The year SEO stopped talking about links and started talking about AI
Eighteen years of one SEO show, transcribed and counted. AI went from zero mentions before 2023 to every episode of 2026. What it displaced is measurable — and the two topics it displaced went different ways.
AI and LLM terms went from zero before 2023 to appearing in 100% of 2026 episodes. Link building did not survive that as a lesser topic — it fell on both measures, from 61% of episodes in 2022 to 21% in 2026, with its airtime falling further still. Social media is the topic that shows the more interesting pattern: it is still raised in 82–95% of episodes, but the time spent on it fell nearly 90% from its 2011 peak. It became an aside. Link building lost the room.
The shift, in one chart
Two ways to fall, and only one of them is an aside
Episode share asks "did this come up at all this year." Rate asks "how much airtime did it actually get." A topic in genuine decline falls on both. A topic that only sounds like it declined, because episodes got longer and diluted an old habit, would fall on rate alone. Neither link building nor social media fits the second pattern — both fall on rate far faster than episode length grew, which is one of the checks recorded in the method section below.
Link building
Social media
Where the airtime went
AI and LLM topics went from entirely absent to the majority of episodes in four years. AEO and GEO, the more specialized "how do I get cited" vocabulary, took longer to become a distinct topic but is now growing from a real base rather than from zero.
The full record
| 09 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Link building | share % | 44 | 50 | 23 | 41 | 51 | 40 | 27 | 20 | 49 | 42 | 50 | 36 | 42 | 61 | 34 | 26 | 29 | 21 |
| rate /100k | 53.2 | 33.2 | 10.3 | 48.7 | 56.8 | 39.0 | 17.6 | 8.7 | 34.5 | 56.3 | 63.2 | 29.5 | 25.9 | 74.3 | 46.7 | 73.5 | 6.3 | 7.3 | |
| Social media | share % | 56 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 97 | 96 | 97 | 92 | 77 | 79 | 92 | 82 | 95 |
| rate /100k | 132.6 | 331.4 | 414.2 | 308.2 | 277.7 | 274.3 | 258.1 | 306.0 | 240.4 | 148.4 | 194.5 | 123.9 | 79.5 | 95.9 | 79.8 | 70.3 | 63.7 | 49.2 | |
| AI / LLM | share % | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 10 | 34 | 78 | 100 |
| rate /100k | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 13.5 | 7.9 | 47.2 | 76.1 | |
| AEO / GEO | share % | 0 | 2 | 9 | 2 | 3 | 7 | 5 | 2 | 9 | 13 | 7 | 3 | 12 | 10 | 7 | 8 | 38 | 63 |
| rate /100k | 0.0 | 0.4 | 2.9 | 0.4 | 0.5 | 1.8 | 0.7 | 0.3 | 2.4 | 10.8 | 1.2 | 0.4 | 3.7 | 2.1 | 1.2 | 1.1 | 9.3 | 33.0 |
How this was counted
The corpus is the full transcript archive of The Best SEO Podcast: 682 episodes, 2009 through 2026, one show only. Guest appearances on other shows (Matt Bertram co-hosts several) are excluded, and no other show's archive is pooled in — a second show has a different host, a different audience, and a different start date, and mixing them would make the count unauditable.
Each episode's full transcript is matched against a fixed set of terms per topic (for example, link building: "link building", "backlink", "guest post", "anchor text", bare "link"/"links"). A topic is scored two ways for each year:
- Episode share: the percentage of that year's episodes in which the topic's terms appear at least once. An episode counts once no matter how many times the topic comes up, so this measure cannot be inflated by one long episode and is not affected by the show's episodes getting longer over time.
- Rate: the number of term matches per 100,000 words that year, across every episode's transcript pooled together. This measure is sensitive to how much time is actually spent on a topic, which is exactly what episode share cannot see.
Three explanations for the link-building and keyword decline were tested and rejected before it was reported as a finding. Longer recent episodes diluting the rate was ruled out because episode share, which cannot be diluted by length, declines too. Missing modern vocabulary was ruled out by broadening the term list to include newer phrasing ("digital PR", "guest post", "anchor text") without changing the result. Transcription or tokenization drift was ruled out by checking that substring and word-boundary matching track each other in every year. What remains is displacement: AI and LLM terms occupy the airtime that link-building and keyword terms used to.
The published Buzzsprout show-notes summaries carry a boilerplate blurb appended to recent episodes that would otherwise inflate every 2025–26 episode's topic count regardless of what the episode is actually about; that boilerplate is stripped before matching, the same rule used on the AI-search pivot chart on the About page.
What's true about AI search today
The chart above shows that the conversation shifted. It does not by itself say what is actually true about AI search and AEO right now. The claims below are drawn from EWR's Knowledge Bank, a set of findings verified against primary studies before being marked reliable enough to publish.
AI Overviews now cover roughly half of all Google searches
Independent trackers converge on AI Overviews triggering on roughly 55–60% of Google searches as of mid-2026, up from about 25% in mid-2024, with rate rising sharply by query intent: comparison queries trigger an AIO more than 95% of the time, question-format queries around 86%, informational queries in the high 30s, and transactional or branded queries only around 5–8%. AIO is now the default experience for research-shaped queries, not an edge case.
Ranking is probabilistic, not positional
Google's AI Mode and AI Overviews break a query into dozens or hundreds of parallel sub-queries ("query fan-out") and assemble an answer from a corpus retrieved across all of them, not from the single head-term ranking. A page that covers more of that fan-out set is measurably more likely to be cited, and a majority of AI Overview citations do not come from pages that rank in the top 10 for the query that triggered the answer. Coverage of a topic beats matching a single keyword.
A single measurement is not a measurement
AI answers are stochastic enough that re-running the identical prompt on the same engine seconds apart returns a different set of cited sources most of the time. Reliable AI-visibility measurement requires roughly 7–8 runs per prompt across at least 8 prompts, reported as a visibility percentage with a confidence interval, never as a rank or position. Any single-run before/after test, including one run on this page's own claims, is noise until it clears that floor.
Recent content is favored, and the strength of that bias varies by engine
Answer engines skew citations toward recently published content, but not equally: AI Overviews shows the strongest recency preference, Perplexity is close behind, and ChatGPT tolerates older, durable sources the most. The bias is also topic-dependent, sharpest in fast-moving verticals and weakest on evergreen, technical subject matter.
Each AI engine pulls from a different source mix
Answer engines do not share a citation pool. ChatGPT leans on Reddit, mainstream press, and Wikipedia; Google's AI surfaces favor YouTube and schema-structured web content; Perplexity rewards primary sources and named experts. Only a small fraction of AI citations match a page's own Google top-10 ranking, so being optimized for classic search does not transfer automatically to being cited by an AI engine, and being cited by one AI engine does not mean the others will follow.
Schema markup does not appear to increase AI citation rate
Current best evidence, one matched-control study
The one controlled experiment run on this question, comparing pages that added schema markup against matched pages that did not, found no citation lift on Google AI Overviews, AI Mode, or ChatGPT, and a small negative effect on AI Overviews specifically. Schema still earns its place for entity disambiguation and Google's rich results; the evidence just does not support selling it as an AI-citation tactic.
llms.txt shows no measurable citation effect on major engines
Current best evidence, early and thin
The one implementation study available found no measurable traffic or citation change on 8 of 9 sites that added an llms.txt file, and no major AI platform has publicly confirmed reading it for citation decisions. Its one confirmed use case is helping IDE coding agents and other tools that browse a site directly, not general AI search visibility.
Frequently asked questions
Why does this study only use one show?
Matt Bertram co-hosts other shows (the Oil and Gas Global Network podcast among them), but pooling shows with different hosts, audiences, and start dates would break the measurement and trade a defensible single-corpus claim for a number nobody could check. This study is one show, 18 unbroken years.
What counts as a mention?
A term match anywhere in the episode's transcript for episode share; a raw count of term matches per 100,000 words for rate. See “How this was counted” above for the exact term lists and the checks run against alternative explanations.
Is the raw data available?
Yes. The full year-by-year table is published as JSON at /research/link-building-to-ai-search.json under CC BY 4.0, so it can be checked or reused without re-scraping this page.
Does this mean link building stopped working?
This study measures conversation, not ranking outcomes. It shows link building is discussed far less at length on this show than it used to be, and mentioned in fewer episodes than it once was. Whether links still move rankings is a separate, ranking-side question this corpus does not answer.
Found an error, or want the full corpus for your own analysis? Tell us.