals" are scores the researchers defined, not values Google exposes, so you cannot test your own page against them exactly. And correlation runs both ways: pages that already rank well also tend to be complete, sourced and well linked.
The weak domain authority result is the most useful one for a small site. A precise, complete answer to a narrow question can be cited next to much bigger brands.
How E-E-A-T weighs against completeness and format
Google does not publish this, and none of the pages that rank for the question answer it well. Here is the reading that fits the facts above.
Think of it as a gate, then a choice. Trust decides whether a page is in the pool at all. Reliable information, reviews and spam detection are all core ranking systems, and in the Wellows data 96% of cited sources show strong E-E-A-T signals. Once a page is in the pool, completeness and how well a passage fits the question decide whether it is the one quoted.
So adding images or entities will not lift a page Google does not trust. And a trusted page that answers vaguely loses the quote to a trusted page that answers precisely.
If your site is new, fix trust first. Put a real author and a real company page behind the content, source every number, and earn links from sites in your field. Then work on how each section answers.
Which passage gets picked when sources look equal
This is the other question the ranking pages leave open. When several trusted pages cover a question equally well, these are the tie-breakers we would bet on, in order.
- The passage answers a fan-out question exactly. The model matches related sub-questions, not your title. A section headed by the sub-question and answered in its first sentence is an exact match.
- It stands alone when quoted. A passage that needs the paragraph above it to make sense is hard to cite. Wellows reports extracts cluster at 134 to 167 words, about one tight section.
- It adds something the others lack. With deduplication and site diversity in core ranking, a second copy of the same point has little reason to be shown. First-hand detail is the tie-breaker you control.
- It is current and checkable. Freshness systems are part of core ranking, and Google frames retrieval as fetching up-to-date pages from the Search index. A dated fact with its source beats an undated claim.
Why it is worth the work
AI Overviews now sit on a large share of searches. Wellows cites an Ahrefs study from 2025 that found them on over 60% of all searches.
They cost clicks when you are left out. Search Engine Land reported in 2025 that organic click-through rate drops by 61%, from 1.76% to 0.61%, on searches that trigger an AI Overview. Being cited wins part of that back: cited pages earn 35% more organic clicks and 91% more paid clicks than competitors that are not cited.
What to do first
A short order of work for a small site:
- Check that each page you care about is indexed and that its text is in the HTML your server sends. If Google has not found a page yet, submit your sitemap to Google and Bing.
- List the related questions buyers ask around your topic. Give each its own heading and a first sentence that answers it.
- Add one thing per page that only you can say: your data, your test, your screenshots.
- Source every number and put a date on it.
- Measure whether it worked. AI answers change from one ask to the next, so check the same questions several times, not once. If you build with Claude Code or Codex, Shipfound runs those repeated checks across AI engines, Google's included, and ships the site fixes and pages as pull requests.
Frequently asked questions
What are Google's ranking factors?
Google runs more than 15 core ranking systems, including BERT, freshness, link analysis, passage ranking, reliable information, reviews and spam detection. AI Overviews use those same core systems to retrieve the pages they cite.
Does AI content affect Google ranking?
Not because a model wrote it. Google does not penalize AI-generated content as part of its helpful content system. Publishing a page for every variation of a search is different: Google says that breaks its scaled content abuse spam policy.
What are the top 3 SEO ranking factors?
Google does not rank its own factors. For AI Overviews, the three with the most support are being retrievable by core Search, a first-hand point of view, which Google says it prioritizes over summaries of existing content, and trust, which 96% of cited sources show strongly in one study.
Do I need to rank in the top 5 to be cited in an AI Overview?
No. In one study 47% of AI Overview citations came from pages ranking below position 5. That fits query fan-out, where the model runs related queries to fetch more results than the one typed.
Does domain authority matter for AI Overviews?
Less than for classic rankings, by one measure. Wellows reports that the correlation between Domain Authority and AI Overview rankings dropped to r=0.18. Trust still matters, but a precise answer from a smaller site can be cited.