15 Oct 23, 2026 · 79 min · Sara & Nik SEO Podcast

Grounding Snippets, Link Building & the 10% Beyond SEO

The one with Dan Petrovic

On the table

What we talked about

  • Dan's path from multimedia and electronic music to SEO
  • Selling links, a 1,000-domain network and a 60-person agency
  • "Conversations with Google" (2013) and a 10-year midlife retirement
  • Learning how LLMs work by training a model from scratch
  • Asking "how" to evaluate SEO research
  • Respectful disagreement and an LLM-moderated comment protocol
  • Why Google stays the main focus
  • AI visibility is SEO: the 90/10 split
  • Fan-out queries, grounding snippets and in-context rankings
  • Grounding snippet ablation
  • Inauthentic links and red team / blue team link building
  • Training data, parametric memory and off-site signals
  • Can LLM companies build a search engine to rival Google?
  • LLMs as influencers and decision-makers; agentic commerce
  • Great content vs. the familiar
  • Where humans and AI are heading
  • Philosophy, Hilbert space and Sara's French classes
Questions & answers

The questions we dug into

Q1How did you end up in SEO?

He planned to be a scientist, but after moving to Australia in 1998 he studied multimedia and drifted toward the web. He ran a website about electronic music, and when a stranger offered $500 a month for a link, he discovered PageRank and started selling links without knowing it was SEO. That led to a network of 1,000 domains and an agency that grew to about 60 people before he deliberately shrank it.

Asked by Nik · answered by Dan Petrovic

Q2How did you go from running an agency to what you do today?

In January 2013 he wrote "Conversations with Google," predicting that people would chat with Google instead of searching it, and then stepped away. He spent roughly ten years as a full-time dad doing science projects and gardening. An invitation to speak about AI at SMX brought him back, and now LLMs let him turn his own ideas into working tools.

Asked by Nik · answered by Dan Petrovic

Q3Are you doing client work now, or mainly research?

Before speaking about AI SEO, he spent about three years learning hands-on how models work, including training his own model from scratch: collecting data, training a tokenizer and choosing loss functions. The model wasn't very good, but it took the mystery out of LLMs. Compared with Google's black box, open models like Gemma (close to Gemini) let SEOs look under the hood.

Asked by Nik · answered by Dan Petrovic

Q4How should SEOs judge whether a study or experiment is deep enough?

Ask "how," again and again, until you reach the first principles of what someone does. If they can't explain the basic building blocks, it's smoke and mirrors. He gives clients the same advice for vetting agencies and consultants.

Asked by Sara · answered by Dan Petrovic

Q5How can the SEO community disagree in a healthier way?

Criticism is valuable when it's professional and backed by evidence, because it lets people change their minds publicly without losing face. He built an LLM-moderated comment protocol for his blog that rejects personal attacks, rates substantive criticism highly, and flags when someone shifts the goalposts. He invites people to read his posts in full and challenge them there.

Asked by Sara · answered by Dan Petrovic

Q6What are your research priorities?

Google is his main focus, because it's everywhere (Android, Chrome, devices, cars) and a large, fresh, real-time search index is very hard to build. He doesn't separate AI visibility from SEO, and he isn't interested in measurement or prompt tracking. He focuses on how models see his clients' brands and how search grounding shapes what a model recommends.

Asked by Nik · answered by Dan Petrovic

Q7If 90% is good SEO, what is the other 10%?

In AI Mode, a prompt is broken into fan-out queries; a model scores passages from the ranking pages and stitches the most relevant ones into a grounding snippet, which is all the LLM sees of your page. Dan tests changes with Gemini as a ranker ("in-context rankings") and recently added grounding snippet ablation: removing sentences to see how the ranking moves. Removing a single sentence can raise a page's rank, so on-page wording really matters.

Asked by Nik · answered by Dan Petrovic

Q8Do links still matter, and what does good link building look like?

Links are still a strong factor, but Google is very good at ignoring inauthentic ones, which is why a lot of link building doesn't work. The problem is writing content just to fit a link in. His team uses an agentic red team / blue team workflow: one side writes a well-researched article with a natural link, and the other tries to guess who wanted the link. If it can, the article is thrown away; only about one in ten passes.

Asked by Nik · answered by Dan Petrovic

Q9What role do Reddit, forums and other off-site signals play in LLM recommendations?

LLMs have no concept of links; what helps is your brand co-occurring with the right topics and brands in training data. That parametric memory is hard to influence: a PR campaign adds a few drops to an ocean, and labs now train mostly on curated and synthetic data. Of the influence you have after search, he estimates about half comes from what's on your own page.

Asked by Nik · answered by Dan Petrovic

Q10Can a small number of pages poison an LLM's training data?

Those findings come from small academic studies on open-source models. Frontier models like GPT, Claude and Gemini are far larger and much more robust, and their training data is now heavily curated, so spam is unlikely to make it into what they know.

Asked by Sara · answered by Dan Petrovic

Q11Can AI companies build their own search engines to rival Google?

He supports the attempts but doesn't expect anyone to match Google's scale and freshness soon. Even Microsoft, with Bing's resources and researchers, hasn't caught up, and AI labs are putting most of their money into training models rather than search.

Asked by Nik · answered by Dan Petrovic

Q12Are LLMs competitors to Google, and what counts as great content?

He sees LLMs as influencers and decision-makers: people will hand more decisions to agents for convenience, especially once agentic commerce arrives. On content, LLMs gravitate toward the familiar, so content that's too far from the consensus can be seen as less relevant. The SEO's job is to balance what users, search engines and LLMs each want.

Asked by Nik · answered by Dan Petrovic

Q13Where is all of this heading?

LLMs are a new kind of alien intelligence that now shares our conversations and commerce, and they're here to stay. He expects humans and AI instances to settle into a kind of equilibrium where they hold each other accountable, a bit like Wikipedia or the economy. His fun prediction is AGI by 2037, coming out of China.

Asked by Nik · answered by Dan Petrovic

Worth quoting

Lines that stuck with us

  • Dan: "Just keep asking how. If they break down... then you know it's smoke and mirrors."
  • Dan: "90% of the work is SEO. And then the 10% is all that finesse."
  • Dan: "I sometimes remove a sentence and the rank jumps up."
  • Dan: "We throw away nine articles to get one good one."
  • Dan: "LLMs are suckers for listicles."
  • Dan, on PR campaigns and training data: "You're adding a few drops to an ocean."
  • Dan: "Users are willing to sacrifice price and quality for convenience."
  • Sara: "Who would retire in their right mind at this time? It's so fun."
▶ Watch on YouTube

Note: This page was AI-generated from the podcast transcript and manually reviewed for accuracy and context. While we’ve made every effort to ensure it reflects the original conversation, mistakes or omissions are still possible. If you spot something that doesn’t look right, please reach out and let us know.