How On-Site and Off-Site SEO Work After AI Ate the SERP

A 1,900-word pillar page on the primary money term is four years old, clean on Core Web Vitals, and still holds position seven in Google. It has been earning less every quarter. The AI Overview above it now answers the same question with a 600-word post from a lower-authority domain, and the ChatGPT citation for the topic belongs to a competitor that was not in the category two years ago. Nothing on the page is broken. Nothing in it is retrievable in isolation, either, and that is the costly mistake. On-site and off-site SEO both still work; the mechanics that move the needle in the ChatGPT and Perplexity era are not the ones that moved it in 2022.

On-Site SEO Became Retrieval Engineering

The page-level job used to be ranking a URL for a query. The retrieval job is surfacing a self-contained passage inside that URL as an answer. Those are different optimization problems. A retriever chunks your page, embeds each chunk, and scores it against the user's question. A chunk that depends on a heading two scrolls up, or an "it" that points to a noun introduced earlier, produces a noisy embedding and loses to a cleaner passage on a smaller domain.

The practical rewrite is specific:

• Front-load the direct answer in the first 40 words of each section.
• Write H3s that name the entity and the question together, not short keyword stubs.
• Attach FAQPage, HowTo, and Product schema where the content actually matches it, and leave it off where it does not.
• Publish an llms.txt pointer that maps your canonical answers to the engines that read one.
• Include author schema with sameAs links to verified bylines, so the model has an entity to attach the content to.

The deeper change is semantic. Google's crawler indexed URLs; a retriever indexes passages and entities. Topical authority in the LLM sense is being the entity the model associates with a concept, not just the URL that ranks for the keyword. Entity SEO for LLMs is less about keyword density and more about repeated co-mention of your brand, your experts, and your category together across the web and inside your own site, in ways a graph can parse. Zyppy's analysis of 23 million internal links found that pages with at least one exact-match anchor pointing to them earned five times the Google traffic of pages without — the internal link graph tells both Google and a retriever which URL is the authoritative answer in your domain.

Off-Site SEO Shifted From Link Count to Corpus Presence

Backlinks still correlate with ranking. Backlinko's analysis of 11.8 million search results found the #1 position has 3.8 times more referring domains than positions #2 through #10. That signal is not dead. What has died is the idea that off-site SEO means only acquiring links to a URL.

Models retrieve from a corpus, and the corpus is the open web. A brand that gets discussed on Reddit, reviewed on YouTube, profiled in B2B trade press, and asked about on Quora is a brand the model has seen described in dozens of ways by sources it already trusts. Unlinked brand mentions in forum answers or long-form video reviews can move the needle in ways a DR-30 guest-post link no longer does. The weighting is still debated, but the correlation is strong enough that it belongs inside the off-page plan rather than parked with the PR team.

Two shifts to make:

• Digital PR targets citation bait, not only link equity. The deliverable is a figure, benchmark, or named framework that becomes the quotable detail in an AI answer. Journalists link it; models pull it back out.
• Domain trust is now a retrieval signal, not only a ranking signal. Google's site reputation abuse policy already punishes host sites that lend their ranking to unrelated third-party content. On the retrieval side, candidate documents are weighted partly on the embedded reputation of their domain, so a site flagged for parasite content loses trust across both surfaces at once.

The Signals That Now Correlate With AI Citations

A 2026 SE Ranking study of 2.3 million pages across 295,000 domains ranked the factors that predict citation in Google's AI Mode. The headline finding: domain-level organic traffic from Google is roughly three times more predictive of AI citation than any single on-page signal. High-traffic domains averaged 6.4 citations per topic; low-traffic domains averaged 2.4.

The implication is unforgiving. Publishing more pages to win AI citations will not work on a domain the engines do not already cite. The on-site rewrite matters, but it compounds on top of an organic footprint the retriever already trusts. A thin content farm that never earned Google baselines cannot citation-bait its way into ChatGPT.

That same footprint maps differently across engines. Ahrefs studied 75,000 brands and found YouTube mentions were the single strongest predictor of AI brand visibility, outperforming referring domains and branded search volume across ChatGPT, AI Mode, and AI Overviews. The off-site corpus is not symmetrical. An industry heavy on YouTube explainers feeds Perplexity much better than one living on PDF whitepapers.

Traffic patterns across verticals have split along the same fault line. Transactional categories like travel, ecommerce, real estate, and food are seeing AI referral traffic rise sharply even as Google organic slips; informational verticals like education and news are losing on both sides as models answer research queries directly.

The Retrieval Pipeline Your Content Must Clear

Perplexity's source selection runs as a five-stage pipeline: intent interpretation, live retrieval, ranking and filtering, LLM synthesis with inline citations, and follow-up generation. Each stage removes candidates. Content designed for the Google SERP has to pass filters it was not built for.

Two gates are worth naming. The retrieval gate is semantic, not lexical. BM25 keyword matching still runs, but dense retrievers score your passage against the query embedding. A page that keyword-matches but answers tangentially will lose to one that paraphrases the question and answers it in the first sentence. This is why answer-ready formatting now outperforms the long, winding intro that worked for dwell-time optimization.

The reranker gate is where domain reputation and entity trust compound. Pages from domains the model already associates with the topic survive. New content on a new domain rarely clears it on the first cycle, no matter how clean the schema. That is why the question "where do our brand mentions live?" now belongs inside the on-site brief, not parked with a separate team — a split that Claude citation data makes especially visible.

Where should you spend first? If your domain is already cited in Google results, invest in the on-site rewrite: chunking, schema coverage, entity markup, llms.txt. If the retriever does not yet know your domain, start with the corpus: Reddit explainer threads, YouTube walkthroughs, trade-press mentions, and a first-party benchmark that journalists will travel. The one-sided plan is the costly one, and it is the missing side that the engineering-led GEO practice at seo.co is built to close.

Back to the Pillar Page

The page from the opening is fixable inside a single sprint. Chunk the body into self-contained H3 answers, add FAQPage and Article schema with author sameAs, and publish an llms.txt pointer to the canonical answers. Commission a verified Reddit thread tied to the same entity, record a 12-minute YouTube walkthrough that names the brand in the first 30 seconds, and pitch a first-party benchmark to three trade publications so the figure travels. By the next crawl cycle, the passage starts showing up as a ChatGPT citation, the AI Overview sources the framework, and position seven in Google begins converting again. The page did not need more words. It needed to be built for the way the model reads it.